STOP Customizing
STOP Technical Debt
START Leveraging AI
Why Your Platform Customization Spend Has Zero AI Leverage — And What to Do Instead
When AI can build unlimited code overnight, the economics flip from "generic SaaS + customization" to "define what you want and regenerate it."
What You'll Learn
- ✓ The Leverage Test — How to identify which platforms trap your investment
- ✓ The Three Tiers — What to keep buying vs what to build with AI
- ✓ Sunk Cost as Spec — Turn your existing config into AI-ready requirements
- ✓ Overnight Sprints — From 6-week cycles to morning PRs
Scott Farrell
LeverageAI • January 2026
The Stupidity Tax
You're already building custom software. You're just doing it badly, inside someone else's prison, with zero AI leverage.
"You're paying $500K/year to build custom software inside someone else's prison — and calling it a 'platform investment.'"
The Hidden Reality: You're Already Building Custom
Here's the dirty secret of enterprise platforms: you're building custom software anyway. You just don't call it that because it happens inside someone else's system.
When you buy Salesforce, HubSpot, or ServiceNow, you're buying a starting point — not a solution. The "out of the box" promise rarely survives first contact with actual business requirements.
The platform becomes the constraint, not the capability. Every business-specific requirement that doesn't fit the vendor's model becomes a "customization project." Every competitive advantage you want to build requires fighting against what the platform assumes you should do.
Salesforce implementation costs in 2026 typically range from $10,000 to over $200,000 for complex projects1 — and that's just the implementation. Before the ongoing customization even begins. One analysis firm warns that "many projects that started at $50K ended up at $150K or more."2
The Stupidity Tax Formula
Let's name this pattern with a formula — because naming it makes it visible, and visibility is the first step to escape.
The Stupidity Tax
"Made for the average" + "Pretend you're not writing custom software" = Stupidity Tax
"Made for the Average"
Platform vendors optimise for their median customer. Your business is not their median customer. Every industry-specific need, every competitive advantage — that's a deviation from average. And deviations cost.
"Pretend It's Not Custom"
The psychological trap: "We bought software." But when you hire consultants, you're building — just with extra constraints. When you configure workflows, you're coding — just in a worse language. When you integrate systems, you're architecting — just without version control.
"= Stupidity Tax"
You pay full price for generic software. Then you pay again to make it specific. And you get zero AI leverage on either investment. Triple loss: money, time, future leverage.
The Real Cost Breakdown
Let's make the economics visceral with a real example. This is what platform ownership actually costs when you count everything.
When Salesforce Costs $200K/Year...
That's nearly $1 million to make generic software fit your specific workflows.
The consultant rates alone tell the story: junior Salesforce consultants charge $60 to $100 per hour, mid-level $100 to $150 per hour, and senior consultants $150 to $250 per hour.3 Professional consulting firms handling end-to-end implementation charge $90 to $300 per hour.4
| What You Think You're Paying For | What You're Actually Paying For |
|---|---|
| Software license | Making generic software specific |
| Platform capabilities | Consultants to configure it |
| "Best practices" | Staff to maintain the customizations |
| Vendor roadmap | Workarounds when their roadmap doesn't match yours |
| Enterprise support | Translation layer between your needs and their system |
The average Salesforce administrator salary is around $85,000 annually, with U.S. salaries ranging from $80,000 to over $130,000.5 Freelance rates range from $80 to $150 per hour. That's the cost of institutional archaeology — maintaining tribal knowledge of "why the system works this way."
Key person risk compounds this: if your admin leaves, knowledge walks out the door. No version control, no documentation, no tests. Just oral tradition about which workflow triggers which automation.
Where the Money Actually Goes
Follow the dollar to understand why this is a trap. The platform vendor's moat is your prison.
The Vendor's Economics
Of a $100K/year SaaS contract, maybe $10-20K actually goes to software delivery. The rest is overhead you're subsidising:
- • Sales costs: 20-30% of revenue
- • Marketing costs: 15-25%
- • Channel partner margins: 20%+
- • Gross margin requirements: 20%+
You're funding their growth, not your capability.
The Consultant's Economics
Implementation partners bill hourly or daily. Their incentives don't align with yours:
- • Their goal: Maximise billable hours
- • Your goal: Minimise configuration complexity
Every "best practice" they recommend = more configuration to maintain.
The No-Leverage Trap
Here's the thesis that will reshape how you think about platform investment: where does your customisation dollar compound?
Platform configuration returns are linear, diminishing, and locked-in. Every dollar you pour into platform configuration:
- ✗ Doesn't benefit from model improvements
- ✗ Doesn't compound with your team's learning
- ✗ Doesn't transfer to your next system
- ✗ Locks you deeper into someone else's roadmap
AI can only help a little bit with Salesforce in practice. Platform customisation sits where AI is weakest: proprietary configuration languages that aren't in AI training data, closed ecosystems with no API access for AI tools, and real-time configuration changes that are high-stakes, one-shot interactions.
The Contrast: Same Dollar, 16x Difference
Platform Configuration
AI does maybe 5% of the work. Proprietary languages, closed systems, no leverage.
Spec-Driven Development
AI does 80% of the work. Open code, version control, compounding leverage.
In January 2026, Andrej Karpathy observed: "I rapidly went from about 80% manual coding and 20% agents in November to 80% agent coding and 20% edits in December. I really am mostly programming in English now."
The Escalating Trap
It gets worse over time, not better. SaaS inflation compounds the stupidity tax every year.
SaaS inflation rate vs general market inflation
Vertice 2026
SaaS pricing increase 2024-2025
SaaStr 2025
SaaS cost per employee (up from $7,900 in 2023)
Vertice 2026
SaaS inflation is now nearly 5x higher than the standard market inflation rate of G7 countries — and is ever increasing whilst the consumer inflation index drops.6 B2B pricing surged approximately 11.4% compared to 2024, versus a 2.7% average market inflation rate.7
The compound effect is brutal: platform licenses rise 10-12% annually, consultant rates 5-10%, admin salaries 3-5%, and integration maintenance scales with complexity. Total platform spend compounds at 8-15% annually.
Meanwhile, AI development costs have collapsed 70-90%. The economic inversion is complete: custom AI development costs have plummeted while SaaS prices rise.
Chapter 1 Summary: The Stupidity Tax
"The stupidity tax is: paying full price for generic + paying again to make it specific + getting none of the AI leverage."
- • You're already building custom software — just badly, inside constraints, with extra translation layers
- • The Stupidity Tax formula: "Made for average" + "Pretend not custom" = Stupidity Tax
- • Real cost is 3-5x the license fee when you count consultants, admins, and integration
- • Your customisation dollars go to translation, not capability building
- • Platform configuration offers zero AI leverage — AI can barely touch proprietary systems
- • Costs compound while alternatives get cheaper — the inversion is accelerating
But wait — does this mean abandon all SaaS? No. Some things should stay SaaS. The question is: which ones?
Next: Chapter 2 — The Three Tiers
The Three Tiers
What to keep buying, what to build instead — and how to tell the difference.
"The plumbing — auth, Stripe — is real-time, efficient, cheap. We aren't customising it anyway. If it's cheap enough and we aren't customising, sure, buy it."
This Isn't Anti-SaaS
Chapter 1 was provocative. Now let's earn the credibility. Some software absolutely should stay SaaS.
The question isn't "buy or build?" — it's "where does customisation burden sit?"
The problem isn't SaaS itself. The problem is that platform vendors extended the SaaS promise to categories where it doesn't work. They took a model that works brilliantly for payments and authentication, and tried to apply it to CRM and workflow management.
The distinction that matters:
- ✓ Buying software works when you're plugging in, not fitting
- ✗ Building software becomes necessary when customisation burden exceeds benefit
What shifted: AI collapsed build costs, so the threshold moved. What was once clearly "buy" territory is now often "build" territory.9
Tier 1: Commodity Plumbing — Keep Buying
This tier stays SaaS because AI can't beat their operational depth. These vendors have something you'll never replicate: global infrastructure, regulatory relationships, and operational muscle built over decades.
Tier 1 Characteristics
| Category | Examples | Why It Stays SaaS |
|---|---|---|
| Authentication | Auth0, Okta, Firebase Auth | Security expertise, compliance certifications, real-time critical |
| Payments | Stripe, Braintree, Square | Fraud detection requires scale, PCI compliance, carrier relationships |
| Communications | Twilio, SendGrid, Mailgun | Carrier relationships, deliverability expertise, scale economics |
| CDN/Infrastructure | Cloudflare, Fastly, CloudFront | Global presence, DDoS protection, operational depth |
| Basic Storage | S3, GCS, Azure Blob | Commodity, zero customisation, operational excellence |
Why doesn't AI change this? These systems require operational depth that AI can't replicate:
- • Global infrastructure presence
- • Compliance certification programs
- • Real-time fraud/abuse detection
- • Carrier/network relationships
The value isn't in the code — it's in the operational muscle. Building this would cost more than subscribing forever.
Tier 2: Horizontal Platforms — The Flip Zone
This is where the economics have inverted. This is where your customisation dollars have zero AI leverage. This is the flip zone.
Tier 2 Characteristics
Expensive base cost: $50K-500K+ annual license fees
Massive customisation projects: Implementation partners, consultant retainers
Internal admin staff: Dedicated employees just to maintain the platform
"Implementation partners": Euphemism for "army of people making it work"
The dirty secret: You're paying for translation, not software.
| Category | Examples | Why It's Flip Zone |
|---|---|---|
| CRM | Salesforce, HubSpot, Dynamics 365 | Massive customisation burden, workflow configuration complexity |
| ITSM | ServiceNow, Jira Service Management | Custom workflows, integration requirements, process alignment |
| HCM | Workday, SuccessFactors, BambooHR | Localisation, policy implementation, integration needs |
| ERP | NetSuite, SAP Business One | Process-specific configuration, vertical customisation |
| Marketing Automation | Marketo, Eloqua, Pardot | Campaign logic, scoring models, integration requirements |
The economics are brutal. For a typical Tier 2 platform:
- • Platform license: $100K-200K/year
- • Implementation: $300K-1M (amortised)10
- • Internal admins: $150K-300K/year11
- • Consultant retainer: $50K-150K/year12
- • Integration maintenance: $50K-100K/year
Total ownership: 3-5x the license fee.13 We saw this in Chapter 1: a $200K Salesforce deployment costs nearly $1M when you count everything.
Why does AI change this? Spec-driven development: AI does 80% of code generation.14 Platform configuration: AI does ~5% (proprietary languages, closed ecosystems). Same dollar, 16x difference in AI leverage. The translation layer was the expensive part — AI eliminates it for code, not for config.
Tier 3: Industry Verticals — Also Flip Zone
Industry-specific platforms have the same problem as Tier 2, with extra deception. They promise specificity but deliver averages.
The Promise
"Built for your industry!"
The Reality
"Built for the average of your industry"
The core platform is still 60-70% generic — the "vertical" is just a skin.15 Your industry specifics don't match their average, so you're back to customisation work. But now you're paying premium pricing because "industry expertise."
| Category | Examples | The Gap |
|---|---|---|
| Insurance CRM | Applied Epic, Vertafore | Your specific products, regulations, distribution differ |
| Healthcare Scheduling | Athenahealth, Epic MyChart | Your specific workflows, compliance requirements differ |
| Legal Practice | Clio, PracticePanther | Your specific matter types, billing structures differ |
| Real Estate | Yardi, AppFolio | Your specific property types, market practices differ |
| Construction | Procore, Buildertrend | Your specific project types, trade relationships differ |
Why "industry-specific" is often worse:
- • You're STILL customising, just with industry-specific constraints
- • The platform's "best practices" may not match your actual practices
- • Switching costs are higher (industry data formats, regulatory compliance)
- • Fewer alternatives within the vertical = less negotiating leverage
The Decision Heuristic: The 50% Threshold
Let's make this concrete. Here's a rule you can apply to your own stack:
The 50% Threshold
If customisation burden > 50% of total cost → build with AI
The "Buy" Component
License/subscription fees — what you pay the vendor directly.
