The Make-vs-Buy Inversion

Don't Buy Software.
Don't Hire Experts.
Build AI Instead.

What Karpathy and Theo proved individually has a radical organisational implication: the make-vs-buy calculus has inverted.

A Strategic Framework for Mid-Market Leaders Ready to Stop Renting Software and Start Building Competitive Advantage

What You'll Learn

✓ Why building is now often cheaper than buying (the economics have crossed over)
✓ The Cognition Ladder framework for deciding what to build vs buy
✓ How to audit your SaaS portfolio and encode expertise in AI systems
✓ The minimal capability required to execute this strategy (less than you think)

Scott Farrell · LeverageAI

January 2026

Part I: The Inversion

The 10x Signal

When the smartest developer alive admits he's falling behind, what does that mean for your organisation?

In December 2025, something remarkable happened. Andrej Karpathy—former OpenAI founding member, former Tesla AI chief, one of the most influential figures in modern artificial intelligence—posted a confession to social media that sent ripples through the technology world.

"I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. 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 and a failure to claim the boost feels decidedly like skill issue."

— Andrej Karpathy, December 2025¹

If he feels behind, what chance does a mid-market CFO have?

But this isn't a story about developers. This is a story about what Karpathy's confession reveals for organisations—and why every executive should be paying attention.

The Alien Tool With No Manual

Karpathy didn't stop at describing his own experience. He painted a vivid picture of what's happening across the entire profession:

"Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession."¹³

The "alien tool" metaphor is apt. AI-assisted development isn't a better hammer—it's a fundamentally different way of working. Like moving from assembly language to high-level programming, except the transition is happening in months, not decades.

The Numbers Behind the Signal

If Karpathy provides the conceptual frame, Theo Browne provides the operational proof. Creator of the T3 Stack and influential developer educator, Theo runs engineering teams at scale—not in theory, but in daily production.

His numbers are striking:

90%

of Theo's personal code is now AI-generated

70%+

AI-generated code in his production teams

Similar

numbers across companies he advises and invests in

"Coding has changed forever. We are past that point now. Getting into it now isn't getting into it early anymore. Getting into it now is getting into it late."

— Theo Browne, "You're Falling Behind"²

These aren't projections. They're not optimistic forecasts. They're current operational reality—documented, measured, and replicated across multiple teams and companies.

The Research Confirms It

Independent studies corroborate what Karpathy feels and Theo measures:

GitHub Copilot Impact Study

Developers completed tasks 55.8% faster with GitHub Copilot. Average completion time dropped from 2 hours 41 minutes to 1 hour 11 minutes.

Source: arXiv research study³

AMD Developer Productivity Study

AI-assisted development showed an 82% reduction in predefined application development time.

Source: AMD Technical Blog⁴

What This Really Means

Most people read the Karpathy post and think: "Developers need to learn AI tools."

That misses the point entirely.

The real insight is what these individual productivity gains mean at the organisational level. If one developer with AI tools can do what a team did before, then every assumption about what to buy, what to build, and who to hire needs to be revisited.

Real Companies Are Already Acting

This isn't just individual developers experimenting. Real companies at scale are building AI capability into their operations.

Raul, head of applied AI at Ramp—a major financial services company—responded to Karpathy's post with a concrete playbook for organisations:

"You are guaranteed to lose if you fall behind. Give all engineers their pick of harnesses, models, background agents. Claude Code, Cursor, Devin, with closed/open models... If agents are being held back because of lack of context, that's your fault."

— Raul, Head of Applied AI at Ramp⁵

Ramp isn't theorising. Their internal "inspect bot" can identify the 20 most common errors in their codebase and spin up pull requests to fix each one—automatically.⁵ Not a product they're selling; just internal tooling they built because, as Raul notes, "doing this is now way more justifiable than ever."

What About "My Work Is Too Complex for AI"?

Theo addresses this objection directly:

"Tell it to the people building compilers with it. Tell it to the people who are building languages and systems and crazy applications with it. Tell it to the CEO of Railway who's rebuilding their deployment system with it. Tell it to Karpathy who's one of the smartest developers that ever lived that feels like he's falling behind."

