← All writing

Build, buy, or wait

June 2026 · 7 min read · vallesAI

The standard AI build-versus-buy analysis compares a vendor's list price against an internal build estimate, finds the vendor cheaper, and stops. It's usually wrong on both sides of the comparison — and it always omits the third option, which is sometimes the right one: wait.

What "buy" actually costs

The vendor quote is the beginning of the cost, not the end. Add the integration work (real, and yours), the data pipeline the tool needs, the workflow changes it imposes, and — the number nobody prices — the switching cost you're accepting. In a market where capability doubles and prices halve on a rhythm of quarters, a hard-to-leave vendor contract is a bet that this specific product tracks the frontier. Ask the vendor two questions: can we export everything, in a usable format? and what happens to our workflows the day we leave? The answers price the lock-in.

What "build" actually costs

Internal estimates fail in the other direction: they price the prototype and forget the system. The model-API part of a build is genuinely cheap now. What costs money is everything around it — evaluation, monitoring, failure handling, the human-review loop, and above all ongoing ownership. A built system is a product your organization now maintains, forever, in a fast-moving field. If no team will own it in eighteen months, you're not choosing "build," you're choosing "abandon later."

The grid that actually decides it

Strip away the noise and two questions decide most cases:

• Is this capability differentiating for you? If it's how you win — the workflow that is your product, the data advantage competitors can't copy — build, because a vendor sells the same capability to your competitors. If it's commodity plumbing (transcription, generic summarization, OCR), buy the plumbing.
• Is the problem stable enough to invest in? If your own process is still changing monthly, or the model capability you need is visibly one generation away, neither build nor buy — wait, and spend a fraction of the budget preparing: collect the data, define "correct," build the evaluation set. Waiting with preparation is a position; waiting without it is just delay.

Vendors never recommend waiting, and internal teams rarely recommend buying. Notice who's giving you the analysis.

The uncomfortable part

That last line applies to us too: a consultancy's incentive is to recommend the option with the most engineering in it. Two things keep us honest. Advisory engagements are fixed-fee and end with a written recommendation — we get paid the same whether the answer is build, buy, or wait. And the recommendation ships with its reasoning and its numbers, so you can check the work. That's the standard you should hold anyone to, including us.