The honeymoon is over. Through 2024 and 2025, enterprises funded AI on faith. In 2026, finance is asking a blunt question: what did we get for it? And the honest industry answer is uncomfortable — only a minority of organisations report significant returns from their AI and agent programmes, even as spending climbs.
That gap between spend and proven return is exactly why a discipline is emerging that we'd call the FinOps of AI: treating AI cost and value with the same rigour cloud spend finally earned. The firms that master it won't just save money — they'll be the ones still funded next year.
Why AI economics are slippery
AI cost behaves unlike traditional software. Three things make it hard to pin down:
- Usage-based and variable. You pay per token, per call, per GPU-hour. Costs scale with success — a popular feature can quietly become an expensive one.
- Hidden in the margins. The model API is often the small line item. Retrieval infrastructure, evaluation, human review, and rework frequently cost more than inference.
- Value that's real but indirect. Hours saved, errors avoided, and faster cycle times are genuine — but only if someone measured the baseline before you started.
Without deliberate measurement, you end up with a large, growing bill and a vague sense that "AI is helping." That's not a position finance will fund twice.
Measure the number that matters
The fix starts before the build, not after. For every AI initiative, agree one primary business metric and capture its baseline:
- Hours removed from a process
- Cost per case or per transaction
- Cycle time from request to resolution
- Error or exception rate
Then track the fully-loaded cost against it — inference plus infrastructure, evaluation, and human oversight. The ratio of those two numbers is your real ROI. Everything else is anecdote.
Control cost by design
Once you can see the economics, most of the savings are architectural — the same right-sizing we wrote about recently:
- Match the model to the task. Don't pay frontier prices for work a small model does just as well.
- Cache and reuse. Many requests repeat; answering the same question twice is pure waste.
- Set budgets and alerts per use case, so a runaway cost surfaces in hours, not at month-end.
- Kill what isn't working. The strongest FinOps move is often stopping a project that never moved its number — freeing budget for one that will.
Governance and ROI are the same conversation
There's a neat overlap here: the controls that keep AI safe also keep it affordable. Least privilege, evaluation, and a named owner with a metric are governance controls — and they're also exactly what stops an unmeasured system quietly burning money. As we argued in our governance playbook, accountability isn't a tax on AI; in 2026 it's how you protect both trust and the budget.
The bottom line
AI in 2026 has to earn its place on the balance sheet like everything else. Pick one metric per initiative, measure the fully-loaded cost against it, right-size the architecture, and cut what doesn't pay. Do that, and when finance asks "what did we get for it?", you'll have an answer — with numbers.
If you want a clear-eyed view of what your AI is actually returning, that's our kind of problem.