The "Customisation" Component
- • Implementation costs (amortised)
- • Consultant and contractor fees
- • Internal admin/developer time
- • Training and certification
- • Integration development and maintenance
- • Change request overhead
Salesforce Deployment
License portion: 27%
Customisation burden: 73%
Verdict: Build candidate
Stripe Deployment
License portion: 91%
Customisation burden: 9%
Verdict: Keep buying
The Three Questions to Ask
Operationalise the framework with these diagnostic questions. Apply them to each platform in your stack.
Question 1: "Are we plugging in or fitting in?"
Plugging In (Tier 1)
- • Standard API integration
- • No business logic in the platform
- • Platform does one thing well
- • We use it as a utility
Fitting In (Tier 2-3)
- • Extensive configuration
- • Business logic encoded in platform
- • Platform does many things adequately
- • We adapt our processes to it
Question 2: "Where does our differentiation live?"
If Differentiation Is In The Platform
- • Competitive advantage depends on configuration
- • Unique workflows encoded in proprietary system
- • Customer experience shaped by platform constraints
→ Flip zone — build instead
If Differentiation Is Elsewhere
- • Platform is pure plumbing
- • Competitive advantage in product/service
- • Platform invisible to customer
→ Keep buying
Question 3: "What happens when we need to change?"
Fast Change (Tier 1)
- • API update, minimal integration work
- • No configuration archaeology required
- • Changes are version-controlled code
Slow Change (Tier 2-3)
- • Requirements meetings to understand impact
- • Sandbox testing, UAT cycles
- • Implementation partner engagement
- • Production deployment windows
→ Flip zone if changes are frequent
The Cognition Ladder Connection
The Three Tiers map to a broader framework about where AI creates value.
Framework Cameo: The Cognition Ladder
The Cognition Ladder describes three rungs of AI value creation.16 Rung 1 (Don't Compete) is real-time, high-stakes territory where AI has 70-85% failure rates — you're competing with humans at what they do well. Rung 2 (Augment) is batch processing with 10-100x more thinking per task. Rung 3 (Transcend) is overnight hyper sprints — work that was never feasible.
Tier 1 software operates in Rung 1 territory — where vendors have operational depth you can't replicate. Tier 2-3 software operates where Rung 2-3 AI excels — where AI multiplies your capability, not the vendor's.
| Tier | Cognition Rung | AI Success Rate | Recommendation |
|---|---|---|---|
| Tier 1 | Rung 1: Don't Compete | 15-30% | Keep buying |
| Tier 2 | Rung 2-3: Augment/Transcend | 60-80% | Build instead |
| Tier 3 | Rung 2-3: Augment/Transcend | 60-80% | Build instead |
Chapter 2 Summary: The Three Tiers
"The question isn't 'buy or build?' It's 'where does customisation burden sit?'"
- • Not anti-SaaS — some software absolutely stays SaaS
- • Tier 1 (Commodity Plumbing): Auth, payments, messaging — keep buying
- • Tier 2 (Horizontal Platforms): CRM, ITSM, HCM — the flip zone
- • Tier 3 (Industry Verticals): "Built for your industry" that's still 60-70% generic — also flip zone
- • The 50% threshold: If customisation burden exceeds license cost, build instead
- • Three diagnostic questions: Plugging in or fitting? Where's differentiation? How fast can you change?
But even if you know which platforms to escape, there's a deeper problem. Platform configuration creates a different kind of technical debt — one with no solution path.
Next: Chapter 3 — Platform Debt vs Code Debt
Platform Debt vs Code Debt
Two types of technical debt. Only one has a solution path.
"This is why Agile got so popular — it's a way to sort of deal with the 'it's too hard' problem. Too complex. Too much technical debt. Just focus on a tiny thing and try to get that done."
A Coping Mechanism, Not a Solution
Here's an uncomfortable truth about Agile methodology: it was never a solution. It was a coping mechanism.
The standard origin story tells us Agile emerged as a response to waterfall failures. The deeper truth? Agile emerged because software got too complex to manage. "It's too hard" became the unstated premise. The answer: just focus on a tiny thing and try to get that done.
This matters for platform configuration because it has the same complexity problem. But platform config didn't inherit the coping mechanisms that code developed. Agile, CI/CD, testing, version control — all these emerged from code complexity. Platform configuration inherited the complexity without the solutions.
When artificial intelligence can generate implementations from specifications in minutes rather than months, the cost of iteration approaches zero.9 Spec-driven AI development makes complexity manageable again. And that changes everything.
Two Types of Technical Debt — Different Futures
Technical debt isn't one thing. There are two distinct types with completely different trajectories.
Platform Configuration Debt
- ✗ Compounds over time — every change harder
- ✗ No version control in the meaningful sense
- ✗ No tests (just "hope it works")
- ✗ Requirements meetings become archaeology expeditions
- ✗ AI can barely help — proprietary languages, closed ecosystems
Codebase Technical Debt
- ✓ Also compounds — BUT we have tools
- ✓ 50+ years of engineering discipline
- ✓ Version control (git), CI/CD, automated testing
- ✓ Clear interfaces, type systems, documentation
- ✓ AI is brilliant — understands, navigates, refactors
How Platform Configuration Debt Accumulates
The Complete Comparison
Let's make the contrast explicit. This table captures why platform configuration debt spirals while codebase debt can be managed.
| Dimension | Platform Config Debt | Codebase Debt |
|---|---|---|
| Version control | Primitive (snapshot-based) | Mature (git, branching, merging) |
| Testing | Manual, hopeful | Automated, systematic |
| Refactoring | Point-and-click archaeology | Structured, tool-assisted |
| AI assistance | ~5% of tasks | 80%+ of tasks |
| Documentation | Tribal knowledge | Code-as-documentation + docs |
| Change cost | Escalating over time | Flat or decreasing |
| Model upgrade benefit | Zero | Free improvement on regeneration |
| Error recovery | Rollback to snapshot (if remembered) | Git revert, branch reset |
| Knowledge transfer | Key person dependency | Codebase speaks for itself |
| Compliance audit | "What does this field do?" | "Read the commit history" |
of IT budgets consumed by tech debt
McKinsey
annual tech debt cost (US)
Netguru/McKinsey
cost multiplier: early fix vs late fix
Software Improvement Group
Signs You Have Platform Configuration Debt
These symptoms show up in every organisation with significant platform customisation. How many do you recognise?
1. Archaeology Expeditions
- "Let me find out who built this workflow..."
- "I think this was from the 2019 implementation..."
- "We can't touch that — no one knows what it does"
2. Key Person Dependency
- • One admin who "knows where everything is"
- • When they're on leave, nothing gets changed
- • Their departure would be catastrophic
3. Shadow Documentation
- • Spreadsheets explaining what the platform does
- • "The real requirements" live outside the platform
- • Confluence pages that are always out of date
4. Change Request Terror
- • Simple changes require weeks of impact analysis
- • "That sounds easy but it touches a lot"
- • Sandbox testing for trivial modifications
5. Integration Fragility
- • "Don't update that — it'll break the sync"
- • Integrations held together with hope and Zapier
- • No one knows the full data flow
6. Conflicting Automations
- • Workflow A triggers when B also triggers
- • Race conditions in platform logic
- • "Why did this record change twice?"
Common symptoms include: proliferating custom fields, conflicting automations, manual spreadsheet workarounds, brittle integrations, and slow admin cycles for simple requests.10
Why Code Has Solutions (And Platforms Don't)
The tools we have for code didn't appear magically — they solved real problems. Platform configuration lacks these tools because vendors didn't build them.
Version Control
Automated Testing
Continuous Integration
AI Assistance
50 years of engineering discipline for code; 0 years for platform config. These disciplines emerged because code complexity demanded them. Platform configuration inherited the complexity without the solutions.
What AI Can Do for Code (But Not Config)
This connects directly to the leverage thesis. AI creates massive value for code because it has the right "shape" for AI assistance. Platform configuration doesn't.
What AI Does Brilliantly with Code
Understanding structure
Parses ASTs, navigates dependencies, understands function signatures, types, interfaces, recognises patterns across codebases
Making changes
Refactors while maintaining behaviour, adds tests to validate changes, updates related code (imports, types, documentation)
Catching errors
Static analysis integration, test generation and execution, suggesting fixes for failures
Generating from spec
Converts requirements to implementation, produces boilerplate, scaffolding, patterns
What AI Can't Do with Platform Config
Understanding structure
Proprietary config formats not in training data, no standard AST for Salesforce workflows, platform-specific "languages" are documentation-poor
Making changes
Can't click through config UIs, no API for most configuration changes, no way to test changes safely
Catching errors
No static analysis for platform config, no way to run "platform tests", can't predict cascade effects
Generating from spec
Can't generate Apex → Flow → Workflow → Process Builder chains, no regeneration path, config isn't derivable from requirements
The Leverage Gap
80%20
AI for code
~5%
AI for platform config
AI can only help a little bit with Salesforce in practice. The 80% vs "a little bit" gap is the leverage difference.
The New Development Model
If Agile was coping with "too complex," what happens when spec-driven AI makes complexity manageable? A new model emerges.
Waterfall Per Increment
Detailed spec per slice
AI can execute from precise specifications — precision pays off
Ruthless evaluation per slice
AI can run comprehensive tests — every slice is validated
Agile between slices
Humans still drive requirements and priorities — flexibility preserved
Each "slice" can be a night's work, not a sprint's work. We'll explore this in Chapter 6.
Why can't platform configuration adopt this model?
- ✗ No spec language for platform configuration
- ✗ No way to regenerate config from spec
- ✗ No automated evaluation for config changes
- ✗ Stuck with "hope and archaeology"
The Diverging Paths
The two types of debt have completely different futures. Only one has a path to improvement.
Platform Configuration Debt Trajectory
Year 1
Clean implementation, understood by implementers. Changes are fast and predictable.
Year 3
Original implementers gone. Some archaeology required for changes. Starting to hear "we can't touch that."
Year 5
Major archaeological expeditions for any change. Key person dependency (the "Salesforce whisperer"). Shadow documentation everywhere.
Year 7+
Complete ossification. "We need to re-implement from scratch" — but re-implementation is on the same platform (repeating the cycle).
Codebase Debt Trajectory (with discipline)
Year 1
Clean implementation, tests pass. Changes are fast and predictable.
Year 3
Some debt accumulation (normal). Refactoring sprints address it. Tests continue to validate changes.
Year 5
Continuous improvement culture. AI tools help with refactoring. Can regenerate components from specs.
Year 7+
Codebase is actually cleaner than year 3. Model improvements make regeneration cheaper. Debt doesn't compound — it gets managed.
Platform config: Debt compounds forever. | Codebase: Debt can be managed, reduced, or regenerated away.
Chapter 3 Summary: Platform Debt vs Code Debt
"Platform configuration has the same complexity as code — without 50 years of engineering discipline to manage it."
- • Agile emerged as a coping mechanism for "too complex" — not as a solution
- • Platform configuration debt compounds with no solution path — no version control, no tests, no refactoring
- • Codebase technical debt has 50+ years of discipline: git, CI/CD, testing, refactoring
- • AI can do 80% of code work; "a little bit" of platform config work — 16x leverage gap
- • Platform config trajectory: ossification and key person dependency
- • Codebase trajectory: manageable, improving, regeneratable
So we know the trap (Ch1), we know where it applies (Ch2), we know why it compounds (Ch3). Now we need the decision tool: Where does your customisation dollar get AI leverage?
