Build or buy AI: how to decide

Sooner or later every AI conversation turns into a build-or-buy conversation. There are more AI products than anyone can keep track of, and most companies already pay for several. Yet the problems that matter most to a business are often the ones none of those products quite fits.

There’s no general answer. It depends on the problem.

When buying makes sense

Buy when a product handles the job well enough. Not perfectly, but well enough that the remaining gap isn’t worth paying to close. The cost should be reasonable once you count licences, setup and the people needed to run it. And the product shouldn’t force your team into a worse way of working just to fit the tool.

General work is usually well served by products: drafting, summarizing, meeting notes, help with code. So are problems that look the same in every company, such as routine accounting tasks or common customer questions.

When building makes sense

Build when the problem depends on how your organization specifically works, meaning your processes, data, rules or customers. Products made for everyone tend to fit these jobs badly, and the effort of bending them into shape can end up costing more than building something that fits.

It’s also worth building when doing the job well is part of how you compete. If you win business because you price faster or onboard customers more smoothly than others, you probably want to own that software and be able to change it.

There are some signs that a product isn’t going to be enough. People use the tool, then copy the results into another system to finish the job. Configuration gets the product close to the real process but misses the exceptions. The data the problem needs sits in systems the product can’t reach. Or the rules change every quarter and the product can’t keep up.

The middle option

Many of the best solutions are partly bought and partly built. They use a bought foundation, such as a cloud platform, a commercial language model or an existing system of record, and add custom software for the parts specific to you: the integrations, the business logic, the interface and the safeguards.

That’s the usual shape of custom AI software. Nobody sensible trains their own large language model to process invoices. The work is in fitting everything around it to the job.

Deciding for a specific problem

Make the decision problem by problem, and only after looking closely at how the work gets done. Write a plain description of the process as it really runs, including the awkward cases (see scoping an AI use case for how). List what any solution would need to do. Then try the best available products against that list, not against the vendor’s demo.

Compare lifetime costs, including what it costs to work around a product’s gaps. Pick one, and write down why. Products improve and needs change, so it’s worth being able to revisit the reasoning in a year.

Sometimes the honest conclusion is that buying wins. An AI strategy engagement should be willing to tell you that.