AI Budget Reviews: A 10-Point Checklist Before You Sign the Invoice
Archi3 min read
Contents
- The invoice is the policy document now
- 1. Map every AI line item to a named owner
- 2. Separate per-token costs from per-seat costs
- 3. Price the human review layer
- 4. Check what the contract says about model changes
- 5. Audit data retention clauses against your storage bill
- 6. Budget for re-certification, not just deployment
- 7. Include the exit cost
- 8. Reconcile procurement with engineering
- 9. Put a cap on experimentation
- 10. Schedule the review before the renewal
- What this looks like in practice
The invoice is the policy document now#
Most AI policy arguments end up in a budget line. A vendor contract with an indemnity clause costs real money to negotiate. A model that hallucinates on 3 percent of queries costs real money in refunds and support hours. A retraining run that nobody scheduled costs real money in GPU hours. If you want to know what an organization's AI policy actually is, read its quarterly spend, not its ethics statement.
This checklist is for the people who have to reconcile the two. It assumes you already have a policy. It asks whether the money matches it.
1. Map every AI line item to a named owner#
Pull the last two quarters of spend. Group it into four buckets: vendor subscriptions, API usage, internal compute, and headcount (annotation, review, red-teaming). Assign a person to each line. Any line with no owner is either a shadow deployment or a billing error, and both cost you at audit time.
2. Separate per-token costs from per-seat costs#
Seat licenses are predictable. Token billing is not. A single team experimenting with a long-context model can outspend its entire department's subscription budget in a week. Get the usage curve for each API contract and find the top three callers. If nobody in procurement can name them, you have an uncontrolled spend channel.
3. Price the human review layer#
Every deployed model has a human in the loop somewhere, even if that human is a support agent cleaning up bad output. Estimate the hours. Multiply by loaded salary. This number is usually larger than the model bill and almost never appears in the AI budget request.
4. Check what the contract says about model changes#
Vendors update models. Sometimes silently. Ask each supplier: what notice do we get before a model version changes, and what happens to our evaluation results when it does? A contract that allows unannounced version swaps transfers the cost of revalidation onto you.
5. Audit data retention clauses against your storage bill#
Many AI vendors retain inputs for a defined period. Your side keeps logs, prompts, and outputs for evaluation. Duplicate retention is a storage cost and a legal exposure. Reconcile the two retention schedules. Where they conflict, the shorter one usually wins on cost and the longer one wins on risk, so pick deliberately.
6. Budget for re-certification, not just deployment#
Any model that touches regulated decisions (credit, hiring, medical triage) needs periodic revalidation. Put a number on it. A common pattern is one full evaluation cycle per year plus one out-of-cycle run per model change. If your vendor ships quarterly, budget four runs.
7. Include the exit cost#
Ask what it takes to leave. Export formats, fine-tuned weights, prompt libraries, evaluation datasets. A vendor that will not return your evaluation data has priced your switching cost at whatever it wants. Get the export clause in writing before renewal, not after.
8. Reconcile procurement with engineering#
Engineering teams often have cloud credits that bypass procurement entirely. Ask for a list of active cloud accounts and the AI services enabled on each. The gap between the procurement ledger and the cloud bill is where surprise spend lives.
9. Put a cap on experimentation#
Give each team a monthly ceiling for model experiments, with a simple approval path above it. Caps do not stop research. They make the cost visible at the point of decision instead of at quarter close.
10. Schedule the review before the renewal#
Most AI contracts renew annually with short notice windows. Put a calendar entry 90 days before each renewal date. At that meeting, compare usage to forecast, check the last evaluation results, and decide whether to renegotiate. A renewal that auto-processes is a policy decision made by default.
What this looks like in practice#
A mid-sized company with three deployed models and roughly a dozen experiments might spend low six figures a year on inference, a comparable amount on seats, and more than either on human review. The surprise is rarely the model bill. It is the review layer and the storage of evaluation data nobody deletes.
Run the checklist once. It will produce a short list of items with no owner, a longer list of contracts with weak change-notice clauses, and at least one cloud account nobody remembered opening. Fixing those three things usually pays for the audit.
The point is not to slow AI adoption. It is to make sure the money follows a decision someone actually made.
More articles