Next: Chapter 4 — The Leverage Test
The Leverage Test
Where does your customisation dollar compound? The new question that changes everything.
"I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December. i.e. I really am mostly programming in English now."
The Paradigm Shift: Karpathy's Confession
When elite programmers change their 20-year workflow in weeks, the ground has shifted. This isn't future speculation — it's current practice.
"This is easily the biggest change to my basic coding workflow in ~2 decades of programming and it happened over the course of a few weeks." The shift happened in weeks, not years. Elite programmers don't change 20-year habits for marginal improvements.
Most telling: "I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last ~year." Even Karpathy feels behind. "A failure to claim the boost feels decidedly like skill issue."
If the best programmers feel they're not capturing the full advantage, what about platform configuration teams?
The Evidence Is Everywhere
Karpathy isn't alone. The data is consistent across sources.
of all code written in 2025 is AI-generated
Index.dev
of developers use or plan to use AI tools in 2025
GitHub Survey
faster task completion with Copilot
McKinsey
increase in "vibe coding" searches, spring 2025
Collins Dictionary
Elite Developer Patterns
- • Theo Brown: 11,900 lines without opening an IDE, 6 parallel Claude Code instances25
- • 90% of personal code now AI-generated among leading developers25
- • 25% of YC's Winter 2025 batch had codebases that were 95% AI-generated26
These aren't experiments — they're production workflows.
The Core Question: Where Does Your Dollar Compound?
It's time to reframe the build vs buy question as a leverage question.
The Old Question
"Should we build or buy this software?"
- • Which is cheaper upfront?
- • Which has lower TCO over 5 years?
- • This was the right question when build was expensive
The New Question
"Where does our customisation dollar get AI leverage?"
- • Where does investment compound?
- • Where do model improvements help us for free?
- • This is the right question when build costs have collapsed
The Leverage Test
Where does your customisation dollar get AI leverage?
Platform Customisation Spend → No Leverage
- • Every dollar goes into a system AI can barely touch
- • Consultants doing manual translation (intent → platform config)
- • Returns: linear, diminishing, locked-in
- • Model improvements: zero benefit
AI-Enabled Team Spend → Massive Leverage
- • Every dollar compounds because AI multiplies capability
- • Staff become 20-100x more effective
- • Returns: compounding, improving, portable
- • Model improvements: free capability upgrade
Why This Just Happened
Why is January 2026 different from January 2024? Model capabilities crossed a threshold.
The Capability Threshold (Late 2025)
Agentic coding became reliable for complex tasks. Models can now:
Karpathy's shift happened "over the course of a few weeks" in Nov-Dec 2025.
Programming is no longer just about languages, data structures, and algorithms. A completely new layer has been added on top. Today's leverage comes from knowing how to prompt and orchestrate AI agents, give them memory and context, and connect them to codebases via tools.
The 20x-100x Claim
Let's quantify the leverage difference. Where does "20x to 100x more effective" come from?
Speed Multiplier
3-10x- • Tasks that took days take hours
- • Tasks that took hours take minutes
- • Overnight batch processing enables 24/7 development
Quality Multiplier
2-5x- • More thorough exploration of solutions
- • Better test coverage (AI generates tests)
- • Fewer regressions (AI catches consistency issues)
GPT-3.5 with agentic workflows (95%) outperforms GPT-4 alone (48%) on HumanEval27
Availability Multiplier
4x- • Your Salesforce admin is awake 40 hours a week
- • Your AI team is awake 168 hours a week
- • Overnight sprints while humans sleep
The Combined Effect
Conservative estimate:
3x speed × 2x quality × 4x availability = 24x
Optimistic estimate:
10x speed × 5x quality × 4x availability = 200x
The 20x-100x claim is supported by this multiplication. Platform configuration: 1x (human speed, no leverage).
The Economic Inversion
SaaS costs are rising while build costs have collapsed. The economics have flipped.
SaaS Cost Trajectory ↑
- • SaaS inflation: 5x higher than general market inflation6
- • 2025 SaaS pricing: up 11.4% vs 20247
- • Platform costs: rising 10-15% annually
Build Cost Trajectory ↓
- • Custom AI development costs: collapsed 70-90%
- • CTO of Vercel: "Opus + Claude Code now behaves like a senior software engineer"28
- • "The cost of software production is trending towards zero"28
5 years ago: Build was clearly more expensive. Today: Build is often cheaper.
The crossover happened in 2024-2025. Most organisations haven't noticed yet.
The Leverage Test Applied
Let's make the test actionable. For any software category, ask three questions:
Can AI help with customisation?
If no → platform work has no leverage
What's the customisation burden?
If >50% → build candidate
What's the model improvement trajectory?
If improving → build advantage grows
Example: Salesforce CRM
| Can AI help with customisation? | Barely — proprietary config, closed ecosystem |
| Customisation burden? | 73% (from Ch1 calculation) |
| Model improvement trajectory? | Zero benefit to platform config |
| Verdict | Flip zone — build instead |
Example: Stripe Payments
| Can AI help with customisation? | Not relevant — no customisation needed |
| Customisation burden? | 9% (minimal integration) |
| Model improvement trajectory? | N/A — commodity plumbing |
| Verdict | Keep buying — Tier 1 commodity |
The One-Liner
"If AI can't help you with it, you're building the wrong way."
Platform config:
AI can barely help
Spec-driven development:
AI does 80%
Chapter 4 Summary: The Leverage Test
"The question isn't 'is it cheaper to build?' It's 'where does my dollar compound?'"
- • Karpathy went from 80% manual → 80% AI in weeks (defining moment)
- • 41% of code is now AI-generated; elite developers run 6 parallel Claude Code instances
- • The Leverage Test: Where does your customisation dollar get AI leverage?
- • Platform spend: No leverage — linear, diminishing, locked-in
- • AI team spend: 20x-100x leverage — compounding, improving, portable
- • Economic inversion: SaaS costs rising 10-15%/year, build costs collapsed 70-90%
Now we understand WHY to escape the platform trap. Part II shows HOW — starting with the counterintuitive insight that your sunk cost is actually your requirements document.
Next: Chapter 5 — Sunk Cost as Spec
Sunk Cost as Spec
Your current platform configuration IS your requirements document. Extract it, translate to spec, regenerate as software you own.
"You don't start from zero. You start from 'what we actually built, expressed as platform configuration.' Export that, translate to spec, regenerate as software you own."
The Sunk Cost Reframe
The standard view: Sunk cost is wasted money, don't let it trap you. The new view: Sunk cost is captured requirements, use it as an asset.
Old Thinking
"We can't leave Salesforce — we've spent $2M configuring it"
The sunk cost traps you.
New Thinking
"We have $2M of requirements documentation sitting in Salesforce"
The sunk cost becomes your spec.
Organisations resist platform escape because "we've invested so much." But the investment isn't wasted — it's encoded knowledge. Your current configuration documents exactly what you need. The archaeology has been done; the requirements have been captured.
The money is spent either way. The question is: what do you extract from it?
What Your Configuration Actually Documents
Every configuration element you've built represents a requirement you discovered. Let's make that mapping explicit.
Custom Fields = Data Requirements
Every custom field describes data you need to capture. Field name, type, validation rules = a specification.
Example: Contract_Renewal_Date__c (Date, Required) → "We need to track when contracts renew"
Workflow Rules = Business Rules
Every workflow encodes a business decision. "When status changes to Closed Won, do X" = business process specification.
Example: Auto-assign to queue when priority = High → escalation business rule
Reports = Output Specifications
Every report documents what you need to see. Groupings, filters, calculations = analytical requirements.
Example: "Sales by region by quarter" → regional sales tracking with quarterly aggregation
Integrations = Interface Contracts
Every integration defines data flow between systems. Field mappings, sync frequency, error handling = requirements.
Example: Salesforce → ERP sync = "These fields must flow to financial system daily"
Validation Rules = Data Quality Requirements
Every validation encodes a business constraint. These constraints are often the hardest to discover — they're already documented.
Example: "Discount cannot exceed 40% without VP approval" → approval workflow spec
Page Layouts = UX Requirements
Which fields appear where = information architecture. Required vs optional = data capture rules. Field groupings = conceptual organisation.
The Extraction Process
Here's the strategic pattern for extracting your requirements from platform configuration.
Configuration Audit
Export all configuration metadata: objects, fields, workflows, validation rules, reports. Most platforms have metadata APIs or export tools. Goal: A complete picture of what's been built.
Implicit → Explicit Translation
Configuration is implicit specification. Translation makes it explicit and portable. "Custom field" → "Data requirement with name, type, rules." "Workflow rule" → "Business rule with trigger and action."
Gap Identification
What configuration serves no purpose? (Dead code equivalent). What requirements are missing? (Tribal knowledge). What conflicts exist? This audit often reveals how much complexity is unnecessary.
Spec Document Creation
Structure the requirements in spec format: data model, business rules, integrations, views. This becomes the regeneration source. The spec is now portable — not locked in any platform.
You're not reverse-engineering requirements. You're extracting requirements that were already captured. The platform was the spec repository; now you're liberating the spec.
Why AI Excels at This Translation
AI is good at exactly this kind of transformation.29 Structure → Structure is an AI sweet spot.
What AI Does Brilliantly
Pattern recognition
Identifies common configuration patterns, maps platform-specific concepts to generic concepts
Structure transformation
Converts one structured format to another — platform metadata → spec document
Consistency enforcement
Ensures the spec is internally consistent, identifies contradictions in configuration
Gap detection
Identifies what's missing or incomplete compared to common patterns
What Humans Still Do
Intent clarification
Why was this configured this way? What problem was it solving?
Priority setting
Which requirements are essential vs nice-to-have? What can we simplify?
Future direction
What do we want to add? What should we stop doing? What's changed since configuration?
Existing Behaviour Defines Acceptance Criteria
Here's a counterintuitive principle: Your current system, with all its quirks, is the acceptance test.30 If the new system does what the old one did, it's correct.
The Principle
Existing behaviour = acceptance criteria. "The new system should do what the old system does" is a valid spec.
- • Plus: Documented improvements and simplifications
- • Minus: Identified unnecessary complexity
Users already know the current system. Edge cases are already handled (or you'd have heard about them). The current system is the reference implementation.
Pain Points as Explicit Constraints
The things that frustrate users are requirements too. Capture them so the new system doesn't repeat them.
Process Friction
- "It takes 6 clicks to do something simple"
- "I have to update 3 screens for one change"
→ These become UX requirements in the new spec
Performance Issues
- "Reports take forever to load"
- "The system is slow on Monday mornings"
→ These become performance requirements
Integration Gaps
- "I have to manually copy data between systems"
- "The sync is always out of date"
→ These become integration requirements
Missing Capabilities
- "I wish I could do X, but the platform doesn't support it"
- "We built a workaround because the system can't..."
→ These become net-new requirements
The Pain Point Interview
- ? "What do you wish the system did differently?"
- ? "What workarounds have you built?"
- ? "What tasks do you dread because of the system?"
These answers often reveal requirements more important than configuration.
The Spec Format
What does the extracted spec actually look like? Here's the high-level structure.
Spec Document Structure
1. Data Model
Entities, attributes, constraints, relationships
Derived from: Custom objects, custom fields
2. Business Rules
Triggers, conditions, actions
Derived from: Workflows, validation rules, triggers
3. Integrations
Source/target systems, field mappings, timing, error handling
Derived from: Integration configurations, middleware
4. Views and Reports
Who sees what, calculations, filtering, grouping
Derived from: Page layouts, reports, dashboards
5. Permissions and Access
Who can do what, role hierarchy, field-level access
Derived from: Profiles, permission sets, sharing rules
Why this format matters: It's platform-agnostic. It's AI-consumable for regeneration.31 It's human-readable for review. It's version-controllable.