Key Takeaways

1 The smartest developers feel behind — if Karpathy admits this, organisations can't pretend the shift isn't real
2 70-90% AI-generated code is operational reality — not future projection, current practice
3 The individual proof reveals the organisational strategy — if developers are 10x more powerful, organisations should rethink what they buy and who they hire
4 "Don't buy software, don't hire experts" — the sharp framing for what comes next

Chapter 1 established the proof: AI has made building dramatically cheaper and faster. Chapter 2 will show the other side: the costs of buying are escalating. Together, they demonstrate the "inversion"—the point where building beats buying.

Part I: The Inversion

The Economics Inversion

Build costs are collapsing. SaaS costs are compounding. The lines have crossed.

Picture a mid-market CFO reviewing annual software costs. Three years ago: $120K in SaaS subscriptions. Today: $200K—same tools, no meaningful new features. And on the desk: an email from the vendor announcing an 18% price increase "to reflect market conditions."

The question every CFO should be asking: "Are we getting 18% more value? Or are we just trapped?"

The Build Side: Development Costs Have Collapsed

The productivity gains from Chapter 1 translate directly into cost reductions. When developers complete tasks 55-82% faster,^3,4 custom development economics fundamentally change.

Custom Development Cost Trajectory

5 Years Ago

$500K

for meaningful custom AI capability

Today

$50K-$150K

for equivalent capability¹⁴

70-90% cost reduction in custom development¹⁴

The math is straightforward: developers are 2-5x more productive with AI tools.¹⁴ What took 12 months now takes 3-6 months. What took a team of 5 can be done by 2.

The Buy Side: SaaS Costs Are Escalating

While build costs are dropping, buy costs are moving in the opposite direction—and faster than most organisations realise.

11.4%

SaaS pricing increase YoY (2025 vs 2024)
vs 2.7% general inflation

Source: SaaStr⁷

SaaS costs are growing at 4.2x the rate of general inflation.⁷ But the headline number understates the reality:

→ Aggressive vendor hikes: 15-25% for major CRM and ERP vendors leveraging lock-in¹⁵
→ PE-backed extremes: Some private equity-owned vendors pushing increases as high as 900%¹⁰
→ Hidden inflation: 20-30% additional costs disguised through unbundling features, stricter API limits, and new "AI taxes"¹⁵

"For many companies, 50%+ of growth now comes from price increases, not expansion or new logos."

— SaaStr Analysis⁸

The Renewal Discount Trap

The most insidious pattern is what happens to your "first-year discount" over time:

Year	List Price (Index)	Effective Discount	What You Pay
Year 1	100	40%	60
Year 3	108	28%	78
Year 5	117	18%	96
Year 7	126	10%	113

Net effect: 88% price increase over 7 years while the list price only rose 26%. Your "exceptional first-year discount" becomes a trap.⁹

The Lock-In Psychology

"Vendor lock-in 'is driving a lot of significant increases, because organisations don't feel they can do anything.'"

— Chris Sawchuk, The Hackett Group¹⁰

The waste compounds the problem:

49%

of SaaS licenses go unused

$21M

wasted annually by companies with 1,000+ employees

Source: The State of Software Costs in 2025¹¹

The Crossover Point

The mathematics have changed. Here's what the shift looks like:

Build vs Buy: The Math Has Changed

Old Equation (Pre-2023)

• Building: $500K dev + $50K/yr maintenance
• Buying: $50K/yr, stable pricing
• 5-year TCO: Build = $750K, Buy = $250K
Clear winner: Buy

New Equation (2025+)

• Building: $75K dev + $10K/yr infrastructure
• Buying: $50K Year 1, growing 15% annually
• 5-year TCO: Build = $125K, Buy = $337K+
Clear winner: Build (for many use cases)

What changed? Three factors converged:

1. Development costs collapsed — AI-assisted development is 2-5x faster
2. Maintenance costs collapsed — "spec-as-asset" means regeneration is cheaper than patching
3. SaaS costs accelerated — PE ownership, lock-in leverage, and "AI taxes"

The lines have crossed. Building is now often cheaper than buying.

Key Takeaways

1 Build costs have collapsed — what cost $500K now costs $50K-$150K
2 SaaS costs are compounding — 11.4% annually vs 2.7% inflation, with hidden costs on top
3 The crossover point has been reached — building beats buying for many categories
4 Your discount is decaying — Year 1's 40% becomes Year 7's 10%
5 The spec is the asset — regeneration beats maintenance in the AI era

Chapter 2 established the economics: building is often cheaper than buying. But not for everything—you need a framework to decide WHERE to build. Chapter 3 introduces the Cognition Ladder: the decision tool for build vs buy.

