How token costs compound — and the five-point audit to run before any AI rollout scales.

The pilot ran beautifully. Twelve weeks, a controlled environment, forty users. The AI tool processed compliance documents faster than anyone expected. The team was proud.

Then someone approved the rollout to four hundred users.

Three months later, finance flagged an invoice that was eleven times the pilot cost. Nobody in the boardroom could explain why.

The vendor could. Nobody had asked them the right question before signing off on the scale.

If you have been in enterprise technology long enough, you have seen this ambush before. It just had a different name. It was called your cloud invoice.

AI tools are not priced like software licences. Most enterprise AI applications run on a consumption model — every query, every document processed, every response generated draws from a running meter called tokens.

Tokens are the invisible unit of compute cost sitting underneath almost every AI product your team is deploying right now. Your vendor quotes you a per-user fee. The per-user fee is not the number that matters.

Here is the arithmetic that gets missed.

Take a compliance workflow running 200 queries per user per working day — not unusual for document-heavy teams once the tool is embedded in daily work. At roughly 3,000 tokens per query and a blended rate of £0.005 per thousand tokens, each user costs approximately £3 per working day. Forty pilot users across sixty working days: £7,200. Manageable, and broadly in line with the approved budget.

Now approve the rollout to 400 users. Add the usage increase that happens when a monitored pilot becomes standard workflow — people stop being careful with it. That same calculation, at realistic production volume, lands between £70,000 and £90,000 in a single quarter. Before a single invoice has been reviewed.

The mechanism is not a bug. Consumption-based pricing scales with consumption. The pilot was not representative of production usage — it never is.

Your organisation already learned this lesson once. Cloud computing was priced the same way. The teams that controlled cloud costs did not do it by understanding the infrastructure. They did it by building governance around the meter before it ran. AI is the same problem with a different name.

The five-point AI spend audit. Use it on any tool moving from pilot to production — or on any live deployment you have not reviewed recently.

1. Cost per transaction, not per user Ask your vendor: what is the average token count per query or document processed, and what is the blended rate per thousand tokens? Calculate the cost per transaction yourself. If they can only quote per-user pricing, the consumption variable is invisible to you. That is where the surprise lives.

2. Background process inventory Are any automated workflows, scheduled runs, or batch processes running queries outside of direct user activity? These are the equivalent of cloud test environments nobody switched off — they appear on the invoice and nowhere in your usage dashboard.

3. Named spend owner Who is the single named person accountable for this tool's monthly cost? If the answer is "IT" or "the vendor manages that," there is no one watching the meter. Accountability without a name is not accountability.

4. Pre-invoice visibility How will you see spend before the monthly invoice arrives? Does the platform have a real-time cost dashboard with alert thresholds? Monthly billing with no mid-cycle visibility means you are always managing cost after the damage is done.

5. The 3× scenario Run the token calculation at three times your projected usage volume. What does the invoice look like if adoption exceeds forecast? If the answer is "we don't expect that," the budget is not stress-tested. The pilot team never expected it either.

If any of these five have no clear answer, you are not ready to approve the scale-up.“What is the projected token cost at full execution volume, broken down by transactions—not by user licenses?"

If your team cannot answer all three, you are not ready to scale.

If this insight would save a peer from an unexpected budget ambush, consider forwarding this email to your transformation team or project steering committee.