The Liberation Moment
Once the spec is extracted, you're free. The platform no longer owns your requirements.
Portability
Spec can generate any implementation. Switch vendors without starting over.
Versionability
Spec lives in git. Changes are tracked, diffable, revertable.
Regenerability
New model → regenerate from spec → better code for free. The spec is the durable asset.
Simplifiability
You can see the full spec at once. Unnecessary complexity becomes visible.
Platform Mindset
"We're locked in; we've invested too much"
→ Leads to platform debt
Spec-Driven Mindset
"We own our requirements; implementation is disposable"
→ Leads to leverage
Chapter 5 Summary: Sunk Cost as Spec
"Your current Salesforce implementation IS your requirements document. Export it, translate to spec, regenerate as software you own."
- • Sunk cost isn't waste — it's captured requirements
- • Configuration translates to spec: Custom fields = data requirements; workflows = business rules; integrations = interface contracts
- • The extraction process: Audit → Translate → Identify gaps → Create spec
- • AI excels at this: Pattern recognition, structure transformation, consistency checks
- • Existing behaviour defines acceptance criteria — plus improvements, minus unnecessary complexity
- • Pain points are requirements too — capture what frustrates users
- • Once spec is extracted, you're free: portable, versionable, regenerable
With the spec extracted, what happens next? This is where overnight sprints enter — the mechanism that takes a requirement and delivers a reviewed PR by morning.
Next: Chapter 6 — The Overnight Sprint
The Overnight Sprint
Requirement to deployed code in one night. PRs in your morning inbox.
"Sprints used to be weeks. Now they're nights."
The Time Compression
What used to take weeks now takes hours. This isn't incremental improvement — it's category change.
Old Timeline: Platform Change Request
Total: 6 weeks minimum
New Timeline: Spec-Driven Overnight
Total: ~12 hours (mostly unattended)
Same Business Outcome
6 weeks → 1 night
Business requirement fulfilled. Tested and reviewed. Deployed to production.
Why Overnight Works: The Batch Advantage
Overnight isn't just "working at night" — it's a different operating model. Batch processing unlocks capabilities real-time can't access.
Deep Reasoning
- • Extended thinking mode36
- • Multiple evaluation passes
- • Self-correction loops
- • No pressure to guess
Multi-Pass Verification
- • Run tests 5 times if needed
- • Retry failed steps with different approaches
- • Exhaustive edge case exploration
- • Quality over speed
Adversarial Checking
- • Second agent reviews first agent's work37
- • Explicit challenge step
- • "What could go wrong?" pass
- • Built-in red team
Progressive Resolution
- • Work coarse-to-fine without time pressure
- • Stabilise structure before detail
- • Each layer checked before advancing
- • No "heads-down drift"
Progressive Resolution Overnight
Work coarse-to-fine — stabilise structure before committing to detail.33 The Jenga Problem: High-resolution changes cascade into unexpected breakage. Solution: Lock each layer before advancing.
| Time | Layer | AI Activity | Stabilisation Gate |
|---|---|---|---|
| 5pm | L0: Intent | Parse business requirement | "Can I explain this in 30 seconds?" |
| 6pm | L1: Architecture | Check if structural changes needed | "Does change fit existing architecture?" |
| 8pm | L2: Component specs | Update affected module specifications | "Each component's role is clear?" |
| 10pm | L3: Interfaces | Regenerate API contracts if needed | "Interfaces are stable and tested?" |
| 1am | L4: Implementation | Generate code from specs | "Code compiles and unit tests pass?" |
| 4am | L5: Integration tests | Run full test suite | "All tests pass?" |
| 6am | Verification | Adversarial review, diff report | "Ready for human review" |
| 9am | Human | Review PR, approve, deploy | "Approved and deployed" |
The Agent Army Model
Not one agent — many agents in parallel. Scale the overnight capacity.
Example Agent Allocation
Source: Theo Brown runs 6 parallel Claude Code instances34 in production workflows
What Overnight Produces
The artifacts that appear in your morning inbox.
1. Reviewed Pull Request
Code changes with full diff. Test results attached. Linked to requirement. Ready for human approval.
2. Evidence Bundle
Specific data that informed decisions. Links to relevant specs and tests. Reasoning trace for key choices.
3. Change Impact Summary
What files changed. What tests cover the changes. Downstream effects. Risk assessment.
4. Suggested Review Focus
"Look closely at X because Y." Known edge cases highlighted. Areas of uncertainty flagged.
Not: "What should I do today?"
Instead: "Should I approve this PR?"
Cognitive load shifted from generation to judgment. Human time on highest-value activity.
Governance Built-In
Overnight sprints are MORE auditable than platform changes. Your AI development inherits CI/CD discipline.35
| Capability | Overnight Development | Platform Changes |
|---|---|---|
| Version control | Every change is a commit, full history | "Who changed that workflow?" |
| Regression testing | Existing tests run against new code | Manual UAT, hope-based QA |
| Canary capability | Deploy to subset, instant rollback38 | All-or-nothing deployment windows |
| Diff reports | What changed and why, before/after | "I think it was changed in 2019..." |
The Availability Multiplier
The Math
40
hours/week
Salesforce admin
168
hours/week
AI agents
4.2x
multiplier
Your Salesforce admin is awake 40 hours a week. Your AI team is awake 168.
The Direct Comparison
| Dimension | Platform Change | Overnight Sprint |
|---|---|---|
| Calendar time | 15-30 days | ~12 hours |
| Human hours required | 20-50 | 2-4 |
| Archaeology required | High | None (spec is source of truth) |
| Test coverage | Manual, incomplete | Automated, comprehensive |
| Rollback capability | Snapshot (maybe) | Git revert (certain) |
| Documentation | Maybe updated | Automatically generated |
Chapter 6 Summary: The Overnight Sprint
"Your Salesforce admin is awake 40 hours a week. Your AI team is awake 168."
- • The compression: 6 weeks → 1 night for the same business outcome
- • Overnight enables: Deep reasoning, multi-pass verification, adversarial checking
- • Progressive Resolution: Coarse-to-fine, layer-by-layer stabilisation
- • Agent army model: Parallel agents, each running until done
- • Morning inbox: PRs ready for review, not "what should I do today?"
- • Governance built-in: Version control, regression tests, diff reports — existing SDLC applies
But what about maintaining the code over time? This is where regeneration economics enter — why patching loses and regeneration wins.
Next: Chapter 7 — Regeneration Economics
Regeneration Economics
Why patching loses. Generation is now cheap. Maintenance is expensive. The economics have inverted.
"After ~5 patches, regeneration almost always wins economically."
The Economics Inversion
The old rules made sense when generation was expensive. Those rules have flipped.39
Old Economics (When Generation Was Expensive)
- • Creating from scratch: Days or weeks of developer time
- • Patching existing: Hours of developer time
Rational strategy: Patch as long as possible, only regenerate when forced
Cost ratio: Generate = 10 units, Patch = 1 unit
Threshold: Patch until 10+ patches
New Economics (When Generation Is Cheap)
- • Creating from scratch: Minutes to hours of AI time
- • Patching existing: Still requires human judgment and time
Rational strategy: Regenerate early, don't accumulate patches
Inverted ratio: Generate = 1 unit, Patch = 2-5 units
Threshold: Regenerate after ~5 patches
Maintenance is now MORE expensive than generation.
The Crossover Rule
| Patch Count | Action |
|---|---|
| 1-2 patches | Patch the output (trivial fixes) |
| 3-5 patches | Consider regeneration |
| 5+ patches | Fix the kernel, regenerate |
Patches 1-2
- • Simple fixes, obvious corrections
- • Low cognitive overhead
- • Faster than regeneration setup
Patches 3-5
- • Patches start interacting
- • Need to remember previous patches
- • Testing becomes more complex
Patches 5+
- • Cognitive overhead exceeds generation cost
- • Knowledge trapped in patch history
- • Model upgrade = applying all patches again
The compound effect: Each patch adds complexity. Complexity makes future patches harder.40 The 10th patch is 3-4x harder than the 1st. Regeneration cost stays constant.
Platform Configuration Can't Regenerate
This is the structural trap. Platform config has patching as the ONLY option.
Why Platforms Are Stuck
No spec to regenerate from
Platform configuration evolved incrementally. No source document. The configuration IS the only representation.
No regeneration path
Platforms don't support "regenerate this workflow." Each element was hand-created. Modifications are always patches.
Patches compound forever
5 years of platform configuration = hundreds of patches. No way to "clean slate" without migration.
Code
After 5 patches → regenerate from spec
Platform Config
After 500 patches → still patching
The platform trap is the absence of regeneration economics.
Model Upgrades = Free Improvement
Regeneration from spec captures model improvements automatically. Platform config gets nothing.
For Spec-Driven Development
- • Run your spec through the new model
- • Get better code automatically
- • No migration effort
Example trajectory:
Jan 2025: Same spec → Claude 3.5 → Good code
Jan 2026: Same spec → Claude Opus 4.5 → Better code, no effort41
The spec appreciates in value as models improve.
For Platform Configuration
- • Zero impact on platform configuration
- • Salesforce workflows don't get smarter
- • ServiceNow processes don't improve
Reality:
New models don't help your existing configuration
AI assistance for config is still minimal
No "regenerate from spec" path
Configuration stays exactly as complex as you built it.
The Delete Test
The Diagnostic
"If I deleted this output, could I regenerate it at will from my kernel?"
If Yes:
- • The kernel (spec) is the asset
- • The output (code) is ephemeral
- • You have regeneration capability
- • Model upgrades will help you
If No:
- • Knowledge is trapped in the output
- • The output has become the source of truth
- • You can't regenerate, only patch
- • You've lost regeneration economics
For Platform Configuration
"If I deleted this Salesforce org, could I regenerate it?"
Almost always no (until you extract the spec per Chapter 5). Knowledge is trapped in configuration. The configuration IS the source of truth.
For Spec-Driven Code
"If I deleted this codebase, could I regenerate it?"
Yes, from the spec + tests.43 Knowledge lives in the spec. The code is a rendering of the spec.
The Kernel-First Discipline
The Anti-Pattern: Output Nursing
- 1. Generate code
- 2. Find issues, patch them
- 3. Repeat patching
- 4. Knowledge accumulates in patches
- 5. Can't regenerate without losing patches
66% of developers spend MORE time fixing "almost-right" AI code than they save in initial writing42
The Pattern: Kernel-First
- 1. Generate code from spec
- 2. Find issues
- 3. Ask: "Is this a spec gap or one-off fix?"
- 4. If spec gap: Update the spec, regenerate
- 5. If one-off: Patch this output only
Knowledge flows into spec, not patches.44
"Is this a problem with THIS output, or a problem with how I'm specifying outputs?"
If the former: Fix the output. If the latter: Fix the spec.
The Diverging Trajectories
Year 1
Spec documents everything. Changes are spec updates + regeneration.
Year 1
Configuration works as designed. Changes are fast (click and save).
Year 3
Spec has evolved. No archaeology (spec is documentation). Team can read the spec.
Year 3
Some archaeology required. Starting to hear "we can't touch that." Original implementers gone.
Year 5
Model has improved 3-4 times. Each regeneration produces better code. Spec evolved with business.
Year 5
Significant archaeology burden. Key person dependency. Shadow documentation everywhere.
Year 7+
Codebase is cleaner than year 3. Models continue to improve. Debt doesn't compound.