Part I: The Inversion

The Cognition Ladder

A framework for deciding where AI wins—and where you should still buy.

An executive watches their AI chatbot fail in real-time. A customer asks a nuanced question; the bot gives a generic response. The customer escalates, frustrated. The executive's question: "We spent $200K on this. Why doesn't it work?"

The answer: they deployed AI where it's weakest.

Why 70-85% of AI Projects Fail

70-85%

AI project failure rate when deployed in real-time, high-stakes contexts

Most AI projects start with the same idea: "Let's put AI in front of customers." Chatbots. Live support. Real-time recommendations. These are the hardest contexts for AI to succeed¹⁶—yet companies start there.

Why real-time is hard:

• No time for error correction
• High stakes (customer-facing)
• One-shot interaction (no iteration)
• Humans have optimised for these contexts over millennia

The Cognition Ladder Framework

AI value comes in three distinct rungs. Each has different time scales, success rates, and economics. Most organisations fail because they start at Rung 1.

Rung	Name	Time Scale	Success Rate	Build vs Buy
1	Don't Compete	Seconds (real-time)	15-30%	Still buy
2	Augment	Hours (batch)	60-80%	BUILD HERE
3	Transcend	Overnight	High (new capability)	THE OPPORTUNITY

Rung 1: Don't Compete (Real-Time)

Live, high-stakes, one-shot interactions. Customer service chat, live sales calls, immediate decision-making.

Why AI fails here: "In low-latency, high-stakes, high-ambiguity contexts, humans still win on accuracy, judgement, and social handling."¹⁷

Decision: For Rung 1 contexts, still favour buying. SaaS vendors have optimised for real-time.

Rung 2: Augment (Batch)

Batch processing with latency tolerance. Ticket queues, overnight analysis, document processing—anywhere time flexibility exists.

Cost savings: 40-60% vs real-time¹⁸

ROI: $3.70-$10.30 per dollar invested¹⁹

Decision: This is where you start building instead of buying.

Rung 3: Transcend (Overnight)

Work that was never feasible before. Not automating existing processes—enabling new capabilities.

"What becomes rational when cognition is abundant?"

Decision: Build. This is what SaaS vendors will never deliver.

What Rung 3 Makes Possible

Old Constraint	New Possibility
Strategic projects need 10 people × 3 months	Hyper sprints: thousands of AI calls overnight, human review in morning
Customer segmentation too expensive to personalise	Marketplace of One: per-customer offers, service levels, risk rules
Committee-think under time pressure	Systematic exploration of more possibilities than humans could attempt

Applying the Ladder

The Decision Framework

Step 1: Identify the rung

What time scale does this need operate on?

Step 2: Apply the default

Rung 1: Buy · Rung 2: Question buying · Rung 3: Build

Step 3: Consider specificity

Does this need match vendor's average use case? → Maybe buy
Does it require your specific workflow/data? → Build

"Version 3 Thinking"

The mental shift that separates organisations getting 5-10% gains from those getting 10x returns:

Version 1

"How do we make this process faster?"

Version 2

"How do we apply more cognition to this process?"

Version 3

"What becomes rational when cognition is abundant?"

Key Takeaways

1 70-85% of AI projects fail — because they deploy AI where it's weakest (Rung 1)
2 Rung 2 is the sweet spot — batch contexts, latency tolerance, 40-60% cost savings
3 Rung 3 is the opportunity — capabilities that were never feasible, what vendors can't build
4 Ask the right question — not "where to automate" but "what becomes rational when cognition is abundant"

Chapter 3 established the framework. Chapters 4-5 will apply it in detail: a flagship example of building instead of buying, and encoding expertise instead of hiring.

Part II: Building the Replacement

Anatomy of a Build Decision

A complete walkthrough: from SaaS renewal shock to owned capability.

The email from the vendor lands in the CIO's inbox. Subject: "Important Notice: Pricing Update for Your Renewal." Current spend: $50K/year for CRM enrichment automation. Proposed new price: $75K/year—a 50% increase.⁷ The vendor's explanation: "To better serve you with our enhanced AI capabilities..."

The CIO's reaction: "We've been asking for custom scoring logic for 18 months. They gave us price increases instead."

