An AI pilot can save a few minutes on a task and still lose money after review, corrections, setup, and exceptions. A useful return-on-investment worksheet starts with the current process, counts the whole attempted workload, and calculates the break-even point without assuming that every predicted saving becomes cash.
This is a proposed Inquory worksheet for a small service business. It is educational, not accounting or investment advice. The example is fictional, excludes taxes and financing, and does not promise that an AI tool will reduce costs. Research and drafting were AI-assisted; see how Inquory uses AI.
Start with one decision
Write the decision at the top of the sheet: continue the limited pilot, revise it, or stop it. Name one task and one eligible case definition. “Use AI across the office” is too broad. “Draft a staff-reviewed summary for complete, typed job notes” is testable.
Keep the manual process as the baseline. Measure the same work, for the same period, with the same staff role and case mix. If the pilot quietly excludes the difficult cases that the baseline must handle, show those exclusions instead of claiming an overall comparison. The AI pilot measurement guide explains how to define eligible cases and keep every attempt in the denominator.
NIST's voluntary AI Risk Management Framework Core recommends choosing measures tied to the most significant risks, documenting test sets and metrics, comparing performance with benchmarks, and recording limits on generalization. This worksheet uses that approach for a business decision. NIST does not endorse these formulas or any vendor.
Copy the input worksheet
Use one consistent period, such as four weeks. Record observed values where possible and label estimates. Keep vendor charges in their billed units, then convert them to the worksheet period.
| Input | Symbol | What to enter |
|---|---|---|
| Eligible cases attempted | N | Every case sent through the pilot, including failures |
| Baseline minutes per case | Mb | Median or average staff time under the current process; state which |
| Pilot handling minutes | Mp | Staff time to prepare, review, and finish an ordinary pilot case |
| Exception minutes | Me | Staff time to investigate and repair an exception |
| Exception cases | E | Cases that required repair, fallback, or reconciliation |
| Loaded labour cost per hour | L | Internal planning rate chosen by the business |
| Fixed pilot cost for the period | F | Setup allocation, subscription minimums, training, and fixed support |
| Variable technology cost per case | V | Usage, integration, and other case-linked costs |
| Fixed measurable benefit | B | A defensible benefit for this period that does not change with case volume; enter zero if none |
Loaded labour cost is an internal planning input, not necessarily an employee's wage or an immediate cash saving. Decide which payroll, benefit, overhead, and contractor components belong in it with the person responsible for the business's finances. Do not publish an employee rate or confidential vendor term in a shared worksheet.
Calculate the observed period
First calculate baseline labour cost:
Baseline labour cost = N × Mb ÷ 60 × L
Then calculate pilot labour cost:
Pilot labour cost = (N × Mp + E × Me) ÷ 60 × L
Calculate total pilot cost:
Total pilot cost = pilot labour cost + F + (N × V)
Finally calculate net value and the ROI ratio:
Net value = baseline labour cost + B − total pilot cost
ROI ratio = net value ÷ total pilot cost
Show the currency, time period, inputs, and rounding. A positive ratio means the quantified benefits exceeded the counted pilot cost for this worksheet period. It does not prove future savings, revenue growth, or cash available to spend. A negative result can still identify a repairable bottleneck, but it does not become a success by changing the label.
Work a fictional example
Suppose a four-week pilot attempts 160 eligible cases. The manual baseline takes 12 minutes per case. Ordinary pilot handling takes 7 minutes, 16 exception cases need another 18 minutes each, and the internal loaded labour rate is CAD $42 per hour. Fixed pilot costs allocated to the period are CAD $420, variable technology cost is CAD $0.35 per attempted case, and fixed measurable benefit B is CAD $0.