Year 7+
"We need to re-implement from scratch" — but re-implementation is still platform config. Cycle repeats.
The Compound Advantage
Regeneration economics compound over time. This is the long-term leverage argument.
What Spec-Driven Compounds
- • Spec improvements: 2-3 per month
- • Model improvements: 3-4 per year45
- • Combined: Significant annual improvement
- • Over 5 years: Transformative advantage
What Platform Config Compounds
- • Complexity
- • Debt46
- • Key person dependency
- • Fear of change
The opposite of leverage.
Every dollar you pour into spec-driven development compounds.
Every dollar you pour into platform configuration accumulates.
Chapter 7 Summary: Regeneration Economics
"Generation is now cheap. Maintenance is expensive. The economics have inverted."
- • Old economics: Generation expensive, patching cheap → Patch as long as possible
- • New economics: Generation cheap, maintenance expensive → Regenerate after ~5 patches
- • The crossover rule: 1-2 patches acceptable, 3-5 consider, 5+ regenerate
- • Platform configuration can't regenerate — that's the trap
- • Model upgrades = free improvement for spec-driven code, zero benefit for platform config
- • The delete test: "Could I regenerate this from my spec?" If no, knowledge is trapped
With the doctrine established (Part I) and the mechanism explained (Part II), Part III shows how this applies to specific contexts: CRM, industry verticals, and solo operators.
Next: Chapter 8 — The CRM Escape
The CRM Escape
You didn't become a sales ops leader to spend 60% of your time on admin archaeology.
"You didn't become a sales ops leader to spend 60% of your time on admin archaeology."
The RevOps Reality
Let's paint the specific pain of CRM-heavy roles. Does this sound familiar?
Sales reps: time on administration
Only 34% on selling
SDRs: time on research
Only 20% on outreach
RevOps leaders: platform maintenance
Not revenue operations
The Leverage Test Applied to CRM
Question 1: Can AI help with customisation?
CRM Platform Configuration
- • Workflows: Proprietary visual builder, AI can't navigate
- • Custom fields: Point-and-click, no AI assistance
- • Reports: Drag-and-drop interface
- • Integrations: Platform-specific connectors
Verdict: ~5% of tasks
CRM as Spec-Driven Code
- • Data model: AI generates from spec
- • Business rules: AI implements from requirements
- • Reporting: AI builds query layers
- • Integrations: AI writes API connections
Verdict: 80%+ of tasks
Question 2: What's the customisation burden?
Typical Salesforce CRM deployment:1,2,3,4,5
License: $100K-200K/year
Implementation (amortised): $30K-100K/year
Admin FTEs: $150K-300K/year5,11
Consultant retainer: $50K-100K/year3,4,12
Integration maintenance: $30K-80K/year
Total: $360K-780K/year
25-35%
License portion
65-75%
Customisation burden
Well above 50%
Flip zone
Question 3: Model improvement trajectory?
Platform Configuration
Salesforce releases → May break your config. New features → Usually cost more. "Einstein" is limited. Your config stays as complex as you built it.
Spec-Driven CRM
Model improvements → Better generated code (free). Each quarter → Your CRM gets smarter. AI features → You build them yourself.
The Three Tiers for Your CRM Stack
Tier 1: Keep Buying
| Component | Examples | Why It Stays |
|---|---|---|
| Email delivery | SendGrid, Mailgun | Deliverability expertise, carrier relationships |
| Payment processing | Stripe | PCI compliance, fraud detection |
| Calendar/scheduling | Calendly | Simple utility, minimal customisation |
| Phone/SMS | Twilio | Carrier relationships, real-time infrastructure |
Tier 2: Flip Zone — Build Instead
| Component | Currently Using | What to Build Instead |
|---|---|---|
| CRM screens | Salesforce UI | Custom UI on your data model |
| Workflow automation | Process Builder | Spec-driven business rules |
| Pipeline management | Opportunity stages | Custom pipeline logic |
| Reporting | Salesforce Reports | Query layer on owned data |
What Overnight Sprints Enable for CRM
Adding a Custom Field
Platform Approach
Request → Impact analysis → Sandbox → Configuration → Testing → Deployment approval → Production
Timeline: 3-7 days
Overnight Sprint
Spec updated → Overnight: Data model updated, migration generated, tests run → Morning: Review PR, approve
Timeline: ~12 hours
Changing Workflow Logic
Platform Approach
Requirements meeting → Current state documentation → Impact analysis → Sandbox development → UAT testing → Production deployment window
Timeline: 1-3 weeks
Overnight Sprint
Spec updated with new business rule → Overnight: Implementation generated, tests run → Morning: Review logic, approve
Timeline: ~12 hours
What Changes for the RevOps Leader
Before: Platform-First
Daily Activities:
- • Debug why workflow fired twice
- • Investigate integration failure
- • Build one-off report for VP
- • Train new rep on workarounds
- • Meet with consultant about upcoming project
Time Allocation:
60% platform admin, 20% firefighting, 15% meetings about the platform, 5% actual revenue operations
After: Spec-Driven
Daily Activities:
- • Review morning PRs (changes from overnight)
- • Update specs based on business needs
- • Analyse pipeline and conversion data
- • Design revenue experiments
- • Work with sales on strategy
Time Allocation:
15% spec maintenance and review, 5% code review, 80% actual revenue operations
Before: Platform administrator → After: Revenue operations strategist
The Escape Path for CRM
Phase 1: Parallel Run
Months 1-3Keep Salesforce running. Build spec layer. Extract requirements. Stand up development environment. Build first module (read-only to start).
Phase 2: Feature Parity
Months 3-6Shift read operations: Reporting from new system. Dashboards on owned data. Maintain Salesforce for writes. No user training required yet.
Phase 3: Transition Writes
Months 6-12New UI for daily operations. Better UX than Salesforce allowed. Training on new system. New system becomes system of record.
Phase 4: Sunset
Months 12-18Historical data exported. Subscription cancelled. Consultant retainers ended. Admin FTE redeployed to higher-value work.
Chapter 8 Summary: The CRM Escape
- • RevOps reality: 60%+ time on platform maintenance instead of revenue operations
- • Leverage Test for CRM: 5% AI help for config vs 80%+ for code; 65-75% customisation burden
- • Three Tiers for CRM: Keep Stripe/SendGrid (Tier 1), build CRM screens/workflows (Tier 2)
- • Overnight sprints: New field 3-7 days → 12 hours; workflow change 1-3 weeks → 12 hours
- • Escape path: Parallel run → Feature parity → Transition writes → Sunset (12-18 months)
CRM is a horizontal platform trap. But what about industry-specific software that promises "built for your industry"? Same economics, extra deception.
Next: Chapter 9 — The Industry Vertical Trap
The Industry Vertical Trap
"Built for your industry!" Reality: "Built for the average of your industry" — and you're not the average.
"Industry solutions aren't custom enough — you spend a fortune fitting them correctly."
The Promise vs The Reality
The Promise
- "Built specifically for insurance/healthcare/legal/construction"
- "We understand your industry's unique needs"
- "Best practices from thousands of firms like yours"
- "Faster time to value than horizontal platforms"
Implies: Less customisation required, pre-built workflows that match your needs.
The Reality
- Core platform is 60-70% generic49
- "Industry features" are a vertical skin on horizontal architecture
- Your specific requirements don't match their 80th percentile customer
- Competitive advantage = deviation from their average
Reality: Same stupidity tax, different packaging.
How the Pattern Plays Out
Insurance Vertical Software
Promise: "Built for insurance carriers and MGAs"
Reality: Your product structures don't match their templates. Your state-specific regulations need custom handling. Your distribution model doesn't fit their channel assumptions. Your underwriting rules are competitive advantage, not standard.49
Result: Heavy customisation to make "insurance software" work for your insurance.
Healthcare Scheduling
Promise: "Built for healthcare workflows"
Reality: Your specific provider types have unique scheduling needs. Your compliance requirements depend on state and specialty. Your integration needs vary by EHR vendor. Your patient flow is your differentiation.
Result: Heavy customisation to make "healthcare software" work for your practice.
Legal Practice Management
Promise: "Built for law firms"
Reality: Your matter types don't match their templates. Your billing structures are firm-specific. Your workflows vary by practice area. Your client experience is your differentiation.
Result: Heavy customisation to make "legal software" work for your firm.
Construction Project Management
Promise: "Built for construction companies"
Reality: Your trade relationships are unique. Your project types have specific requirements. Your bidding process is competitive advantage. Your field workflows don't match generic templates.
Result: Heavy customisation to make "construction software" work for your company.
Why Industry Verticals Are Often Worse
Counter-intuitive: Industry-specific can be harder than horizontal. Here's why.
Reason 1: False Fit Assumptions
With horizontal platforms: Everyone knows it's generic. Customisation is expected from day one.
With industry verticals: Vendor claims it fits. Implementation teams assume less work. Gaps are discovered late. "Wait, you don't do X like everyone else?"
Reason 2: Less Flexibility
Horizontal platforms: Built for maximum configurability. Extensive API access. Large ecosystem.
Industry verticals: Built for a specific use case. Limited API access. Smaller ecosystem. Fewer consultants. When you hit a wall, you really hit a wall.
Reason 3: Higher Switching Costs
Industry-specific data formats: Your data is stored in their schema. Export is often poor or proprietary.
Industry-specific compliance: Vendor may hold certifications you need. Moving means re-certification. Lock-in is tighter, not looser.
The Leverage Test for Industry Verticals
Question 1: Can AI help with customisation?
Same proprietary configuration problem as horizontal. Often LESS documentation than Salesforce. Smaller user base = less training data for AI.50 Verdict: 5% or less — even worse than horizontal platforms.
Question 2: What's the customisation burden?
Higher license cost (premium for "industry fit"). Implementation still required. Fewer admins available. Higher consultant rates. Verdict: Often 50-80% — the premium isn't justified.
Question 3: Model improvement trajectory?
Same story as horizontal: Zero benefit. Vendor's AI features are generic add-ons. Your competitive advantage gets no AI help. Verdict: No leverage from model improvements.
The "Built for Your Industry" Deceptions
Deception 1: "Pre-built workflows"
Claim: "Workflows built for your industry out of the box"
Reality: Pre-built for their MEDIAN customer. Your workflows differ at the edges. The edges are where you compete. Pre-built saves maybe 20% of the work.
Deception 2: "Compliance built-in"
Claim: "Compliance requirements handled automatically"
Reality: Compliance varies by jurisdiction. Your specific needs require configuration. The "built-in" is a starting point, not a solution.
Deception 3: "Faster time to value"
Claim: "Deploy in weeks, not months"
Reality: Basic deployment may be fast. Customisation to your needs takes months anyway. Total project duration: Similar to horizontal.
Deception 4: "Best practices from peers"
Claim: "We've learned from thousands of firms like yours"
Reality: "Like yours" means broadly in your industry. Best practices = average practices. You don't want to operate like everyone else.
The Spec-Driven Alternative
What You Actually Build
Not:
An entire industry platform from scratch
Actually:
- • Data model for YOUR business
- • Workflows for YOUR operations
- • Integrations to YOUR systems
- • UX for YOUR users
YOUR industry expertise goes into YOUR spec. AI implements from your spec. No gap between "industry knowledge" and implementation.
Why Your Expertise Wins
What the Vendor Knows
- • How their median customer operates
- • Industry patterns from 3-5 years ago
- • General compliance requirements
- • Standard terminology and concepts
Frozen in their software.
What You Know
- • How YOUR business actually operates
- • Current market conditions and competition
- • Your specific regulatory environment
- • What your customers actually want
- • Where you're trying to differentiate
Evolves with your market.
Your team has the industry expertise that matters. That expertise should go into YOUR spec — not into hoping the vendor got it right.