The Scenario: CRM Enrichment Tool

What It Does

• Enriches CRM contacts with company data, social profiles
• Processes overnight (batch)
• Feeds sales team prioritised leads each morning
• Used for 3 years

The Pain Points

• Price: $35K → $45K → $50K → now $75K¹⁵
• Features: Custom logic request ignored 18 months
• Lock-in: Vendor-specific export format¹⁰
• Waste: Paying for 10K enrichments, using ~4K¹¹

Applying the Cognition Ladder

Step 1: Identify the Rung

CRM enrichment runs overnight with no real-time requirement.

Verdict: Rung 2 (Augment) — batch processing with latency tolerance¹⁸

Step 2: Apply the Default

Per the Cognition Ladder: Rung 2 = "Question buying, likely build"

Step 3: Consider Specificity

Need custom ICP scoring, proprietary intent signals, different logic per sales team.

High specificity — vendor's average use case doesn't serve us.

Decision: BUILD

The Build Analysis

Old World (Pre-AI)

• 2 senior developers × 3 months
• Development cost: ~$150K
• Ongoing maintenance: $30K/year
5-year TCO: $300K

New World (AI-Assisted)

• 1 developer × 3 weeks¹⁴
• Development cost: ~$15K
• Infrastructure: $3K/year
5-year TCO: $30K

The 5-Year TCO Comparison

Continue with Vendor
Year 1	$75,000	Cumulative: $75,000
Year 3	$90,750 (+10%/yr)⁷	Cumulative: $248,250
Year 5	$109,808	Total: $457,883

Build and Own
Initial development	$15,000
5-year infrastructure	$15,000
Year 3 refresh	$8,000
Total 5-Year TCO	$38,000

$419,883

saved over 5 years

What You Gain Beyond Cost

Custom Logic

Scoring algorithm reflects YOUR ICP, not vendor's generic assumptions. Different logic for different sales teams.

Data Ownership

Your format, your systems. No export fees, no migration projects, full audit trail.

Controlled Upgrades

Model improvements incorporated on YOUR timeline. Not locked into vendor's roadmap.

Spec-as-Asset

Specification is durable. Code regenerates as AI improves.¹⁴ No technical debt in traditional sense.

Addressing the Objections

"What about maintenance?"

Maintenance = regeneration in AI-era development. Bug fix: modify spec, regenerate. Model upgrade: regenerate everything—often BETTER than before.

"We don't have developers"

One developer with AI = what a team did before.³ See Chapter 7 for minimal capability requirements.

"This seems risky"

Risk of building: project takes longer. Risk of NOT building: vendor increases prices indefinitely, custom needs never addressed. Use parallel run pattern to de-risk.

Key Takeaways

1 Apply the Cognition Ladder first — CRM enrichment is Rung 2, so "question buying, likely build"
2 Calculate real 5-year TCO — include vendor price increases, not just current price
3 Value specificity — vendor's average use case isn't your competitive advantage
4 De-risk with parallel run — don't hard-switch, build confidence through comparison

Chapter 4 showed building instead of BUYING (software replacement). Chapter 5 shows the other half: building instead of HIRING (expertise encoding).

Part II: Building the Replacement

The Expertise Encoding Pattern

When "don't hire experts" makes sense—and when it doesn't.

The HR dashboard shows an open position: Compliance Specialist, ISO 27001 / SOC 2. Salary range: $160K-$190K. Status: Open for 9 months. Candidates interviewed: 23. Offers extended: 2 (both declined for higher offers elsewhere).

Meanwhile, compliance gaps are blocking two major deals.

The COO's question: "Do we need a person? Or do we need the knowledge applied consistently?"

The "Don't Hire Experts" Insight

Scott Farrell's distillation: "Don't hire experts—write software for it."²⁵

But where does this apply?

