Skip to content
All insights

AI Strategy

Build, buy, or leave it alone: choosing per workflow, not per company

The build-versus-buy debate is usually held at the wrong altitude. It's not a company-level decision. It's a decision you make separately for every workflow.

NXTVIS Engineering7 min read

Organisations tend to adopt a posture, 'we build' or 'we buy', and then apply it uniformly. This is almost always wrong, because the correct answer changes workflow by workflow, and applying one policy across all of them guarantees overpaying in some places and over-engineering in others.

Buy when the problem isn't yours

Transcription, OCR of standard documents, generic translation, common object detection. These are solved problems where a vendor amortises the model across thousands of customers. Building your own is a way of paying more for something worse. If the problem looks identical at your company and your competitor's, buy it.

Build when the problem is the business

Your defect taxonomy, your demand patterns, your fraud signatures, your routing constraints. These are shaped by decisions specific to your operation. A vendor product will be trained on an average of everyone, and average is exactly what you are trying to beat. If the accuracy of the model is what creates the advantage, that model has to be yours.

Competitors have the identical problemBuy
Your data is the differentiatorBuild
Regulatory explainability requiredBuild
Needed in weeks, not quartersBuy, then revisit
Volume too low to repay engineeringLeave it alone

The third option nobody puts on the slide

Leaving it alone is a legitimate and frequently correct answer. Plenty of workflows are annoying without being expensive. Automating a task that occupies two hours a month is a way of feeling productive while spending engineering capacity that had a better use somewhere else.

If the workflow wouldn't make the top ten list of things costing you money, it doesn't belong in an AI roadmap, however irritating it is.

The hybrid that usually wins

In practice the strongest architecture is often bought foundations with a built layer on top: a commodity vision or language model handling the general capability, wrapped in logic, thresholds and integrations that encode how your operation actually works. You get the vendor's economies of scale on the hard general problem and keep ownership of the part that constitutes your advantage.

Deciding which layer is which requires knowing the workflow in detail. That is the analysis the audit produces, and it is the reason the answer differs line by line in the report rather than being stated once at the top.

Want this looked at in your own systems?

The audit is free, the report is yours to keep, and there is no obligation to build anything with us.