Checklist · 4 minute read
AI Budget Review Checklist: Defending the Line Item
AI spending now competes with other operational costs, which means the review needs attribution, full cost including review hours, evidence of return against a baseline, and honest forecast assumptions. Programmes without a baseline taken before deployment struggle here regardless of how well they work.
AI spending is moving onto operational budgets where it competes with everything else in the department. This checklist covers a defensible review, drawn from FISTA Solutions' AI enablement delivery work.
What does the review need?
Six inputs, all of which take time to assemble if they do not exist.
| Input | Why it is needed |
|---|---|
| Cost attribution per feature | Locates the spend |
| Full operating cost | Makes the figure credible |
| Pre-deployment baseline | Enables return to be shown |
| Measured outcome | The return itself |
| Forecast with assumptions | Supports the ask |
| Cut list | Prevents arbitrary reduction |
Attribution
Locate the spend before defending it. See LLM cost control checklist.
- Model spend attributed per feature and workflow
- Infrastructure costs attributed
- Vendor subscriptions allocated to the systems they serve
- Cost per task computed for each significant operation
- Trend over at least three periods
- Unattributed spend identified and investigated
- Attribution methodology documented
Full operating cost
The figure that survives scrutiny. See the operating cost of intelligence.
- Human review hours measured and costed
- Evaluation maintenance effort costed
- Corpus and content upkeep costed
- Monitoring and on-call load accounted for
- Periodic revalidation after model changes included
- Support and training effort included
- Total compared against the model bill to show the ratio
Return evidence
Requires a comparison, which requires a baseline. See how to calculate AI ROI.
- Baseline metrics from before deployment available
- Current metrics measured the same way
- Change attributed carefully, with confounders noted
- Return expressed in the units the business uses
- Systems with no baseline identified honestly
- Qualitative benefits stated separately, not mixed into the number
- Methodology stated so the figure is interpretable
Forecast
Assumptions stated, because they are where forecasts go wrong.
- Volume growth assumption stated
- Price change assumption stated
- Review rate assumption stated and justified
- Planned efficiency work and its expected effect
- New systems planned and their cost
- Sensitivity to the main assumptions shown
- Comparison against last period's forecast accuracy
Efficiency opportunities
Show the work already identified.
- Routing opportunities identified with expected saving
- Caching opportunities identified
- Prompt and context size reductions available
- Batch conversion candidates identified
- Vendor renegotiation opportunities noted
- Effort required for each estimated
- Priority order proposed
Cut list
Prepare it before you are asked for it.
- Systems with usage but no measured outcome listed
- Exploratory work with no decision produced listed
- Duplicate or overlapping tooling identified
- Consequence of each cut stated
- Systems that must not be cut, with the reason
- Order of cuts proposed
- Reversibility of each cut noted
What are the most common failures?
Presenting the model bill as the cost. No baseline, so no return. Forecasts with hidden assumptions. And arriving without a cut list, which means the cut is decided by someone with less information.
Who should own this?
The business owner of the AI programme presents; finance provides the framework; system owners supply their own numbers. A review assembled centrally without system owners produces figures nobody will defend.
How often should it run?
Quarterly alongside the operational budget cycle, with attribution maintained continuously. Annual-only review means problems are found a year after they started.
What evidence should it produce?
The attribution report, the full cost model, baseline and current outcome metrics, and the forecast with its assumptions. That package is what makes the ask defensible.
What if there is no baseline?
Say so, and establish one now for the next comparison. Reconstructing a baseline after the process changed is not credible and attempting it damages the rest of the report.
Where a comparable process still runs manually, that can serve as a contemporaneous comparison. Otherwise, state the limitation plainly and use the qualitative evidence you have. See why AI budgets are moving to operations.
What should you do first?
Attribute last month's model spend to features. The distribution is usually surprising and it is where the budget conversation will start.
How FISTA Solutions helps
FISTA Solutions builds and operates production AI systems through AI agents, AI enablement, and forward deployed engineering: spend attributed per feature, full operating cost including review hours, and baselines captured before deployment so return can be demonstrated later, decisions documented with their reasoning, and handover that leaves your team able to maintain what was delivered. The record is 150+ projects for 50+ companies across 12+ countries.
To adapt this checklist to your environment, message FISTA on WhatsApp, or read how to calculate AI ROI.
Share-ready article cover
Download the generated social format.
Clear answers
Questions raised by this field note.
Straightforward guidance for evaluating scope, fit, and the next step.
01What does a defensible review contain?
Cost attributed per feature, full operating cost including people, measured return against a pre-deployment baseline, and a forecast with stated assumptions. Anything less invites arbitrary cuts.
02Why is the baseline so important?
Because return is a comparison. Without numbers from before the system changed the process, no improvement can be demonstrated, and the spending looks unjustified however well it works.
03What is usually missing from the cost?
Human review time, evaluation maintenance, and corpus upkeep. Together these frequently exceed the model bill, and omitting them produces a figure that does not survive scrutiny.
04What assumptions should be stated?
Volume growth, price changes, review rate, and any efficiency work planned. A forecast that assumes review disappears or prices fall should say so explicitly.
05What should you cut first?
Systems with usage but no measured outcome, and exploratory work that has not produced a decision. Identify these before the conversation, because the alternative is a proportional cut across everything.
Continue exploring
Related capabilities
Start with the hard problem
Need the outcome owned, not merely analyzed?
Tell us where delivery is constrained. We’ll map the fastest credible path from intent to verified production.