Rung 1 (Real-Time Judgement): High-stakes decisions under time pressure, novel situations, political navigation → Human expertise still critical^16,17
Rung 2 (Batch Application): Applying known frameworks to recurring situations, reviewing against established criteria → Encode in AI systems¹⁸
Rung 3 (Scale): Check EVERY document, continuous monitoring, personalised guidance for every employee → Build new capabilities^20,21

Traditional Expert Model vs AI-Encoded Model

Traditional Model

1. Identify need for expertise
2. Write job description
3. Recruit for 6-12 months
4. Pay $150K-$250K salary + benefits
5. Expert applies knowledge repeatedly
6. Expert leaves → back to step 2

AI-Encoded Model

1. Identify need for expertise
2. Engage expert to ENCODE knowledge
3. Build AI system that applies knowledge
4. Expert reviews edge cases
5. System scales indefinitely
6. Expert improves system over time

The Economics

Model	Year 1	Years 2-5	5-Year Total
Traditional Hire	$250K	$1M (4 × $250K)	$1.25M
Encoded + Review¹⁴	$100K build + $50K review	$200K ($50K/yr)	$300K
Savings			$950K

Where This Pattern Applies

Compliance Review

Traditional: Officer reviews documents manually, samples 10% of contracts, catches issues after the fact. $180K/year.

Encoded: EVERY document reviewed automatically, issues flagged before signing, expert handles edge cases. $30K build + $20K/year.¹⁴

Result: 10x coverage at 70% lower cost

Research Synthesis

Traditional: Analyst reads sources manually, summarises 50-100 per project. $120K/year.

Encoded: EVERY relevant source processed, analyst reviews synthesis. $20K build + $30K/year.¹⁴

Result: Comprehensive coverage, faster delivery

Training Delivery

Traditional: Expert delivers training, limited by calendar, inconsistent quality. $150K/year + travel.

Encoded: On-demand delivery, consistent quality, personalised pace, expert updates content. $40K build + $30K/year.¹⁴

Result: Unlimited scale, consistent quality

"AI knows the expert advice. Your job is to work out what is valuable, dream up what a better replacement looks like, and run it through AI."²⁵

— Scott Farrell

When NOT to Encode Expertise

Don't Encode When:

• Rung 1 dominant: Work is primarily real-time judgement calls
• Rapidly changing: Knowledge changes faster than you can encode
• Relationship-dependent: Value is in expert's relationships, not knowledge
• Truly novel: Each situation is genuinely unique

Do Encode When:

• Rung 2-3 dominant: Applying known frameworks at scale
• Stable knowledge: Rules/criteria change slowly
• Volume problem: More instances than experts can handle
• Consistency matters: Variation is a bug, not a feature

The Expert's Role Changes

Experts don't disappear. Their role shifts:

Before: 80% Application

Expert spends most time applying knowledge repeatedly to similar situations

After: 80% Improvement

Expert spends most time on high-value Rung 1 work and system improvement

Key Takeaways

1 "Don't hire experts" needs the Cognition Ladder — Rung 1 still needs humans; Rungs 2-3 can be encoded
2 The traditional model is expensive and fragile — $200K+ salaries, 9-month searches, knowledge walks out
3 The encoded model scales — one expert encodes, system applies at unlimited scale
4 ROI is typically 6-12 months²⁴ — and the system improves forever

Chapters 4-5 showed the flagship examples. Part III applies the same framework to your existing portfolio and capability building.

Part III: Applying the Framework

The SaaS Portfolio Audit

Systematically apply the Cognition Ladder to your existing software spend.

The CFO reviews the annual software spend report. Total SaaS: $420K/year across 47 tools. Finance's estimate of utilisation: "Hard to measure." IT's estimate: "Most tools are well-used." Industry reality: 49% of licenses go unused¹¹.

The realisation: "We're probably wasting $200K and don't even know on what."

The Portfolio Problem

49%¹¹

of SaaS licenses go unused

$21M¹¹

wasted annually by 1,000+ employee companies

SaaS accumulates like sediment:

• Year 1: Marketing buys HubSpot ($24K)
• Year 2: Sales adds Salesforce ($36K)
• Year 3: Ops adds Monday.com ($18K)
• Year 4: HR adds Workday ($45K)
• Year 5: Each tool has annual increases, new tools added
• Year 6: $420K total, 47 tools, no one has the full picture

Each tool has its own renewal cycle. Pricing increases happen individually. Nobody owns the "should we keep this?" question.

The Cognition Ladder Audit

For each SaaS tool in your portfolio:

Identify the Rung

Apply Default

Calculate 5-yr TCO

Assess Specificity

Decide

Category-by-Category Analysis

Rung 1: Keep Buying

Core Infrastructure

Cloud hosting, CDN, databases

Real-Time Comms

Slack, Zoom, phone systems

Commoditised

Payroll, expense, basic accounting

Rung 2: Question Buying

Reporting & Analytics

BI dashboards, custom reporting

CRM Enrichment

Lead scoring, data enrichment

Document Processing

Contract analysis, extraction

Ask: Could we build this in batch context at lower 5-year TCO?

