STOP Customizing, STOP Technical Debt, START Leveraging AI

Why your platform customization spend has zero AI leverage — and what to do instead.

Andrej Karpathy — one of the most respected AI researchers alive — just announced he went from 80% manual coding to 80% AI-generated code in a matter of weeks. “This is easily the biggest change to my basic coding workflow in ~2 decades of programming,” he wrote. “It happened over the course of a few weeks.”¹

The Vercel CTO put it more bluntly: “The cost of software production is trending towards zero.”²

Meanwhile, you’re paying $200K/year for Salesforce, plus $150K for two admins, plus $300K for implementation partners, plus $100K in annual consulting retainers.³

That’s nearly a million dollars in year one — to make their generic platform fit your specific workflows.

Here’s the uncomfortable question: Where is AI in that equation?

The answer: Nowhere. AI can barely touch Salesforce configuration. Every dollar you spend on platform customization offers zero AI leverage.

The Stupidity Tax

Let me name what’s actually happening.

You bought software “made for the average.” Then you spent a fortune making it un-average. You hired consultants to translate your requirements into what the platform allows. You hired admins to maintain those translations. You paid integration partners to connect it to your other systems.

And you called this a “platform investment.”

But here’s what you actually did: You built custom software — badly — inside someone else’s prison.

If you’re spending more to customize software than on the software itself, the economics have inverted. And you’re on the wrong side.

Why Platform Configuration Has No AI Leverage

Here’s what Karpathy described about his shift to AI-assisted development:

“I am mostly programming in English now, a bit sheepishly telling the LLM what code to write… in words. It hurts the ego a bit but the power to operate over software in large ‘code actions’ is just too net useful.”

— Andrej Karpathy¹

AI is brilliant at code. It understands codebases. It can navigate them, refactor them, test them, improve them. In fact, 41% of all code written in 2025 is now AI-generated.⁴

But Salesforce configuration? ServiceNow workflows? HubSpot automations?

AI can barely help. Why? Proprietary configuration languages. Closed ecosystems. No version control in any meaningful sense. No automated testing infrastructure. No CI/CD pipelines.

Platform configuration debt accumulates without any of the tools we’ve built over 50 years of software engineering.

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, and locks you deeper into someone else’s roadmap.

Every dollar you pour into spec-driven AI development gets better as models improve (for free), compounds as your team masters AI tooling, transfers to any future system, and builds assets you own.

The Three Tiers: What to STOP vs What to START

Not all software is the same. Here’s the framework for deciding what stays “buy” and what flips to “build”:

Tier 1: Commodity Plumbing — Keep Buying

• Examples: Auth0, Stripe, Twilio, SendGrid, Cloudflare
• Characteristics: Real-time critical, deeply operational, cheap per-unit, zero customization anyway
• The test: You’re not fitting them to your business — you’re plugging them in

These stay. They’re efficient, cheap, and you aren’t customizing them anyway. The plumbing works.

Tier 2: Horizontal Platforms — The Flip Zone

• Examples: Salesforce, HubSpot, ServiceNow, Workday
• Characteristics: Expensive base cost, massive customization projects, consultant armies, internal admin staff
• The dirty secret: You’re not paying for software — you’re paying to make generic software fit your specific case

This is where the economics have inverted. If you’re going to customize it, parameterize it, pay consultants to configure it, and hire staff to maintain it — you may as well throw it out and go AI coding.

Tier 3: Industry Verticals — Also Flip Zone

• Examples: “Insurance CRM,” “Healthcare scheduling,” “Legal practice management”
• Characteristics: Claims to fit your industry but is still 60-70% generic
• The vendor promise: “Built for your industry!”
• The reality: “Built for the average of your industry”

STOP Playing the Averages

Here’s the deeper problem with SaaS platforms.

When you buy Salesforce, you’re buying software designed for the average of all Salesforce customers. When you buy a vertical industry solution, you’re buying software designed for the average of your industry.

You’re playing the averages.

But your competitive advantage doesn’t come from being average. It comes from being different. From serving your specific customers in your specific way with your specific workflows.

“If you and your competitor are all using the same service, you have no edge over each other. Their AI and your AI against each other — I don’t know who’s going to win.”

— Mehdi Paryavi, CEO, International Data Center Authority⁵

Using the same SaaS as your competitors = same capabilities = no differentiation.

Building custom with AI = your specific workflows = competitive advantage.

