Buy the system of record. Build the system of intelligence.
Every serious AI team hits the same fork on every project: buy this, or build it? Case by case it turns into a hundred small arguments nobody wins. There is a cleaner rule underneath, and it decides where your scarce engineering compounds. Your advantage lives in the layer you build, never in the one you bought.
The discipline
Buy the system of record. Build the system of intelligence.
The principle
The system of record is the infrastructure everyone in your industry needs and no one wins on, so buy it. The system of intelligence is the layer that turns your proprietary data and your particular way of working into advantage, so build that. Get the line right and you spend your scarce engineering where it compounds, instead of rebuilding what you could have licensed in a week. And know that the line has moved: AI has made building faster and cheaper than it has ever been, so the honest default now leans toward build more than it used to.
One company builds everything and pours its best people into a worse version of software it could have bought. The other buys everything, ends up with the same stack as every competitor, and wonders why the tools handed it no edge.
Both spent real money. Neither bought advantage, because advantage does not come from the tool. It comes from what only you can put on top of it.
Buy the system of record
Buy when the problem is broadly shared, the vendors are strong, differentiation is thin, speed matters, and maintaining your own version would add nothing. That covers most of the stack: the cloud, the general-purpose AI environments, the electronic lab notebook, the LIMS, the document and quality systems. These are systems of record. They hold the work, they have to be reliable and well integrated, and the market already builds them better than you will. Buying them is not a concession. It is how you free your best people to work on the part that matters.
Build the system of intelligence
You build for two reasons, and the bar for both has dropped. You cannot buy it: the capability is genuinely unavailable, or your proprietary data and your particular workflow are the whole point and no vendor has them. Or you will not buy it: it is bespoke and tailored enough that a vendor implementation would cost more work and compromise than building it yourself. Either way this is the system of intelligence, the layer that reads across your own data, encodes your hard-won method, and produces a decision no vendor can sell you. With AI in hand, a focused build is often faster than the procurement it would replace, so build more of this than you used to. The only thing to refuse is a rebuild of pure commodity: if the thing could simply be bought and wins you nothing, you are rebuilding a record, slower and at your own expense.
Self-service first, and two more options open
Before buy or build there is a front door: self-service. Give people secure access to powerful general-purpose tools, paired with real education and clear accountability, and a surprising amount of value arrives with no project at all. Then two more options belong on the table, not in the footnotes. Redesign, when the honest fix is not a tool but a different workflow or decision structure, because automating a bad process only gives you a faster bad process. And do nothing, because not every problem needs AI, and declining the weak requests protects your attention and your credibility. Self-service, buy, build, redesign, do nothing. Routing each request through those five is most of the discipline.
What it is, and what it is not
This is a routing logic, not a maturity ladder. You do not graduate from buying to building. A healthy company does all of it at once, matched to the problem in front of it. The test never changes: is this a capability the market shares, or one that is distinctively yours? And the line is not fixed. What is a differentiating build this year can become a commodity you should buy in two, once the vendors catch up. Revisit it, or you will find yourself maintaining a proud, expensive system the market now sells for a fraction.
In biotech
Biotech makes the line concrete. Buy the sequencing plumbing, the lab systems, the enterprise AI environment. No patient is better served because you wrote your own document management. But the model trained on your assay data, the tool that reads across your own programs to surface a signal a competitor cannot see, the workflow built around your particular science: that is the system of intelligence, and it is worth building precisely because no one can sell it to you. Which data is yours to build on, and where it is allowed to go, is a data question before it is an engineering one. That is the same instinct behind Permission, People, Programs, where the sensitivity of the data decides the environment it lives in.
The old caution was building too much, rebuilding the commodity you should have bought. The newer mistake is worse, and more common: only building big, swinging for the fences while the small problems people have go unsolved.
Adoption is the whole game. Solve one person's real problem, and they bring you a bigger idea than they had before. Solve for the individual, and you unlock the institutional.
Route the loud, then build the small.
Take your three loudest AI requests and route each honestly through self-service, buy, build, redesign, do nothing. If they all came back build, check you are not rebuilding commodity you could license.
Then do the harder thing. Find the three smallest problems your best people keep hitting, the ones too minor to make any roadmap, and build one this week. That is where adoption starts, and adoption is the whole game. Give the routing to an AI Product Partner whose job is to know the difference.