Buy when a vendor product already covers most of the requirement and your own data is not what makes the output good. Build when the value sits in data only you hold, in an integration surface no vendor prices for, or in a workflow that changes faster than a product roadmap. Most build-versus-buy arguments go wrong before the numbers appear, because the two options get costed on different lines: the licence is compared against engineer day rates, and the data work that both options need never enters the buy column.
Cost both columns the same way
Take a 24-month horizon and put identical lines in each column.
| Cost line | Buy | Build |
|---|---|---|
| Licence or subscription | €40k to €200k a year for mid-market AI and data tooling | none |
| Integration and data preparation | 30 to 80 engineer days | 30 to 80 engineer days |
| Core build | none | two to three engineers, four to eight months |
| Change and run | vendor roadmap, plus 10 to 20 days a year of your own team | 15% to 25% of the original build effort a year |
| Exit | re-platform, two to four months | the code stays in your repository |
The second line settles more cases than the first. Getting your data into a shape a model can use costs roughly the same either way, because no product reads twelve years of free-text CRM fields on your behalf. If your evaluation shows integration at five days for the bought option and sixty for the built one, the evaluation is wrong, not the options.
When buying wins
- The capability is commodity: transcription, OCR, document classification, standard demand forecasting, a BI semantic layer.
- The product covers roughly 80% of the requirement as shipped. Below about 60%, the gap becomes a build with a licence fee attached.
- You are in a regulated function and want a vendor’s certification and contractual indemnity to point at during an audit.
- Your engineering team is under about fifteen people and every internal system already competes for the same two seniors.
When building wins
- Your data is the differentiator. A shared model trained on everyone’s data gives you a competitor’s answer.
- The integration surface spans five or more internal systems, which is where per-connector pricing and vendor connector backlogs start to bite.
- Per-seat or per-call pricing punishes success. A product at €0.05 a call is trivial at 100,000 calls a month and a board-level line item at 50 million.
- The logic changes monthly. Waiting two quarters for a vendor to ship your edge case costs more than the licence ever saved.
- Retrieval quality decides adoption. Chunking, ranking and evaluation on your own corpus are where the accuracy comes from, which is the argument set out in our guide to how retrieval-augmented generation is put together.
The hybrid that costs more than either option
The expensive outcome is buying a platform and then building a second system around it to compensate for the parts that do not fit. It carries both cost structures and neither set of benefits. Watch for one signal: the share of engineer time spent on workarounds rather than features. Past about 40%, you have paid for a product and built one anyway, and the cheaper correction is usually to keep the product for the narrow thing it does well and move the rest into your own code.
For the build column, day rates and blended team costs move the answer more than any other input, so use current figures from our published daily rates and size the effort with the project cost calculator before the debate becomes philosophical. The drivers behind those numbers are set out in our note on what actually moves the cost of an AI build.
Where renting engineers is the wrong answer
Four cases where you should not call us:
- The requirement is genuinely commodity. Buy the product, integrate it in six weeks, spend your engineers on something a vendor cannot sell you.
- Nobody internal owns the roadmap. Rented engineers build what you specify. With no specifier, you get an expensive prototype and a maintenance liability.
- The work is under about twenty days. Onboarding overhead eats the benefit.
- You want a single throat to choke on a fixed outcome. That is a delivery contract, not time-and-materials capacity.
If none of those apply, the decision is arithmetic, and the arithmetic is usually decided by the integration line rather than the licence.