Rung 3: Build Instead

Competitive Differentiation

Custom analytics no vendor offers

Deep Customisation

Workflows vendor won't build

Vendor Won't Build

Too niche for their market

These create competitive moats vendors can't deliver.

The Audit Checklist

For Each SaaS Tool:

Basic Information

☐ Annual cost (current)
☐ Renewal date
☐ Contract terms (auto-renew, escalation)
☐ Number of licenses/seats
☐ Actual utilisation

Decision Criteria

☐ Rung assigned (1/2/3)
☐ 5-year TCO projected
☐ Build alternative estimated
☐ Specificity assessed
☐ Decision: Keep / Replace / Build

Using Build Alternative as Leverage

Before

Vendor: "15% increase"

Customer: "That's a lot"

Vendor: "Market rates"

Customer: "Okay"

With Build Alternative

Vendor: "15% increase"

Customer: "We've evaluated building. Here's the 5-year TCO comparison."

Vendor: "Let me talk to my manager about a loyalty discount"

Key: You need to actually be willing to build. Empty threats don't work.

Key Takeaways

1 Portfolio view is essential — individual tools compound into massive spend no one owns
2 Apply Cognition Ladder systematically — Rung 1 (keep), Rung 2 (question), Rung 3 (build)
3 Calculate real 5-year TCO — including compounding price increases
4 Shadow IT reveals unmet needs — these are build candidates, not governance problems

Chapter 6 showed how to audit existing SaaS. Chapter 7 addresses the capability question: what do you need internally to execute "build instead of buy"?

Part III: Applying the Framework

Building Organisational AI Capability

You need less than you think. Here's the minimum viable capability.

The executive meeting about audit results. CFO: "So we've identified $200K/year in SaaS we could build instead. What do we need to execute?" CTO: "We need three developers, a DevOps engineer, and probably a product manager."

CEO: "That's $600K in new headcount to save $200K?"

The real answer: they need far less than they think.¹²

The Capability Myth

What Organisations Think They Need

• "We need a development team"
• "We need DevOps infrastructure"
• "We need product management"
• "We need security review"
• "We need ongoing maintenance staff"

Assumed cost: $500K-$1M/year

What's Actually Required (Post-AI)

• One capable developer with AI tools
• AI tooling access ($20-50/month)
• Agent documentation (agents.md)
• Specification capability
• Willingness to experiment

Actual cost: $160K-$330K/year

The Real Requirements

1. One Developer with AI Capability

Not "a development team"—one capable developer with AI tools.¹⁴

What "capable" means:

• Comfortable with AI coding tools
• Understands spec-first development
• Can review AI-generated code

Cost:

$100K-$150K/year^26,27

(vs $500K+ for a team)

2. AI Tooling Access

Give developers their pick: Claude Code, Cursor, Devin.⁵

Per-developer

$20-$50/month²⁹

API access

$1K-$5K/month³⁰

Infrastructure

$2K-$10K/month^31,32

3. Agent Documentation (agents.md)

Codebase-specific docs that compound over time.⁵

"Every manual edit you make is an opportunity for agents.md improvement."⁵

— Raul @ Ramp

4. Specification Capability

The durable asset is the spec, not the code.

Who has this: Product managers, business analysts, subject matter experts—the people who know the problem best. You don't need developers to write specs.

5. Willingness to Experiment

"Now more than ever, ask forgiveness not permission feels almost essential."²

— Theo Browne

The Raul @ Ramp Playbook

Use coding agents

Give all engineers their pick of harnesses, models, background agents⁵

Give agents tools to ALL dev tooling

Linear, GitHub, Datadog, Sentry, internal tooling. "If agents are held back by lack of context, that's your fault."⁵

Invest in codebase-specific agent docs

agents.md file in every repo, updated when AI makes mistakes⁵

Figure out security issues

"Stop being risk averse and do what is needed to unblock access to tools."⁵

The Spec-as-Asset Principle

Why maintenance burden is lower than organisations expect:¹⁴

Old World	New World
Code is the asset	Specification is the asset
Maintaining code is expensive	Regeneration is cheap
Developer leaves = knowledge loss	Spec remains, regenerate code
Model upgrades = migration project	Model upgrades = free improvement

Minimum Viable Capability

To Execute "Build Instead of Buy":

One developer with AI skills	$100K-$150K/year^26,27
AI tooling	$36K-$60K/year^29,30
Infrastructure	$24K-$120K/year^31,32
Total	$160K-$330K/year

If SaaS audit identified $200K/year in build candidates, breakeven in Year 1.

