The AI proof-of-concept was fast, cheap, and impressive. So the programme feels affordable. The demo did its job a little too well.
A pilot and a production rollout have almost nothing in common financially. The pilot ran on a handful of users, a slice of data, one team's enthusiasm, and a free or heavily discounted tier. Production is the opposite of all four. At scale, the per-query cost stops being a rounding error and becomes a monthly line that recurs for as long as the thing is switched on.
That's before the parts the demo never showed: the integration into systems that weren't built for it, the data plumbing, the security review, the monitoring, and the people needed to keep the output honest. The cost didn't just grow. It changed shape, from a one-off experiment into an operating expense that arrives every month, forever.
This is where the CFO walks in, and rightly so. The strategy named the ambition. Someone still has to name the payback, the unit economics at full volume, and who owns that recurring line when the pilot's sponsor has moved on to the next thing. "Why now" had a confident answer. "What does it cost at ten thousand users" often doesn't.
So the questions worth asking before you scale aren't about the model. What is the marginal cost per unit of real work at full volume. What recurs versus what was one-off. What is the payback, and against which honest baseline. Who owns the bill in year two.
The pilot is cheap by design. That is what makes it a pilot. The number that decides whether this was wise is the one that turns up after it ships.