Every few weeks a vendor, a peer, or a board member tells you to “do something with AI,” and the question arrives pre-framed: should you buy it, build it, or hire for it? That framing is a quiet trap, because it starts at the solution. The question that actually decides the outcome comes one step earlier. Is the thing you want to automate a commodity capability, the kind hundreds of companies need in roughly the same shape, or a workflow that is specific to how your business really runs? Get that distinction right and buy-build-hire mostly answers itself. Get it wrong and you tend to pay for it twice: once to build what you could have bought, or once to buy a tool that never touches the work that makes you money.
Spending is not the hard part
The real gap is not between companies that invest in AI and companies that sit it out. It is between companies that get value and companies that get a bill. BCG’s 2025 survey of 1,250 executives found only about 5 percent of firms are “future-built” for AI, while roughly 60 percent of firms, the group BCG calls laggards, reported minimal revenue or cost gains. MIT’s NANDA group, reviewing more than 300 disclosed initiatives, put it more starkly: around 95 percent of organizations were getting zero measurable return on generative AI. Both figures are self-reported and correlational, so treat them as direction, not decimals. What matters is that both reports land on the same cause, and the cause is the useful part. The divide was not driven by model quality or budget. It was driven by approach: the winners wired AI into one real workflow and let it improve, while the rest bought a capability and waited to be paid back. Your buy-build-hire decision is really a decision about which approach you are set up to run.
Buy when the capability is a commodity
Most of what an operator wants from AI in year one is a commodity: transcription, document extraction, first-draft copy, support triage, meeting notes. Many vendors sell each in nearly the same shape, the capability is well understood, and switching costs are low. Buy these. The enterprise buyers are already voting this way. a16z’s survey of 100 CIOs found a marked shift toward buying third-party apps over the prior year, as the ecosystem matured, and a candid reason for it: internally built tools were “difficult to maintain and frequently don’t give them a business advantage.” The deployment math points the same direction. In MIT’s sample, tools bought through vendor partnerships reached full deployment about twice as often as internally built ones. The report’s own caution applies, that this may reflect the capabilities of the firms as much as the choice itself, so do not over-read the ratio. But the logic holds: when a capability is generic, the maintenance, the model upgrades, and the security reviews are work you want to be someone else’s problem. (a16z is a venture firm that funds AI app startups, so weigh its “buy” enthusiasm accordingly; the CIO quotes still ring true.)
Build when the workflow is the moat
Building earns its keep in exactly one place: where the workflow itself is your advantage. If the process encodes something competitors cannot easily copy, your proprietary data, a hard-won operational sequence, a quality bar specific to your customers, then a generic tool will flatten the very thing that makes you good. That is the case for building, and it is narrower than it feels at 9 a.m. with a board deck open. The cost of getting it wrong is not the build; it is the upkeep. Those same CIOs flagged internal tools as hard to maintain, and an AI system is not a project that ends, it is a system that drifts: models change, prompts rot, the data shifts under it. So the discipline is to build only the thin layer that is genuinely yours, and buy everything around it. Custom orchestration over a bought model, not a bought workflow over custom infrastructure.
Hire (or partner) when the gap is capability, not tooling
Sometimes the missing piece is not a product at all. It is someone who can tell a good eval from a vanity metric and knows why the demo that dazzled in the meeting will fail in production. Here the market is unkind to small companies. ManpowerGroup’s 2026 survey of more than 39,000 employers found that AI skills had, for the first time, become the single hardest capability to hire for, ahead of traditional engineering. For an SMB, a full-time senior AI hire is expensive, slow to find, and easy to under-use once the first project ships. So weigh three honest options against the work, not the org chart. Upskill someone you trust who already understands the business, which is cheapest and slowest. Hire, when AI will be load-bearing enough to keep a specialist busy. Or partner with a studio or fractional team for the build-and-handoff, which trades some control for speed and avoids carrying a rare salary you cannot keep fed. None is the “right” answer; the right answer is whichever matches how much AI work you will actually have twelve months out.
The sequence, not the slogan
So run the decision in order, and let each answer gate the next. First: is this capability a commodity, or is it yours? If it is a commodity, buy the best-fit tool, wire it into one real workflow, and measure whether it moved anything. If it is genuinely yours and you have the capability in-house, build the thin proprietary layer and buy the rest. If it is yours and you do not have the capability, hire or partner before you build, because a custom system without someone to maintain it is a liability with a launch date. None of this requires a CTO. It requires resisting the urge to start at “build” because it sounds ambitious, or at “buy” because it sounds safe. The Chamber of Commerce found that 58 percent of US small businesses now use generative AI, up from 40 percent a year earlier. The ones pulling ahead are not the ones who spent the most. They are the ones who picked a single workflow and made it real. Pick yours, measure it, and earn the second one.