Key Takeaways

1 One developer, not a team — AI tools multiply individual capability 2-5x¹⁴
2 Tooling is cheap — $36K-$180K/year for comprehensive AI development capability^29,30,31,32
3 Specification is the capability — you need people who understand problems, not just coders
4 Agents.md compounds — every project makes the next one easier⁵

The Window Is Open

The economics have inverted.¹⁴ Those who recognise this shift build competitive moats—proprietary systems, encoded expertise, capabilities that off-the-shelf SaaS can't deliver.

Those who don't keep paying rent to vendors, watching costs compound,⁷ waiting for features that serve average use cases instead of their specific needs.

What's the first SaaS subscription you're going to question at renewal?

References & Sources

Sources cited throughout this ebook, organised by category.

LinkedIn Commentary

1Andrej Karpathy: Karpathy X Post

10x more powerful quote

https://x.com/karpathy/status/2004607146781278521

13Business Insider: Karpathy "Alien Tool" Quote

Magnitude 9 earthquake, powerful alien tool

https://www.businessinsider.com/openai-founding-member-never-felt-so-behind-programmer-2025-12

5Raul @ Ramp: AI Leader Playbook

No unforced errors playbook

https://x.com/rahulgs/status/2006090208823910573

6LinkedIn: OpenAI Internal Usage

70% more PRs per week

https://www.linkedin.com/posts/justinhaywardjohnson_openai-unveils-o3-and-o4-mini-activity-7318687442868342784-1l3m

Industry Analysis

2Theo Browne: You're Falling Behind

70%+ AI-generated code quote

https://www.youtube.com/watch?v=Z9UxjmNF7b0

7SaaStr: The Great SaaS Price Surge of 2025

11.4% vs 2.7% inflation

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/

9Software Pricing Guide: Enterprise SaaS Pricing Reality

Renewal discount decay

https://softwarepricingguide.com/enterprise-saas-pricing-in-2025-the-7-year-cost-reality/

10CIO Magazine: SaaS Price Hikes

Lock-in driving increases; PE-backed vendors up to 900%

https://www.cio.com/article/4104365/saas-price-hikes-put-cios-budgets-in-a-bind.html

15Medium: Build vs Buy Analysis

15-25% aggressive vendor hikes; 20-30% hidden inflation from unbundling

https://medium.com/@shuaib_18577/build-vs-buy-why-modern-companies-must-own-their-digital-platforms-264fcc82ba31

Primary Research

3arXiv: GitHub Copilot Impact Study

55.8% faster task completion

https://arxiv.org/abs/2302.06590

4AMD: AI for Tech Professionals

82% reduction in development time

https://www.amd.com/en/blogs/2025/ryzen-pro-ai-pcs-for-tech-professionals.html

112-Data: The State of Software Costs in 2025

49% licenses unused, $21M waste; 48% shadow IT app usage

https://www.2-data.com/knowledge-hub/the-state-of-software-costs-in-2025-what-every-business-needs-to-know

12Retool: 2025 Builder Report

80% of builders can ship without support

https://retool.com/blog/2025-builder-report

14LeverageAI: The AI Paradox

Build-vs-buy calculation has flipped; $500K to $50K-$150K cost reduction; 2-5x developer productivity

https://leverageai.com.au/the-ai-paradox-why-68-of-smbs-are-using-ai-but-72-are-failing

LeverageAI / Scott Farrell

Practitioner frameworks and interpretive analysis developed through enterprise AI transformation consulting.

25Scott Farrell: Build vs Buy Insight

"AI knows the expert advice. Your job is to work out what is valuable" - foundational insight on encoding expertise

https://leverageai.com.au

16LeverageAI: Maximising AI Cognition and AI Value Creation

70-85% AI project failure rate; real-time contexts are hardest for AI