Chapter 9 Summary: The Industry Vertical Trap
"Industry vertical software promises to fit your industry. It actually fits the average of your industry — and you're not trying to be average."
- • The promise: "Built for your industry" implies less customisation
- • The reality: 60-70% generic core, heavy customisation still required
- • Often worse than horizontal: False fit assumptions, less flexibility, higher switching costs
- • The alternative: Your industry expertise goes into YOUR spec, AI implements
- • Your expertise wins: Yours is current and specific; vendor's is average and frozen
So far we've discussed escaping existing platforms. But what if you're starting fresh? Solo operators and small teams have a different advantage.
Next: Chapter 10 — The Solo Operator Advantage
The Solo Operator Advantage
What if you don't have legacy platforms to escape? You have the biggest advantage of all: You start with leverage.
"Enterprises start with platform debt and need to escape. Solo operators can start with leverage."
The Clean Slate Advantage
Solo operators and small teams can skip the escape phase entirely. Starting fresh with spec-driven development = pure leverage from day one.
What Enterprises Face
- • Years of platform configuration to understand
- • Sunk cost psychology to overcome
- • Migration risk and downtime
- • Staff retraining costs
- • Political resistance from platform advocates
What Solo Operators Face
- • Nothing to migrate
- • Nothing to overcome
- • Nothing to retrain
- • Pure greenfield opportunity
The Trajectory Comparison
Enterprise Path
Day 1: Buy platform
Year 3: Platform debt accumulating
Year 5: Major archaeology burden
Year 7: "We need to start over"
Solo Operator Path
Day 1: Write spec, generate code
Year 3: Spec evolved, code improved
Year 5: Model upgrades improved code free
Year 7: System cleaner than year 3
Economies of Specificity
The economic advantage has inverted from economies of scale to economies of specificity.51
Old Economics: Economies of Scale
- • Larger companies had cost advantages
- • Spread fixed costs across more customers
- • Standardise to reduce per-unit cost
- • "Needs of the many outweigh needs of the few"
How platforms captured value: Build once, sell to thousands. Optimize for the median.
New Economics: Economies of Specificity
- • Customisation is now cheap
- • Per-customer optimisation is viable
- • Speed of learning matters more than scale
- • "Needs of the one can be met at scale"
What this means: Build exactly what you need. No compromise for the median.
The Coordination Cost Advantage
Small teams avoid the overhead that kills large org AI adoption. Speed of learning beats size.54
Enterprise Coordination Overhead
- • Consensus-building processes
- • Training programs (weeks to roll out)
- • Change management initiatives
- • Multiple approval layers
- • Integration with existing systems
$10M-$50M/year just to exist53
Solo Operator Overhead
- • Your time to decide
- • Your time to implement
- • That's it
$5K-$20K/year for tools
The Learning Speed Difference
Enterprise Learning54
Identify opportunity → Committee review → Pilot proposal → Approval process → Pilot → Evaluation → Rollout decision → Training → Full deployment
Timeline: 6-18 months
Solo Operator Learning
Identify opportunity → Try it → Keep or discard
Timeline: Days to weeks
AI tools change constantly. Best practices evolve rapidly. Solo operators can iterate at AI speed.
The Three Tiers for Small Teams
Same framework, different recommendations. Skip Tier 2-3 entirely, go straight to build.
Tier 1: Still Buy (same as enterprises)
The key: These are utilities you plug in, not platforms you customise.
Tier 2-3: Skip Entirely, Build Custom
Enterprises' path:
Buy Salesforce → Spend 3 years customising → Trapped
Solo operator path:
Never buy Salesforce. Build exactly what you need from day one.
Full AI Leverage from Dollar One
Every dollar compounds from the start. No migration cost, no platform debt.
The Enterprise Math
5 years of compounding lost.
The Solo Operator Math
No catch-up period, no migration tax.
What Solo Operators Actually Need
It's less than you think.
| Capability | Cost | Notes |
|---|---|---|
| AI tooling (API, IDE)55 | $100-300/month | Claude, Cursor, etc. |
| Hosting | $20-100/month | Vercel, Railway, Fly |
| Authentication | $0-50/month | Auth0 free tier, Clerk |
| Database | $20-100/month | Supabase, PlanetScale |
| Total | $150-500/month | Not $500K/year platform |
of YC's Winter 2025 batch had codebases that were 95% AI-generated
revenue, 98% margin, zero employees (Dan Koe)
When to Go Spec-Driven from Day One
Good Candidates
- • You know your domain deeply
- • Your needs are specific (not generic CRM)
- • You can write clear requirements
- • Speed matters more than "best practices"
- • You want to learn and iterate fast
Signals: "I wish this tool did X, but they all do Y"
When Platforms Still Make Sense
- • Your needs are truly generic (basic email marketing, standard HR)
- • You have zero technical inclination
- • You need compliance certifications immediately
- • You're testing a business model (pivot fast)
Apply the Leverage Test: If AI can help with 80% of the build work, and your needs require >30% customisation...
Go spec-driven from day one.
Chapter 10 Summary: The Solo Operator Advantage
- • Clean slate advantage: No migration, no platform debt, no catch-up period
- • Economies of specificity: Customisation is now cheap; optimise for your exact needs
- • Coordination cost advantage: Your learning cycle is days, not months
- • Three Tiers for small teams: Buy utilities (Tier 1), skip platforms entirely (Tier 2-3)
- • Full leverage from day one: Every dollar compounds immediately
- • Requirements: $150-500/month in tools, domain knowledge, clear requirements writing
Whether you're escaping platforms or starting fresh, the destination is the same: PRs in your morning inbox, leverage on every dollar, and time for the work that matters.
Next: Chapter 11 — Conclusion: The Morning Inbox
The Morning Inbox
PRs ready for review. Not "what should I work on today?" but "should I approve this change?"
"Stop funding your vendor's moat. Start funding your AI leverage."
What Monday Morning Looks Like
Old Model (Platform-First)
- • Arrive at work
- • Check what fires need fighting
- • Three Salesforce support tickets
- • Integration broke over the weekend
- • VP wants a report yesterday
- • Day becomes reactive firefighting
New Model (Spec-Driven)
- • Arrive at work
- • Open morning inbox
- • 4 PRs ready for review
- • Each PR addresses a spec update from last week
- • Tests passed overnight
- • Review, approve, deploy before 10am
- • Day is now proactive strategy work
From: "What do I need to DO today?"
To: "What do I need to DECIDE today?"
| Item | What It Contains | Time to Review |
|---|---|---|
| PR #1: New field on Account | Data model change + migration | 5 minutes |
| PR #2: Workflow update | Business rule + tests | 10 minutes |
| PR #3: Report modification | Query layer change | 5 minutes |
| PR #4: Integration fix | API handler + logging | 10 minutes |
| Total morning review | 30 minutes |
Compare to: 4 platform change requests = 4-8 weeks
The Journey We've Taken
Part I: The Platform Trap (Why)
Ch 1 — The Stupidity Tax
"Made for average" + "Pretend not custom" = Stupidity Tax. Real cost: 3-5x the license fee.
Ch 2 — The Three Tiers
Tier 1: Keep buying. Tier 2-3: The flip zone. If customisation >50%, build instead.
Ch 3 — Platform Debt vs Code Debt
Platform debt compounds forever. Code debt has 50 years of engineering discipline.
Ch 4 — The Leverage Test
"Where does your dollar get AI leverage?" Platform: No leverage. AI team: 20x-100x.
Part II: The Escape Path (How)
Ch 5 — Sunk Cost as Spec
Your configuration IS your requirements document. Extract it, translate it, own it.
Ch 6 — The Overnight Sprint
6 weeks → 1 night for same business outcome. Morning inbox: PRs ready for review.
Ch 7 — Regeneration Economics
Old: Patch as long as possible. New: Regenerate after ~5 patches. Model upgrades = free improvement.
Part III: Applying the Leverage Test (Where)
Ch 8 — CRM Escape
RevOps reality: 60%+ on admin, not revenue. Escape path: 12-18 months to sunset.
Ch 9 — Industry Vertical Trap
"Built for your industry" = "Built for the average of your industry."
Ch 10 — Solo Operator Advantage
No platform debt to escape. Full leverage from dollar one.
What Actually Changes
| Before | After |
|---|---|
| Platform administrator | Spec writer and reviewer |
| Configuration archaeologist | Requirements clarifier |
| Vendor relationship manager | AI orchestrator |
| Firefighter | Strategist |
Cost Structure Changes
Before
Platform license: $200K/year1
Consultant retainer: $100K/year3
Admin FTEs: $300K/year5
Integration maintenance: $80K/year
Total: ~$700K/year
After
AI tooling: $36K-100K/year55
Minimal consultant needs
Spec maintenance: Fraction of time
API connections in spec
Total: ~$100-200K/year
From funding vendor's moat → To funding your AI leverage
The Compound Advantage
The advantage grows over time. Model improvements and spec improvements both compound in your favour.
Platform-First (5-Year View)
- • Costs compound upward (10-15% annually)6
- • Debt compounds (harder to change)17
- • No AI leverage gained
- • Eventual "we need to migrate" conversation
Spec-Driven (5-Year View)
- • Costs trend down (AI gets cheaper)56
- • Knowledge compounds (spec gets better)
- • AI leverage increases (models improve)41
- • System is cleaner than year 1
The Questions You Can Now Answer
"Should we build or buy?"
Old answer: "It depends on cost and time"
"Where does your dollar get AI leverage?" If AI can help 80%+ → Build. If customisation >50% → Build. If Tier 1 commodity → Buy.
"How do we reduce platform spend?"
Old answer: "Negotiate harder with vendors"
"Extract your spec and regenerate." Sunk cost = Captured requirements. Export, translate, regenerate. Sunset the platform.
"Why are our AI initiatives failing?"
Old answer: "We need better prompts/models/training"
"You're investing in places AI can't help." Platform configuration = No leverage. The platform IS the blocker.
"How long until we see ROI?"
Old answer: "After this implementation project..."
"Your first overnight sprint." First spec → First regeneration → First morning PR. ROI visible in days, not years.
What to Do Next
This Week
- Run the Leverage Test on one platform
- Calculate total cost (license + consultants + admins + integration)
- Calculate customisation burden
- If burden >50% → It's a flip zone candidate
This Month
- Extract spec from one workflow
- Document what it actually does
- Translate to platform-agnostic requirements
- Run one overnight sprint
- Experience the time compression
This Quarter
- Pilot the escape
- Parallel run: Keep platform, build spec-driven version
- Prove feature parity on one module
- Demonstrate overnight sprint velocity
- Build confidence for full transition
What's at Stake
If You Stay Platform-First
- • Every year, costs rise 10-15%7
- • Every year, debt compounds18
- • Every year, AI capability improves for others, not you
- • The gap with spec-driven competitors widens
If You Go Spec-Driven
- • Every quarter, models improve your code for free
- • Every month, your spec captures more judgment
- • Every day, you have leverage others don't
- • The compound advantage accelerates
Three Years From Now
- • Your morning inbox has PRs ready for review57
- • Your team writes specs, not configuration
- • Your costs have dropped 50-70%
- • Your systems improve automatically with each model upgrade
- • You have time for strategy, not archaeology
"Stop funding your vendor's moat.
Start funding your AI leverage."
The platform trap is real, but it's not permanent.
The escape path exists.
The leverage is waiting.
The Complete Summary
- 1. The morning inbox: PRs ready for review, not fires to fight
- 2. The doctrine: Stupidity Tax → Three Tiers → Leverage Test → Sunk Cost as Spec → Overnight Sprints → Regeneration Economics
- 3. What changes: Staff become strategists, cadence becomes overnight, costs drop 50-70%
- 4. The compound advantage: Model improvements are free; spec improvements compound
- 5. The question answered: Buy vs build → "Where's the leverage?"
