Every AI initiative hits this question. Pick wrongly and you either pay to rebuild a commodity, or rent something that should have been your advantage.
Buy when
- The use case is generic: meeting notes, code assistants, general writing
- Your process matches how the product works
- Speed matters more than fit
- Data can leave your environment under the vendor’s terms
Build when
- The workflow is your competitive advantage
- It needs deep integration with your own systems
- Your data cannot leave your control
- Per-seat pricing gets painful at your volume
- You need control over models, evaluation and behaviour
The hybrid middle
The real answer is usually in between: buy the commodity layers such as models, vector search and observability, and build the workflow and integration layer that makes it yours. Almost nobody should train their own foundation model.
Five questions that settle it
1. Does this differentiate us? If competitors could buy the same result, buy.
2. How unusual is our workflow? The more you must bend the process to fit a tool, the stronger the case for building.
3. What is the three-year cost? Compare per-seat fees with build plus run costs. Include integration and internal time on both sides.
4. What is our capacity to operate it? A custom system needs ongoing care. If nobody will own it, buying or a managed partner is safer.
5. What is the exit? Who owns the data, prompts and logs if you leave? Can you switch models?
Traps
- Pilot pricing. Vendor prices often change at renewal.
- AI features that aren’t. Ask to see evaluation results on your data.
- Building for pride. Engineers love building. That is not a business case.
- Ignoring lock-in. Keep a model abstraction layer, so a switch costs days rather than months.
Answer eight questions in our Build vs Buy Decision Engine for a recommendation, or use the AI Vendor Scorecard to compare options.