-
1Is the problem still unnamed, unowned or unmeasured?
YesStop and frame it. No tool fixes a problem nobody owns. Start at the first rung of the production ladder.
-
2Does a product you already license do this, or nearly?
YesConfigure. Switch on and govern what you have. The cheapest production is the one already in your tenancy, provided permissions are clean.
-
3Is it a common task, done much the same way across your sector?
YesBuy. A mature product will be ahead of anything you build, and its supplier carries the upkeep. Buy on evidence, with an exit.
-
4Is it regulated specialist work: a medical device, a credit decision, a safety case?
YesRefer. Bring in the specialist or the certified product. Your part is the integration and the governance around it.
-
5Is the process part of how you win, on data only you hold?
YesBuild. Small and owned, on your data, through the same production gates as anything bought.
No to allWait. Keep it on the register as a watch item and look again at the next review. Not every process needs AI this year.
Ask the questions in this order. Each “no” falls through to the next. Source: Vardonne
Each answer, in the library
- Configure
- Shoosmiths and Balfour Beatty rolled out assistants they already licensed, and the work was incentives, training and permissions. Library.
- Buy
- Rotherham’s IT service desk agent and Newcastle’s contact-centre prompts are bought products fitted to local knowledge. Library.
- Refer
- Stroke scan support is a regulated medical device, bought with its certification and governed clinically. Library.
- Build
- Garfield.Law and Ocado built because the capability is the business. Library.
The costs each answer hides
Escalating cost is one of the four reasons Gartner gives for projects abandoned after proof of concept.1 The cost that escalates is rarely the one compared at the start.
| Cost | Bought | Built |
|---|---|---|
| Running cost | Licences and usage at full volume, not pilot volume | Hosting and model usage at full volume |
| Integration | Connecting to your systems of record | The same, with no vendor connector to start from |
| Evaluation | Testing the vendor’s claims on your data | Building the test set and keeping it current |
| Ownership | A named owner and a supplier relationship | A named owner and the people to maintain it |
| Exit | Data export and a notice period | Decommissioning, and what replaces it |
Where Vardonne stands
No vendor pays us to recommend it. When the answer is to build, the build may be proposed to ArcSystemsTechnology, the same group’s build arm, and that is stated in the proposal and priced under its own engagement letter, so the advice and the delivery are never confused.
Sources
- Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025. Gartner, 29 July 2024.
- Use cases in production (the entries named in this guide). Vardonne, 21 September 2026.