- 6. The call to action: Leverage Test this week, spec extraction this month, pilot this quarter
- 7. The stakes: Stay → Gap widens. Go → Compound advantage.
For all sources, research, and frameworks referenced throughout this ebook, see the References chapter.
References & Sources
The evidence, research, and frameworks that inform this analysis.
This ebook synthesises primary research from consulting firms, industry analysts, and technology leaders with practical frameworks developed through enterprise AI transformation consulting. All statistics and claims are sourced to their original publications.
Numbered Citations
Citations referenced inline throughout the chapters. Numbers correspond to N tags in the text.
[1] Salesforce Implementation Cost 2026
Source: Codleo
Implementation costs range $10K-$200K+ depending on complexity. Evidence that platform licenses are just the starting point.
https://www.codleo.com/blog/salesforce-implementation-cost
[2] Salesforce Implementation Cost Analysis 2025
Source: Techcronus
"Many projects that started at $50K ended up at $150K or more." Evidence of implementation cost unpredictability and escalation.
https://www.techcronus.com/blog/salesforce-implementation-cost-analysis/
[3] Salesforce Implementation Cost Analysis 2025
Source: Ascendix
Consultant hourly rates: Junior $60-100/hr, mid-level $100-150/hr, senior $150-250/hr. The cost structure of the consultant dependency.
https://ascendix.com/blog/salesforce-implementation-cost/
[4] Salesforce Implementation Cost 2025
Source: LinkedIn/SaaSGuruHQ
Consulting firm rates: $90-$300/hour for end-to-end implementation and support. Professional services markup on platform costs.
https://www.linkedin.com/pulse/how-much-does-salesforce-implementation-really-cost-2025-saasguruhq-iuiyf
[5] How Much to Hire Salesforce Admin
Source: Ascendix
Admin salaries: $80K-$130K+ full-time, $80-150/hr freelance. The ongoing cost of "institutional archaeology."
https://ascendix.com/blog/hire-salesforce-admin/
[6] SaaS Inflation Index 2026
Source: Vertice
SaaS inflation running nearly 5x higher than G7 general inflation. Key economic data showing the platform cost trajectory.
https://www.vertice.one/l/saas-inflation-index-report
[7] The Great Price Surge of 2025
Source: SaaStr
SaaS pricing up 11.4% vs 2024 compared to 2.7% average G7 market inflation. Industry confirmation of aggressive B2B price increases.
https://www.saastr.com/the-great-price-surge-of-2025-a-comprehensive-breakdown-of-pricing-increases-and-the-issues-they-have-created-for-all-of-us/
[8] SaaS Inflation Index 2026
Source: Vertice
60% of vendors deliberately mask rising prices through unbundling, stricter limits, and feature degradation. Making cost clarity in negotiations difficult.
https://www.vertice.one/l/saas-inflation-index-report
[9] Spec-Driven Development Agile Methodology
Source: Corestory
When AI can generate implementations from specifications in minutes rather than months, the cost of iteration approaches zero, enabling true responsiveness to changing requirements.
https://corestory.ai/post/specification-driven-development-agile-methodology-reducing-iteration-cost-in-ai-assisted-software-engineering
[10] Stop CRM Technical Debt
Source: ASC Digital
Platform debt symptoms include proliferating custom fields, conflicting automations, manual spreadsheet workarounds, brittle integrations, and slow admin cycles for simple requests.
https://www.asc.digital/blog/asc-blog-1/stop-crm-technical-debt-before-it-starts-45
[11] How Much to Hire Salesforce Admin
Source: Ascendix
Full-time admin salaries range $80K-$130K+, freelance rates $80-150/hr. The ongoing cost of maintaining platform configuration.
https://ascendix.com/blog/hire-salesforce-admin/
[12] Salesforce Consultant Rates 2025
Source: Ascendix
Consultant hourly rates: Junior $60-100/hr, mid-level $100-150/hr, senior $150-250/hr. The cost structure of consultant dependency.
https://ascendix.com/blog/salesforce-implementation-cost/
[13] Salesforce Implementation Cost 2026
Source: Codleo
Implementation costs range $10K-$200K+ with hidden fees doubling total spend. Total ownership typically 3-5x the license fee.
https://www.codleo.com/blog/salesforce-implementation-cost
[14] Andrej Karpathy on Agentic Coding
Source: LinkedIn
"I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December." The paradigm shift in AI-assisted development.
https://www.linkedin.com/posts/karpathy_i-rapidly-went-from-about-80-manualautocomplete-activity-7285394366070665216-Wqzs
[15] Why Custom AI Solutions Outperform Off-the-Shelf
Source: Medium
Generic platforms built for averages lack industry-specific depth. Core platforms remain 60-70% generic with vertical features as "skin."
https://medium.com/@dejanmarkovic_53716/why-custom-ai-solutions-outperform-off-the-shelf-options-0b9463b9febc
[16] The Cognition Ladder Framework
Source: LeverageAI / Scott Farrell
AI value creation framework mapping real-time vs batch processing. Rung 1 (15-30% success), Rung 2-3 (60-80+ success) guide build vs buy decisions.
https://leverageai.com.au//
[17] McKinsey Tech Debt IT Budget Study
Source: McKinsey Digital
Technical debt consumes 25-40% of IT budgets. Evidence that platform debt diverts massive resources from value creation.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/ai-for-it-modernization-faster-cheaper-and-better
[18] US Technical Debt Cost Analysis
Source: American Enterprise Institute
Technical debt costs US businesses $2.41 trillion annually. The compounding cost of deferred maintenance and patched systems.
https://www.aei.org/technology-and-innovation/inside-techs-2-trillion-technical-debt/
[19] Bug Fix Cost Multiplier Study
Source: Software Improvement Group
Bugs cost $100 to fix in planning stage but escalate to $10,000 in production. The exponential cost of deferred quality.
https://www.softwareimprovementgroup.com/technical-debt-and-it-budgets/
[20] Andrej Karpathy on Agentic Coding
Source: Andrej Karpathy / LinkedIn
"I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in December." The paradigm shift in AI-assisted development capabilities.
https://www.linkedin.com/posts/karpathy_i-rapidly-went-from-about-80-manualautocomplete-activity-7285394366070665216-Wqzs
[21] Developer Productivity Statistics 2026
Source: Index.dev
41% of all code written in 2025 is AI-generated. Evidence of widespread adoption and the paradigm shift in software development.
https://www.index.dev/blog/developer-productivity-statistics-with-ai-tools
[22] Developer Productivity Statistics 2026
Source: Index.dev
84% of developers use or plan to use AI tools in 2025, up from 76% last year. Demonstrates rapid mainstream adoption.
https://www.index.dev/blog/developer-productivity-statistics-with-ai-tools
[23] Capturing AI Potential in TMT
Source: McKinsey
Software developers using Microsoft's GitHub Copilot completed tasks 56 percent faster than those not using the tool. Measured productivity gains from AI assistance.
https://www.mckinsey.com/~/media/mckinsey/industries/technology%20media%20and%20telecommunications/high%20tech/our%20insights/beyond%20the%20hype%20capturing%20the%20potential%20of%20ai%20and%20gen%20ai%20in%20tmt/beyond-the-hype-capturing-the-potential-of-ai-and-gen-ai-in-tmt.pdf
[24] The Vibe Coding Explosion
Source: Law + Koffee
Searches for 'vibe coding' reportedly jumped 6,700% in spring 2025, signaling active experimentation across industries. Named Collins Dictionary Word of the Year 2025.
https://lawandkoffee.substack.com/p/the-vibe-coding-explosion-how-programming
[25] You're Falling Behind
Source: Theo Browne
Elite developers like Theo Brown run 6 parallel Claude Code instances, generating 11,900 lines without opening an IDE. Evidence of elite developer adoption patterns.
https://www.youtube.com/watch?v=Z9UxjmNF7b0
[26] My Predictions for 2026 in AI
Source: Lumberjack
By March 2025, 25% of Y Combinator's Winter 2025 batch had codebases that were 95% AI-generated. Demonstrates production-scale adoption in startups.
https://lumberjack.so/my-predictions-for-2026-in-ai/
[27] Agentic Workflows in The Batch
Source: Andrew Ng / DeepLearning.AI
GPT-3.5 with agentic workflows achieves 95% accuracy on HumanEval, while GPT-4 alone scores 48%. Evidence that workflow architecture multiplies AI capability.
https://www.deeplearning.ai/the-batch/
[28] When AI Writes Almost All Code
Source: Malte Ubl (CTO of Vercel) / Pragmatic Engineer
"Opus + Claude Code now behaves like a senior software engineer... The cost of software production is trending towards zero." Evidence from production engineering leadership.
https://newsletter.pragmaticengineer.com/p/when-ai-writes-almost-all-code-what
[29] Martin Fowler: Understanding Spec-Driven-Development
Source: Martin Fowler
The spec becomes the source of truth for the human and the AI. AI excels at structure-to-structure transformation, making spec extraction from platform configuration an ideal use case.
https://martinfowler.com/articles/exploring-gen-ai/sdd-3-tools.html
[30] ThoughtWorks Technology Radar Vol 33
Source: ThoughtWorks
Tessl Framework takes a more radical approach in which the specification itself becomes the maintained artifact, rather than the code. Existing behaviour can serve as acceptance criteria for regenerated systems.
https://www.thoughtworks.com/content/dam/thoughtworks/documents/radar/2025/11/tr_technology_radar_vol_33_en.pdf
[31] Spec-Driven Development with AI
Source: GitHub Engineering
Specs should live longer than the code. Code becomes a by-product of well-written specifications. Platform-agnostic specs are AI-consumable for regeneration and human-readable for review.
https://github.blog/ai-and-ml/generative-ai/spec-driven-development-with-ai-get-started-with-a-new-open-source-toolkit/
[32] Real-Time vs Batch Processing Architecture
Source: Zen van Riel
Batch AI processing provides 40-60% cost reduction versus real-time with improved depth of analysis. No latency constraints means time for thorough retrieval and reasoning.
https://zenvanriel.nl/ai-engineer-blog/should-i-use-real-time-or-batch-processing-for-ai-complete-guide/
[33] Progressive Resolution Diffusion Architecture
Source: Scott Farrell / LeverageAI
Coarse-to-fine development pattern enabling requirement-to-deployed-code overnight. Work through resolution layers stabilizing structure before detail.
https://leverageai.com.au//progressive-resolution-diffusion-architecture/
[34] You're Falling Behind
Source: Theo Browne
Elite developers like Theo Brown run 6 parallel Claude Code instances in production workflows, generating 11,900 lines without opening an IDE. Evidence of elite developer adoption patterns.
https://www.youtube.com/watch?v=Z9UxjmNF7b0
[35] Nightly AI Decision Builds
Source: Scott Farrell / LeverageAI
CI/CD discipline for AI development including regression tests, canaries, rollback, and diff reports. Design-time AI produces reviewable artifacts before deployment.
https://leverageai.com.au//nightly-ai-decision-builds-backed-by-software-engineering-practice/
[36] Test-Time Compute Analysis
Source: Hugging Face Research
Small model with thinking time can outperform a model 14× larger with instant response, demonstrating why overnight batch processing unlocks AI capability. Extended reasoning improves performance.
https://huggingface.co/blog/Kseniase/testtimecompute
[37] How the Creator of Claude Code Actually Uses It
Source: Boris Cherny / Dev.to
Verification loops improve quality by 2-3x. Second agent reviews first agent's work. Human reviews and approves before deployment. Evidence that adversarial checking is non-negotiable.
