The economics changed
Building AI capability used to mean training models, which meant a data science team. Now it usually means calling an API with good prompts, retrieval and evaluation — which a competent development team can do.
That has shifted the build-or-buy line considerably towards building, particularly where the AI is one step inside a larger workflow.
Buy when
- The problem is standard: transcription, translation, general document extraction, meeting notes
- A mature product exists with a real customer base in your sector
- You need it working in weeks rather than months
- The volume is modest enough that per-unit pricing is comfortable
- Nobody internally will own a bespoke system
Build when
- The AI step sits inside a workflow specific to your business
- Data residency or confidentiality requirements rule out the available products
- Product pricing becomes uneconomic at your volume
- You need it embedded in your own interface rather than in someone else's
- The output must feed systems no product integrates with
The most common trigger we see is the fifth one. The AI part works fine in the product; getting the result into the operational system is the problem, and that is a build.
The hybrid nearly everyone ends up with
Buy the model capability from a provider, build the workflow, evaluation and interface around it. That is what most custom AI projects actually are — not model training, but careful assembly.
It gives you control of the parts specific to your business and someone else's investment in the parts that are not.
Cost comparison, honestly
| Buy | Build | |
|---|---|---|
| Time to working | Weeks | 2–4 months |
| Upfront cost | Low | £15,000–£40,000 typical |
| Running cost | Per seat or per unit | Model API usage, usually lower |
| Control of quality | Limited | Complete |
| Integration depth | What they offer | Whatever you need |
| If you leave | Data export, if provided | It remains yours |
A sensible sequence
Buy something to establish whether the capability helps at all. If it does and the limits bite — pricing, integration, quality on your data — build the version that fits.
That sequence costs a few months of subscription and removes most of the risk from the build decision.