Which activities should the cost audit inspect?
Inspect pre-launch setup, daily manual checks, and post-launch platform ownership against the categories below using real observed task logs.
| Activity | Evidence to collect | Where to count it |
|---|---|---|
| Data cleanup and integration | Internal engineering hours, contractor scopes, and gateway fees | Upfront setup plus ongoing upkeep |
| Review and correction | Actual human minutes across all runs including failures | Deduct directly from gross time saved |
| User training and rollout | Workshop attendance and playbook drafting hours | Setup; ongoing onboarding in monthly ops |
| Exceptions and incidents | Manual fallback and cleanup minutes on bad outputs | Task effort or monthly ops, once only |
| Evaluation and maintenance | Named steward and monthly prompt/QA check logs | Monthly recurring costs |
| Migration or retirement | Export tooling, prompt rewrites, and wind-down effort | Dedicated lifecycle exit line |
How do you avoid double-counting review costs?
Route every minute of staff oversight to exactly one line in the spreadsheet: if draft checking trims your net time-saving rate, never deduct those same review hours a second time inside monthly AI spend.
Keep clean functional boundaries: checking individual outputs sits inside task handling time, whereas centralized monthly QA sampling or prompt tuning sits under recurring operating costs—even when one senior analyst handles both.
Report external cash invoices and internal employee capacity side by side rather than lumping them together, since only supplier invoices directly draw down cash reserves.
How can omitted work change the business case?
Unaccounted QA and upkeep hours easily turn a rosy deck positive-to-negative. Take a team projecting 40 gross saved hours monthly at $40 per hour and 50% capacity utilization, showing $800 monthly value—yielding $600 net after a $200 tool subscription.
Once time logs uncover 15 hours spent verifying facts and editing drafts, net savings shrink to 25 hours worth $500 under the same assumptions. Add eight hours monthly at $40 ($320) to maintain prompts and review logs: recurring spend jumps to $520, flipping monthly net value to −$20 with zero payback possible.
Catching those omitted QA and upkeep hours upfront protects budgets from funding structurally unprofitable deployments.
Why should evaluation remain in the operating budget?
Production accuracy drifts as internal templates change, upstream schemas evolve, and model providers ship new checkpoints. NIST’s AI RMF explicitly couples pre-launch testing with continuous operational monitoring.
Fund recurring hours for golden evaluation datasets, error triage, and regression sign-offs before pushing updates live—a single launch QA pass cannot catch silent regressions six months later.
What should you do when an important cost is unknown?
Time a short pilot sample or enter conservative high/low cost assumptions instead of leaving blank budget cells that calculate as zero.
Tackle the single cost assumption most capable of flipping ROI negative: clock QA time on ten real tasks, get binding connector pricing, and price one knowledge-update cycle—then calculate the exact break-even ceiling before expanding the pilot.
Frequently asked questions
Should existing employees’ time count as an AI cost?
Yes in economic planning, because internal hours carry real opportunity cost—simply separate internal capacity hours from incremental vendor cash payments on finance reviews.
Is a contingency percentage enough to cover hidden costs?
Contingency buffers help absorb API spikes, but they never substitute for mapping actual cleanup, QA, and maintenance tasks upfront.
Should I include the whole IT department in AI costs?
No—charge only incremental IT and security hours actually consumed by the workflow or a proportional share of shared enterprise infrastructure.