Build or buy: how to decide on an AI tool without wasting six months

Build only when one of these is true

If none of those are true, buy, and we will say so. A single product that genuinely covers the whole problem will be cheaper and better maintained than anything built for one business. The question is whether one product actually covers it, or whether you are about to assemble several and call that buying.

• The process is genuinely specific to how you operate, and every product you have tried needs workarounds that recreate the original problem.
• The data cannot leave your systems, for regulatory, contractual or competitive reasons.
• You are paying per seat for a broad platform you use in one narrow way, and the maths has stopped working.
• Three or more tools are currently held together by a person copying between them. That person is the integration, and they are expensive and error-prone.

Why the maths changed

Custom software used to mean six figures and six months, because the hard part was building capability from scratch. That has changed. Understanding language, extracting data from documents and classifying unstructured input are now services you call rather than systems you train.

What remains is the plumbing and the interface, which is well-understood work. A focused internal tool that would once have been a major project now frequently lands between eight and thirty thousand dollars.

The build-versus-buy line moved. Plenty of things that were obviously "buy" three years ago are now genuinely arguable.

The costs people forget on both sides

• Buying: per-seat growth as you hire, data export difficulty at renewal, and the workarounds that quietly become process.
• Building: hosting and model usage, the person who maintains it, and what happens when whoever commissioned it leaves.
• Both: the integration work. It is almost always the largest line item and almost always the one omitted from the estimate.

Questions worth asking a vendor

• What happens to our data, and is it used to train your models?
• Can we export everything, in a usable format, without your assistance?
• What does this cost at three times our current headcount?
• What are the three things your product is bad at? A vendor with no answer has not thought about it or is not being straight with you.

A reasonable default

Buy the platform, build the edges. Use a well-supported product for the core, and build the narrow pieces that connect it to how you actually work. That is where most of the value sits and where off-the-shelf products are weakest, because those pieces are specific to you by definition.