STOP Playing Safe

“No one got fired for buying Salesforce.”

That was true. Past tense.

Here’s what’s changing:

• SaaS inflation: Running 5x faster than general inflation (12.2% vs 2.4% for G7 countries)⁶
• AI coding capability: 41% of all code now AI-generated; 25% of Y Combinator startups ship 95% AI code^4,7
• Development cost collapse: Custom AI development costs have fallen 70-90%⁸
• Paradigm shift speed: Karpathy’s transformation happened in weeks, not years¹

The “safe” choice — buying platforms and customizing them — is becoming the risky choice. You’re investing in systems that can’t benefit from AI improvements, lock you into escalating vendor pricing, accumulate technical debt without solution paths, and offer no competitive differentiation.

Meanwhile, your competitors who are building with AI are getting better every time models improve, paying near-zero marginal cost for new features, shipping in days what takes you months, and building actual competitive advantages.

Your Sunk Cost Is Your Spec

Here’s the operational insight that changes everything.

Your current Salesforce implementation IS your requirements document.

• Every custom field = a data requirement
• Every workflow rule = a business rule
• Every report = an output specification
• Every integration = an interface contract

You don’t start from zero. You start from “what we actually built, expressed as platform configuration.”

Export that. Translate it to spec. Regenerate as software you own.

Your sunk cost becomes your asset — not money thrown away, but requirements documented in detail. The consultants who configured your platform effectively wrote the spec for what you actually need.

Now AI can build it. In code. That you own. That you can test. That improves with every model upgrade.

The Overnight Sprint

Here’s what becomes possible when you move from platform configuration to spec-driven development.

Old model (Salesforce change request):

Week 1: Submit request to admin team
Week 2: Requirements gathering meetings
Week 3: Impact analysis (what does this touch?)
Week 4: Development in sandbox
Week 5: Testing, UAT
Week 6: Deploy to production (hopefully)

New model (Spec-driven overnight sprint):

5pm: Business requirement arrives
     ↓ Spec changes drafted
     ↓ AI regenerates affected code
     ↓ Automated tests run
     ↓ PR created with diff + audit trail
9am: Human reviews PR, approves, deploys

Same business outcome. Six weeks → one night.

This isn’t science fiction. Elite developers are already doing this. Theo Brown generated 11,900 lines of code without opening an IDE.⁹ The creator of Claude Code ships 100% AI-written code.² Companies are building enterprise-grade products with small teams in months.¹⁰

The AI-Enabled Team

The money you’re spending on platform customization already exists. It’s just going to the wrong place.

Instead of:

• 2 Salesforce admins ($150K each)
• Implementation partner ($300K)
• Annual consultant retainer ($100K)
• Integration maintenance ($80K)

Redirect to:

• AI-proficient developers who write specs, not configs
• AI tooling and infrastructure
• Domain expertise that gets encoded into specifications

The difference:

• Platform admins: Linear returns, no AI leverage
• AI-enabled team: 20-100x leverage, compounding with every model improvement

Work towards an AI-first execution team that can power through requirements. Staff who are empowered, not constrained. Who work with AI, not against proprietary configuration languages.

The Harness Is the Moat

One more insight from the frontier.

The people getting the most from AI coding aren’t just using generic tools. They’re building harnesses — domain-specific configurations, workflows, and specifications that make AI dramatically more useful for their specific context.

“The harness is the moat. The harness is the product. The harness is what separates ‘I vibe coded a broken app’ from ‘I built a system that actually works.'”

— Lumberjack¹¹

If you understand your domain — your specific workflows, your specific data models, your specific business rules — you can encode that understanding into specifications that make AI dramatically more powerful than any generic SaaS platform.

Your domain knowledge becomes your competitive advantage. Not your vendor’s generic features.

The Decision

You have a choice.

Path A: Keep customizing platforms

• Zero AI leverage
• Escalating vendor costs (5x inflation)
• Technical debt without solution paths
• Same capabilities as competitors
• Playing safe, playing the averages

Path B: Build with AI

• Massive AI leverage (80%+ AI-generated)
• Costs trending toward zero
• Regeneration beats patching
• Custom = differentiation
• Playing to win

The economics have inverted. The safe choice isn’t safe anymore.

Your Salesforce admin is awake 40 hours a week.

Your AI team is awake 168.

The question isn’t whether to make this shift.

The question is whether you’ll make it before your competitors do.

If AI can’t help you with it, you’re building the wrong way.