What an AI Pilot Costs Beyond the Model Call

An AI pilot’s model bill is only one part of its operating cost. Before continuing or expanding a selected workflow, account for preparation, review, rejected work, recovery and maintenance over a defined period. Keep cash spending, staff capacity and elapsed waiting separate. A reduction in staff minutes is useful evidence about capacity; it is not automatically a cash saving.

For business owners and digital or IT leaders, the useful question is: what must this workflow consume to produce accepted work, and which missing cost could change the continuation decision?

Our guide to finding AI automation opportunities introduces the need to count more than model calls. This worksheet turns that principle into an operating estimate. The categories and allocation rules below are AMF’s proposed decision aid, not measured pilot results.

Set the boundary before collecting numbers

Write down the operating period, eligible work, acceptance criteria and decision owner. Define “accepted” in terms of a useful business result, rather than whether the model produced an answer.

For an explicitly hypothetical weekly report-preparation pilot, each eligible department supplies approved records for one report pack. The assistant prepares a draft. A person checks its figures and statements against those records and accepts, corrects or rejects the pack. Distribution remains a separate human action; the assistant does not send reports.

The continuation decision might be whether to retain that preparation assistance next month. Record the manual alternative and the review capacity available. A cost estimate alone cannot establish whether the reports are sufficiently useful or dependable.

The U.S. Government Accountability Office’s 2020 Cost Estimating and Assessment Guide summary identifies scope, assumptions, data, sensitivity analysis and updates using actual costs as elements of cost estimation. That is general estimation guidance. It provides neither an AI pilot benchmark nor validation of this worksheet.

Give every cost an owner and an evidence label

Copy the table and replace each owner role with a person’s name. Split rows when quantities use different units or have different evidence status. For every entry add a quantity, unit, evidence location and status: actual, estimate or unknown. An empty cell must not silently become zero.

Use this blank entry for each quantity in the worksheet:

Recovery and maintenance arise in the surrounding workflow as well as in model use; the reference architecture identifies relevant responsibilities.

Keep three views of consumption

Cash records spending within the stated boundary. For a shared subscription, show both the allocation rule and whether continuing the pilot changes the bill. An allocated share and an additional cash expense answer different questions.

Capacity records staff hours by activity. If you value those hours using a documented hourly rate, label the result an allocation. Do not add that value to a cash total that already includes the same salary, overtime or contractor expense. Likewise, do not count a correction once as rework and again as exception handling.

Waiting records elapsed time awaiting inputs, tools or review. Waiting can affect a deadline without consuming an equal amount of staff time. Record active intervention separately; never add waiting hours to staff hours as though they were interchangeable.

A retry can legitimately consume both tool spending and staff effort. Counting those distinct resources is appropriate; recording the same tool charge twice is not.

Show setup separately. If a comparison spreads setup across expected operating periods, disclose that assumption and include only the allocated portion. Do not add the full setup cost again.

Count useful outputs without hiding failures

For the hypothetical reporting week, define:

• E: eligible report packs, counted once each.
• A: packs accepted by the reporting cutoff, with A ≤ E.
• C: cash spending within the stated period and boundary.
• H: staff hours within that same boundary.

Report both E and A, plus rejected and unresolved packs. Keep all in-scope costs in C and H, including work that produced no accepted pack. Retries are attempts, not additional eligible packs.

Cash per accepted pack is C / A; staff hours per accepted pack is H / A. These are separate measures. When A = 0, both ratios are undefined: report the totals and “no accepted outputs,” rather than a zero unit cost.

Test the assumption that could reverse the decision

Use low, base and high assumptions for review effort. If m is review minutes per eligible pack and H_other covers other staff work, estimated capacity is:

H = H_other + (E × m / 60)

Calculate that expression using m_low, m_base and m_high. Record the reason for each assumption; these are scenarios, not forecasts. This version assumes every eligible pack receives review; use the actual reviewed count if some do not.

Separately vary accepted packs using A_low, A_base and A_high, holding other inputs fixed to expose the denominator’s effect. If rejection also increases recovery work, add a combined scenario that changes both. Do not imply that acceptance and effort are independent.

Compare the scenarios with the available budget, capacity and required output quality. If only optimistic assumptions support continuation, identify the missing fact and name the person responsible for obtaining it. Retaining limited assistance, revising the workflow or stopping are all legitimate decisions.

Provenance and review status

Prepared by Agent Model Fit with AI assistance from Codex in source inspection, worksheet development and drafting on October 2, 2026. The worksheet and weekly reporting example are proposed guidance, not observed AMF or customer results. The GAO landing-page summary was inspected on that date; this article does not represent a review of the full guide. Human editorial review by the Agent Model Fit team was completed on October 2, 2026.