| Calculation | Fictional result |
|---|---|
| Baseline labour | 160 × 12 ÷ 60 × CAD $42 = CAD $1,344 |
| Pilot labour | (160 × 7 + 16 × 18) ÷ 60 × CAD $42 = CAD $985.60 |
| Fixed plus variable technology | CAD $420 + (160 × CAD $0.35) = CAD $476 |
| Total pilot cost | CAD $985.60 + CAD $476 = CAD $1,461.60 |
| Net value | CAD $1,344 − CAD $1,461.60 = −CAD $117.60 |
| ROI ratio | −CAD $117.60 ÷ CAD $1,461.60 ≈ −8.0% |
The draft saves ordinary handling time, yet the period remains below break-even after exception work and technology costs. That is useful evidence. The next question is whether a narrower case definition, fewer exceptions, or lower fixed cost can improve the result without weakening review or excluding costs from the sheet.
Find the break-even volume
When per-case savings are positive, calculate contribution per attempted case. In this simple break-even equation, B is a fixed benefit for the worksheet period and is independent of N:
Labour contribution per case = (Mb − Mp − exception rate × Me) ÷ 60 × L
Net contribution per case = labour contribution per case − V
Break-even cases = (F − B) ÷ net contribution per case
If a measurable benefit occurs for every case, include its per-case amount in net contribution instead. If benefit changes with volume in another way, define it as B(N) and solve N × net contribution + B(N) − F = 0; do not subtract it as though it were fixed.
In the fictional example, the exception rate is 16 ÷ 160, or 10%. Labour contribution per case is (12 − 7 − 0.10 × 18) ÷ 60 × CAD $42, which is CAD $2.24. After the CAD $0.35 variable cost, net contribution is CAD $1.89 per case. With CAD $420 in fixed cost and B = CAD $0, the simple break-even estimate is about 223 attempted cases.
Round up to a whole case. If net contribution is zero or negative, volume cannot produce break-even under those assumptions. Fix the process or stop; multiplying a loss does not create a gain. If the business has capacity limits, seasonal demand, or tiered vendor pricing, calculate scenarios rather than extending a straight line beyond the observed range.
Run three scenarios without hiding uncertainty
Create conservative, observed, and favourable columns. Change only inputs you can explain. The conservative case might use higher exception time and the full fixed cost. The observed case uses the actual pilot record. The favourable case can use an attainable repair target, not perfect accuracy or zero review.
| Scenario question | Record explicitly |
|---|---|
| Volume | Is the case count observed, contracted, forecast, or merely possible? |
| Time | Was it timed, estimated after the fact, or copied from a vendor claim? |
| Exceptions | Which failures count, and are abandoned cases retained? |
| Costs | Which setup, review, integration, usage, and support costs are excluded? |
| Benefits | Is the value realized in the period, or only a capacity estimate? |
Time released from a task may become faster response, spare capacity, training time, or no realized financial benefit. Label it accordingly. Do not turn saved minutes into new revenue unless the business observed and can attribute that revenue using an agreed method. The live AI automation ROI calculator can help estimate an opportunity from planning assumptions; this worksheet serves a different purpose by recording observed pilot handling, exceptions, and costs. The two tools do not use identical formulas, so keep their results labelled and do not substitute one for the other.
The Office of the Privacy Commissioner of Canada's generative-AI principles advise organizations to establish that a generative-AI use is necessary and proportionate, likely to be effective for the specified purpose, and preferable to more privacy-protective alternatives where appropriate. A positive spreadsheet result does not answer those questions or establish legal compliance.
Make the decision reviewable
Attach the input definitions, raw attempt count, exception log, timing method, formula version, and reviewer to the result. Record who can stop the pilot and how the team returns to the manual process. Recalculate after changes to the task, case mix, model, review step, vendor terms, integration, or staffing assumption.
Continue only if the observed result and operational controls meet the predeclared threshold. Revise when a specific bottleneck has a bounded test. Stop when critical errors remain unresolved, the comparison is not trustworthy, or realistic net contribution stays at or below zero. The worksheet should make an inconvenient result easier to see, not easier to explain away.