Should You Build Your Own AI?

There is a version of this conversation happening in your building right now. An internal team wants to build the thing rather than pay for it. A consultant has quoted you a custom platform and used the word “proprietary” four times. Or a board member asked, in a budget meeting, why you’re renting from vendors when you could own the capability outright. The pitch is always the same underneath. Off-the-shelf is generic. Your business is special. Build it and it’s yours.

The instinct is understandable and mostly wrong. Not because building is beneath you. Because the thing you would be building is almost never the part that matters, and the part you actually own decays faster than anything you’ve owned before.

You are not building AI

Start with what “build your own AI” actually means in 2026. You are not training a model. Almost nobody is, and the handful of companies that should are not reading this to find out. The frontier models cost hundreds of millions of dollars to train and are obsolete inside a year. You are going to call someone else’s model over an API, the same model your competitor calls.

What you would be building is everything around that call. The data pipelines that feed it. The retrieval layer that gives it your documents. The evaluation harness that tells you whether it’s getting worse. The permissions, the logging, the interface your people actually touch. The connective tissue.

That tissue has a name in the engineering literature, and the name is a warning. A mature machine-learning system is at most five percent model and at least ninety-five percent glue.1 The model is the part everyone pictures. The glue is the part you pay for, forever, and it is the part that breaks.

This is where AI departs from ordinary software. When you buy a normal piece of software, the ground under it holds still. The build you commission this year runs on roughly the same assumptions next year. AI glue does not get that. It sits on top of a model that the vendor deprecates on their schedule, not yours. OpenAI commits to a minimum of six months’ notice before retiring a model you’ve built on.2 Six months. Every prompt your team tuned, every quirk they engineered around, every eval you calibrated, aimed at a version with an expiration date. Thirty-seven percent of enterprises now run five or more models in production, and the ones who’ve tried to switch describe it plainly: the prompts were tuned for one provider, and changing models became a task that eats real engineering time.3 You are not buying an asset. You are renting a foundation and paying to rebuild the house on it every time the landlord renovates.

What the numbers say about the instinct

The market has now run this experiment at scale, and the results are not close.

MIT’s NANDA initiative reviewed three hundred enterprise AI deployments and found that ninety-five percent of generative AI pilots produced no measurable impact on the bottom line.4 That number gets quoted as an indictment of the technology. It’s more useful as an indictment of a format. Dig into the same research and a split appears: initiatives built on purchased tools and vendor partnerships reached deployment roughly twice as often as ones built internally, about two in three versus one in three.4 Buying isn’t a little better. It roughly doubles your odds of the thing ever reaching production.

The broader failure data points the same way. RAND, interviewing senior data scientists, put the failure rate for AI projects north of eighty percent, about double the rate of ordinary IT projects.5 S&P Global watched the share of companies abandoning most of their AI initiatives before production jump from seventeen percent to forty-two percent in a single year.6 Gartner expects more than forty percent of agentic AI projects to be canceled by 2027, on cost and unclear value.7 The wreckage is concentrated in exactly the place the “build it, it’s ours” instinct sends you.

The reference case is old enough to be a parable. MD Anderson spent sixty-two million dollars on a custom-built oncology AI with IBM. It never treated a patient.8 The flagship custom build, backed by a real budget and a serious partner, that quietly became a write-off. That story is not rare because it’s extreme. It’s the common story with a big enough number attached to make the trade press.

The thing that used to justify building just got cheap

For years there was a real argument for building: the integrations. Off-the-shelf tools didn’t talk to your systems, so you built the plumbing yourself, and once you were building the plumbing you may as well build the rest. That argument is expiring.

The connective work that used to justify a custom build is turning into commodity plumbing. An open integration standard went from roughly a hundred thousand downloads in late 2024 to more than eight million a few months later, and it’s now supported across every major model vendor.9 The practical effect for you is that “our systems are special and nothing connects to them” is a weaker excuse every quarter. The glue is increasingly something you connect, not something you commission.

This is why the sophisticated money is moving toward buying, not away from it. The people who fund AI companies for a living now describe the enterprise pattern as a marked shift toward buying third-party applications, with more than ninety percent of surveyed CIOs testing purchased tools.10 The default has flipped. Building is the exception you justify, not the baseline you assume.

When building is actually right

None of this means never build. It means the bar is specific, and most build proposals don’t clear it. Two conditions have to hold at once.

The first is that the workflow is something you actually win on. Not something you do. Something that is a source of advantage, where doing it better than competitors is part of why customers choose you. A general model plus a bought tool will match your competitors on all the generic work, because they’re buying the same tool. Build only where a proprietary workflow or a dataset nobody else has is the edge. In a market where the models are commoditizing, your private data and the system of work that captures it are close to the only durable moat left.11 If the thing you want to build is not that, you are spending scarce engineering talent to reproduce, worse and later, something you could have bought this afternoon.

The second condition is that you can check the output cheaply. This is the constraint that quietly kills most builds, and it’s the through-line of why some workflows yield to AI and others don’t. If you cannot mechanically tell whether the system’s output is right, you cannot maintain it, because every model swap and every prompt change becomes a judgment call resolved in a meeting. A build you can’t verify isn’t an asset. It’s a standing liability that degrades silently until a customer finds the bug for you.

So the rule. Build where the workflow is a genuine source of advantage and you own a cheap, trustworthy way to check the result. Buy everything else, which is most things. If a proposal can’t point to both the moat and the verifier, it’s a buy dressed up as a build, and it will cost you three times the sticker to learn that.

Something to carry

Next time someone brings you a build, before you weigh vendors or budgets, make them answer two questions in plain language.

First: if a competitor bought the best tool on the market tomorrow, what would we still be able to do that they couldn’t? If the honest answer is “not much,” the workflow is generic and you should be buying the same tool they will. Second: when this is running a year from now and the underlying model has been replaced twice, how will we know, cheaply and within hours, whether it still works? If nobody can answer that one, you are not being sold an AI capability. You are being sold a maintenance obligation with an AI label, and the person who has to defend that line item to the board in 2027 is you.

Footnotes

  1. D. Sculley et al., “Machine Learning: The High-Interest Credit Card of Technical Debt,” Google, NeurIPS 2015. The “five percent model, ninety-five percent glue” framing predates the current wave but describes it exactly. ↩

  2. OpenAI, “Deprecations,” developer documentation, 2025-2026. Minimum notice for generally available models; specialized variants get less. ↩

  3. Andreessen Horowitz, “How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025.” ↩

  4. MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025.” The build-versus-buy split is drawn from interviews and reads as directional rather than precise, but the direction is unambiguous. ↩ ↩2

  5. RAND Corporation, “The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed,” 2024. ↩

  6. S&P Global Market Intelligence, “Voice of the Enterprise: AI & Machine Learning,” 2025. ↩

  7. Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 2025. ↩

  8. Eliza Strickland, “How IBM Watson Overpromised and Underdelivered on AI Health Care,” IEEE Spectrum, 2019. ↩

  9. Model Context Protocol, “One Year of MCP,” 2025; adopted across OpenAI, Google, Microsoft, and AWS. ↩

  10. Andreessen Horowitz, “How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025.” ↩

  11. Same a16z survey: in a commoditizing model market, proprietary data and the “system of work” that captures it are described as the durable advantage. ↩