Practical guides
Build versus buy for enterprise AI
The question is usually posed as a make-or-buy decision about a system. It is better posed as a question about which parts of your judgement you are willing to rent.
Framed as build versus buy, the answer is almost always buy. Vendors ship faster, cost less on day one, and carry the maintenance. Framed that way, the decision is not interesting.
The framing is wrong, though, because it treats an AI system as one thing. In practice these systems have four separable layers, and the right answer differs at each. Organisations that get this wrong do not usually buy the wrong product. They buy the right product at the wrong layer.
Four layers, four different answers
Buy the parts that are the same for everyone. Own the parts that encode how your organisation decides.
The option most organisations should be using
There is a third answer that gets missed because the question was binary: assemble. Buy the models and the plumbing, build the thin layer of context and evaluation that is specific to you, and use open components for everything between.
This is more accessible than it sounds. Serving stacks like vLLM, open interoperability standards like the Model Context Protocol, and ordinary observability through OpenTelemetry cover most of the middle. What remains for you to write is small, valuable, and yours.
MIT's The GenAI Divide found that builds involving an external partner succeeded roughly twice as often as internal-only efforts, which is often read as an argument for buying. It is better read as an argument about discipline: an outside team is forced to be specific about the workflow before anyone starts. That specificity is available to an internal team too, and the same research shows what happens without it.
When buying is clearly right
Four conditions. If three or more hold, buy and move on.
The capability is not differentiating
If two competitors could use the identical product with the identical result, there is nothing to own.
The workflow is standard
Your process resembles the industry norm closely enough that a configurable product fits without a fight.
Speed is the constraint
The value is time-sensitive enough that six weeks against six months changes the answer.
You cannot staff the operation
Not just the build. Someone has to run evaluation and monitoring for years. If that person does not exist, a vendor doing it badly still beats nobody doing it.
When building earns its cost
The reverse conditions, plus one that people under-weight.
The costs each side hides
Both options are routinely mispriced, in opposite directions.
Buying hides integration, the configuration work that is really development by another name, the renewal negotiated once you are operationally dependent, and the ceiling you meet when the product cannot express your exception. Building hides evaluation and monitoring as permanent lines, the second engineer who has to understand the system, and the process change that has to happen either way.
Compare them over three years, with the recurring lines included on both sides, using the method in how to calculate ROI for an AI initiative. Over that horizon the two totals are usually closer than either advocate expects, and the decision comes down to which layer you are buying.
A decision rule
Ask what would have to be true for a competitor using the same vendor to get a worse result than you. If there is a clear answer, that answer names the layer you should own, and everything else can be bought. If there is no clear answer, buy the whole thing and spend the saved effort somewhere that does differentiate.
Whichever way it goes, keep the parts that let you change your mind: your evaluation sets, your context, and an architecture where the model is a replaceable component. That is the subject of how to avoid AI vendor lock-in, and it is the same argument as vendor-neutral AI architecture.
Which layer are you actually buying.
That is usually the whole decision, and it takes an afternoon to establish. We are vendor-neutral by construction, which is why we are willing to tell you to buy.