https://dev.to/a_shokn/how-the-creator-of-claude-code-actually-uses-it-32df
[38] DevOps Engineering in 2026
Source: Refonte Learning
Tech giants like Amazon deploy code thousands of times per day using CI/CD automation. Canary deployments at 1% enable early detection of issues with instant rollback capability.
https://www.refontelearning.com/blog/devops-engineering-in-2026-essential-trends-tools-and-career-strategies
[39] Stop Nursing Your AI Outputs
Source: Scott Farrell / LeverageAI
Nuke and Regenerate economics framework explaining why regeneration beats patching. After ~5 patches, regeneration almost always wins economically.
https://leverageai.com.au//wp-content/media/Stop%20Nursing%20Your%20AI%20Outputs.html
[40] Why AI-Generated Code Costs More to Maintain
Source: AlterSquare
Patching AI outputs accumulates complexity faster than human-written code. Each patch adds complexity making future patches harder—the compound effect.
https://www.altersquare.io/ai-generated-code-maintenance-costs/
[41] Claude Sonnet 4.5 Release
Source: Anthropic
44% faster processing, 25% improved accuracy, at same price. Demonstrates how model improvements benefit spec-driven regeneration automatically—better code at no extra cost.
https://www.anthropic.com/news/claude-sonnet-4-5
[42] AI Generated Code: The Iron Triangle in 2025
Source: AskFlux
66% of developers spend more time fixing 'almost-right' AI code than they save in initial writing. Evidence that output nursing is counterproductive—the anti-pattern of patching without spec improvement.
https://askflux.ai/ai-generated-code-iron-triangle-2025
[43] Understanding Spec-Driven Development
Source: Martin Fowler
The spec becomes the source of truth for both humans and AI. Code is a rendering of the spec. Explains why spec-driven code passes the delete test while platform config doesn't.
https://martinfowler.com/articles/exploring-gen-ai/sdd-3-tools.html
[44] Software Engineering + AI = Future
Source: Chris Poel / Medium
Weekly regeneration from specs is normal practice. PRs primarily change spec and validations while AI regenerates code. Knowledge flows into spec, not patches—the kernel-first discipline.
https://chrispoel.medium.com/software-engineering-ai-future-f39b3ddb4afa
[45] Open Source vs Proprietary LLMs 2025
Source: WhatLLM
Newer models are both better and cheaper. Spec-driven code benefits from model improvements automatically—quarterly compounding with model releases at no additional effort.
https://whatllm.org/blog/open-source-vs-proprietary-llms-2025
[46] Inside Tech's $2 Trillion Technical Debt
Source: American Enterprise Institute
Technical debt from patched outputs costs US businesses $2.41 trillion annually. Platform configuration compounds debt while spec-driven development compounds value—the opposite of leverage.
https://www.aei.org/technology-and-innovation/inside-techs-2-trillion-technical-debt/
[47] Revenue Grid: Salesforce Customization
Source: Revenue Grid
Sales reps spend 66% of time on non-selling activities (admin, data entry, navigation). Only 34% of time spent actually selling. Evidence of platform administration overhead consuming revenue-generating capacity.
https://revenuegrid.com/blog/enhance-your-experience-how-to-customize-salesforce-for-maximum-efficiency/
[48] SalesForge: AI SDRs and Real-Time Lead Data
Source: SalesForge
Human SDRs spend only 35% of time actively selling, with the rest consumed by research, data entry, and routine outreach. Evidence that CRM-heavy workflows bury sales teams in administration.
https://www.salesforge.ai/blog/ai-sdrs-real-time-lead-data-explained
[49] Why Custom AI Solutions Outperform Off-the-Shelf
Source: Medium
Generic platforms built for averages lack industry-specific depth. Core platforms remain 60-70% generic with vertical features as "skin." Evidence that industry vertical software is still largely generic.
https://medium.com/@dejanmarkovic_53716/why-custom-ai-solutions-outperform-off-the-shelf-options-0b9463b9febc
[50] The Cognition Ladder Framework
Source: LeverageAI / Scott Farrell
AI value creation framework mapping real-time vs batch processing. Rung 1 (15-30% success), Rung 2-3 (60-80+ success) guide build vs buy decisions. Shows why proprietary platforms have low AI leverage.
https://leverageai.com.au//
[51] MIT Sloan Review: The End of Scale
Source: MIT Sloan
Traditional size advantages inverting as AI technologies enable economies of unscale. For more than a century, economies of scale made the corporation ideal, but AI is turning economies of scale inside out.
https://sloanreview.mit.edu/article/the-end-of-scale/
[52] Solo Entrepreneurs Doing $1M+ Revenue
Source: Founderoo
Dan Koe case study: $4.2M revenue, 98% margin, zero employees. Evidence of solo operator economic viability at scale using AI-enabled workflows.
https://www.founderoo.co/resources/one-man-businesses-solo-entrepreneurs-doing-1m-yearly-revenue
[53] Communication Overhead
Source: Personal MBA
As team size increases, communication overhead grows geometrically until the total percentage of time each individual must devote to group communication approaches 100%. Evidence for coordination cost advantage of small teams.
https://personalmba.com/communication-overhead/
[54] The Team of One: Individuals Learn 10-100x Faster
Source: LeverageAI / Scott Farrell
Individuals learn 10-100x faster than organizations due to coordination overhead. Organizational learning loop: 3-4 months from proposal through rollout. Solo operators can iterate at AI speed.
https://leverageai.com.au//the-team-of-one-why-ai-enables-individuals-to-outpace-organizations/
[55] Claude AI Programming Tools Pricing Comparison
Source: Eesel.ai
AI tooling costs for solo operators: Claude Pro $20/month, Cursor Pro $20/month, API access $1K-5K/month. Total AI development capability: $36K-180K/year vs enterprise platform costs.
https://www.eesel.ai/blog/claude-ai-programming-tools
[56] Open Source vs Proprietary LLMs 2025
Source: WhatLLM
Newer models are both better and cheaper. Spec-driven code benefits from model improvements automatically—quarterly compounding with model releases at no additional effort.
https://whatllm.org/blog/open-source-vs-proprietary-llms-2025
[57] How the Creator of Claude Code Actually Uses It
Source: Boris Cherny / Dev.to
Verification loops improve quality by 2-3x. Second agent reviews first agent's work. Human reviews and approves before deployment. Evidence that adversarial checking is non-negotiable for overnight development patterns.
https://dev.to/a_shokn/how-the-creator-of-claude-code-actually-uses-it-32df
Primary Research
Technical Debt: $2 Trillion Problem
Technical debt from patched outputs costs US businesses $2.41 trillion annually — foundational data for the regeneration economics argument.
https://www.aei.org/technology-and-innovation/inside-techs-2-trillion-technical-debt/
McKinsey: AI for IT Modernization
Technical debt consumes 25-40% of IT budgets. Supports the argument that platform debt diverts resources from value creation.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/ai-for-it-modernization-faster-cheaper-and-better
SaaS Inflation Index 2026
SaaS prices rising 5x faster than general inflation. 60% of vendors deliberately mask rising prices. Key data for the platform economics argument.
https://www.vertice.one/l/saas-inflation-index-report
Software Improvement Group: Technical Debt Economics
Bugs fixed in planning cost $100; left as technical debt, they cost $10,000. Supports the "fix the kernel, regenerate the output" approach.
https://www.softwareimprovementgroup.com/technical-debt-and-it-budgets/
Implementation Cost Analysis
Salesforce Implementation Cost 2026
Implementation costs range $10K-$200K+ depending on complexity. Source for the "starting point, not solution" argument.
https://www.codleo.com/blog/salesforce-implementation-cost
Salesforce Implementation Cost Analysis 2025
"Many projects that started at $50K ended up at $150K or more." Evidence for implementation cost unpredictability.
https://www.techcronus.com/blog/salesforce-implementation-cost-analysis/
RT Dynamic: Salesforce Implementation Cost 2026
Costs range $5K-$100K+ based on complexity, user count, and rollout type. Covers hidden costs including sandboxes and AppExchange.
https://www.rtdynamic.com/blog/salesforce-implementation-cost-2026/
Ascendix: Salesforce Consultant Rates
Junior consultants $60-100/hr, mid-level $100-150/hr, senior $150-250/hr. Full-time admin salaries $80K-$130K+.
https://ascendix.com/blog/salesforce-implementation-cost/
Ascendix: How Much to Hire Salesforce Admin
Admin salary data: $80K-$130K+ full-time, $80-150/hr freelance. The cost of "institutional archaeology."
https://ascendix.com/blog/hire-salesforce-admin/
Industry Analysis
SaaStr: The Great Price Surge of 2025
2025 SaaS pricing up 11.4% vs 2024. Industry confirmation of the platform cost trajectory driving the build case.
https://www.saastr.com/the-great-price-surge-of-2025/
Clevyr: Salesforce vs HubSpot ROI Analysis
Training costs and steep learning curve as hidden expenses. Evidence for ongoing overhead burden.
https://clevyr.com/blog/post/how-much-does-salesforce-cost
ASC Digital: Stop CRM Technical Debt
Platform debt symptoms: proliferating custom fields, conflicting automations, manual workarounds, brittle integrations.
https://www.asc.digital/blog/asc-blog-1/stop-crm-technical-debt-before-it-starts-45
LinkedIn: Salesforce Implementation Cost 2025
Consulting firm rates: $90-$300/hour for end-to-end implementation and support.
https://www.linkedin.com/pulse/how-much-does-salesforce-implementation-really-cost-2025-saasguruhq-iuiyf
LinkedIn Commentary
Andrej Karpathy on Agentic Coding
"I rapidly went from about 80% manual/autocomplete to 80% agent in December." The paradigm shift defining the AI coding threshold.
https://www.linkedin.com/posts/karpathy_i-rapidly-went-from-about-80-manualautocomplete-activity-7285394366070665216-Wqzs
LeverageAI / Scott Farrell
Practitioner frameworks and interpretive analysis developed through enterprise AI transformation consulting. Listed for transparency — these frameworks inform the ebook's analysis without being cited inline.
The Cognition Ladder
AI value creation framework mapping real-time vs batch processing. Informs the Three Tiers categorisation.
https://leverageai.com.au//
Waterfall Per Increment
How agentic coding changes everything. Spec-driven development model underlying the Overnight Sprint pattern.
https://leverageai.com.au//waterfall-per-increment-how-agentic-coding-changes-everything/
Progressive Resolution Diffusion Architecture
Coarse-to-fine development pattern. The mechanism enabling requirement-to-deployed-code overnight.
https://leverageai.com.au//progressive-resolution-diffusion-architecture/
Nightly AI Decision Builds
CI/CD discipline for AI development. Governance framework for overnight sprint operations.
https://leverageai.com.au//nightly-ai-decision-builds-backed-by-software-engineering-practice/
Stop Nursing Your AI Outputs
Nuke and Regenerate economics. Why patching loses to regeneration after ~5 modifications.
https://leverageai.com.au//wp-content/media/Stop%20Nursing%20Your%20AI%20Outputs.html
Team of One Economics
Economies of specificity framework. Why solo operators and small teams have structural advantages in AI-enabled development.
https://leverageai.com.au//
Note on Research Methodology
This research was compiled in January 2026 using primary sources from consulting firms (McKinsey, implementation specialists), industry analysts (Vertice, SaaStr), and technology practitioners. All statistics are attributed to their original sources.
Implementation cost ranges represent market conditions as of early 2026 and may vary by region, complexity, and vendor relationship. SaaS pricing trends reflect publicly available index data.
Some linked resources may require registration or subscription access. Where possible, key data points are quoted directly in the text.