Every AI conversation in an enterprise boardroom eventually reaches the same question: do we buy a product, rent a platform, or build something ourselves? The instinct is to buy, because building sounds like a technology company's job. The economics point the other way, provided you are clear about what is being built.
Three ways to spend the same money
Rent. A per-seat or per-transaction subscription to a tool that does one thing: summarise tickets, draft emails, flag anomalies. Fast to start, easy to stop, and worth almost nothing on the day you stop. The vendor owns the roadmap, the data model and, increasingly, the learning from your usage.
Buy. A platform licence plus an implementation. More capable, and much stickier: the implementation encodes your rules in the vendor's way of thinking, and switching means starting over. The asset on your balance sheet is a licence that depreciates to zero.
Build, on a model you own. A system that connects your existing systems into one model of the business, with your rules and your evaluations, in your tenant. Slower to the first result, and the only option that gets cheaper with each subsequent use case, because the model is reused rather than rebuilt.
What the return actually depends on
The return on an AI system is not set by the model provider. Every competitor can rent the same models at the same price. It is set by three things a vendor cannot sell you: how much of your operating knowledge is captured in a form a system can act on, how many use cases share that capture, and whether the system is allowed to act rather than merely recommend. Rent gives you none of the three. Buy gives you some of the first, inside someone else's product. Build gives you all three, and they compound.
The number a CFO should hold
Marginal cost per use case. In a rented stack it is roughly constant: each new tool costs what the last one did. On an owned model it falls, because the second use case inherits the connections, objects, rules and controls of the first. By the fourth or fifth, the cost of a new system is mostly the time it takes to agree the baseline. That curve is the whole investment case, and it does not appear in any vendor's pricing page.
How to de-risk building
- Start from a P&L line you already own, with a baseline measured before anything is built.
- Insist on your own tenant, your own data, and a written statement that the model and the code are yours.
- Run the first system in shadow mode on live data before it acts, and gate what it can do.
- Set a performance gate at month six. If the number has not moved, stop, and you have lost one quarter, not a platform migration.
Rent for experiments. Build for the operation. And be precise about the asset you are paying for: a subscription ends, a model of your business appreciates.