Skip to content
Fritz Gerald ZephirinMenu

Social

Buy the Boring, Build the Different

3 min read

Buried inside MIT’s famous finding that 95 percent of enterprise AI pilots showed no P&L impact is a less-quoted number that’s worth more: in the same study, purchased solutions succeeded roughly twice as often as internal builds. The headline says AI fails. The footnote says building your own AI fails — buying focused tools and embedding them in one workflow is what the successful five percent mostly did.

This matches what I see from the operator’s chair so exactly that I’ve turned it into a standing rule: buy the boring, build the different.

First, why internal builds fail at twice the rate. It is almost never the technology. An internal build is a product, and most companies that aren’t product companies underestimate what a product costs by roughly everything after the demo: maintenance, evals, edge cases, documentation, the person who leaves taking the context with them. The MIT authors called the failure a “learning gap” — the integration and iteration work around the tool, not the model inside it. A vendor amortizes that work across a thousand customers. Your internal platform team amortizes it across one — you — while a procurement-grade version of their roadmap ships quarterly to your competitors. Meanwhile the real motives for building are rarely written in the doc: control feels safe, platforms look good on engineering résumés, and “we’re building our own AI” sounds better at the board meeting than “we bought a tool.” Name those incentives honestly and half the build proposals withdraw themselves.

The test I use is the two-layer rule. Take what customers actually pay you for. You may build that layer and the one directly beneath it — the place where your workflow knowledge is the moat. Everything further down is boring by definition — not unimportant, just undifferentiating — and boring things should be bought from people for whom they are the whole company. In my world: clients pay for resolved conversations and reliable payroll. So we build escalation logic, quality rubrics, the knowledge architecture, payroll-corridor reliability — and we buy ticketing, transcription, models themselves, dashboards, with no more ceremony than we buy laptops. A model, in particular, is the most boring layer of all now: it’s a commodity with a price war attached. Renting it is not a strategic decision. What you wrap around it is.

Two disciplines keep the rule honest:

Every build gets a kill date. A written check-in — six months is right for most — where the build must beat the best buyable alternative on the metrics you set at kickoff, or be replaced without shame. Sunk-cost defense of internal tools is where engineering capacity goes to die.

Every buy gets a workflow change. Buying the tool is the cheap part; a tool bolted onto an unchanged workflow returns nothing. The purchase isn’t done when the contract signs; it’s done when a step has been deleted.

Now the honest caveats, because the rule has edges. The MIT study is one study — self-reported, contested in its details, directional rather than gospel; I trust it because it agrees with the scar tissue, not the other way around. And “buy” loses to “build” in two real cases: when the boring layer becomes strategic (volume so high that vendor margin is your biggest cost line — the moment ADP-style conversions happen), and when no vendor’s tool can hold your actual constraint (regulated data paths, corridor-specific payroll logic nobody else needs). Those exceptions are rarer than every engineering team believes, and both announce themselves with numbers, not vibes.

The quiet truth under all of it: build-versus-buy is not a technology decision, it’s a focus decision. Your company has a fixed budget of attention for being genuinely different. Every boring thing you build spends differentiation budget on something customers can’t see. Buy the boring. Spend the difference where they’re looking.

Sources & further reading