NOTE · JULY 2026 · AI INVESTMENT ECONOMICS
The hidden costs
of AI.
> there are eight. the other six decide whether it pays back.
> open each one below.
> est. read: 3 min.
01 / THE ILLUSION
Two costs go in the spreadsheet. Six arrive later.
Almost every AI business case prices the model usage and the compute, signs it off, and moves on. Then the real bill turns up — assembled from six more costs that were never counted, and at least one that keeps growing after launch. Here's the whole bill. Open any line to see what it actually is and where it bites.
02 / THE FULL BILL
Eight cost categories. Most cases count two.
The first two are the ones everyone budgets. The other six are the ones most cases leave out. Click a line to open it.
What you pay per token — every call, plus any hidden reasoning the model does before it answers. The one cost everyone remembers, and the one that behaves least like they expect.
The cloud, GPU and storage to actually run it — whether you're calling an API at scale or hosting models yourself.
Getting your data clean, connected and usable enough that the AI can be trusted with it. The unglamorous work that decides whether anything downstream is any good.
People's time to check, correct and sign off what the AI produces. Automating a task doesn't remove the verification — it moves it to a person.
Wiring AI into the systems you already run — the CRM, the case management, the tools where the work actually happens. A demo lives in a sandbox; value lives in your stack.
Training and support so people actually adopt it — and redesigning the roles around it. The one most cases forget entirely, and the one that most often decides whether any of the rest pays back.
Fixing drift as models change under you and your own needs move. AI isn't a build-once asset — it's a system that decays quietly if no one tends it.
Keeping it safe, compliant and tested — the monitoring, the red-teaming, the checks that stop a quiet failure becoming a public one.
▸ tap any line to expand · at least one keeps growing long after launch
03 / THE TRAP
It scales with success — not failure.
The counter-intuitive part: the better AI performs, the more it gets used — so the bill grows with adoption, not with problems. Your pilot was cheap precisely because barely anyone touched it. The costs that grow after launch — usage, maintenance, guardrails — are the ones a one-off business case is worst at seeing.
04 / THE ONE MOST CASES FORGET
Budgets don't fail on tokens. They fail on people not adopting.
Forecasting spend is still hard — model a range, not a point, and meter from day one. But the largest waste is rarely the compute. It's whole budgets gone on tools people never learn to use well. Change management is the cost most cases forget, and the one that most often decides whether the other seven were worth paying at all.
05 / WHAT A CREDIBLE CASE DOES
Price all eight. Then plan for the bill to grow.
Put every category in the spreadsheet in real pounds — not just tokens and compute, but data work, human checking, integration, change management, maintenance and guardrails. Meter usage from day one, and model a range, not a single confident number. Expect the growing costs to climb as adoption climbs — and read that as the sign it's working, not a failure to control. One honest footnote for the board: when costs spiral unmanaged, access gets rationed — and a budget problem quietly becomes a fairness problem about who gets the good tools.
David Kolb advises leaders on building AI business cases that survive contact with real usage. This note is written to be forwarded — if it's useful to a colleague, send it on.
One thing you can do this week
Open your current business case and mark which of the eight costs you've included.
What it won't tell you
Finding the gaps is straightforward. Estimating the costs — especially the ones that increase with adoption — is a different exercise.
If this is live for you.
> Send me the thing you can't get a straight answer on. I'll tell you whether I'm useful.
Email me →