The answer at a glance
Measure the time to produce an acceptable result, including review and correction. Apply that difference to the tasks people will actually use AI for. Report recovered hours separately from cash benefits, subtract implementation and running costs, and set a quality threshold before approving a rollout.
Free interactive tool
AI ROI calculator
See the time a workflow could release and the cash it could return. The starting values are a hypothetical example, not a forecast.
No account needed. This calculator does not transmit or store your inputs. All monetary values are in USD.
Your modeled results
Capacity released
- Monthly hours saved
- Monthly capacity value
649.5 AI-assisted tasks/month × 3 net minutes saved/task ÷ 60 = 32.475 hours. Capacity value = hours × loaded hourly cost.
Capacity value is not included in cash benefits or ROI.
Cash return
- Monthly net cash benefit
- Cash benefit − recurring cost
- 12-month net cash benefit
- 12 × monthly net cash benefit − setup cost
- 12-month cash ROI
- 12-month net cash benefit ÷ total cost × 100
- Simple payback period
- Setup cost ÷ positive monthly net cash benefit
Total 12-month cost: $9,600.00 = 12 × recurring monthly cost + setup cost.
Monthly task volume = people × tasks per person per week × 4.33 × adoption ÷ 100. Net minutes = baseline time − AI-assisted time − review and rework. This simple model holds monthly inputs constant for a full year, with setup paid upfront; it excludes taxes, discounting, and ramp-up. It does not establish output quality or prove a cash benefit. Validate both with a measured pilot.
Start with one workflow and one funding decision
Choose a bounded job: summarize a completed support ticket, extract fields from a supplier document, or draft an internal project update. Record where the job starts, what a usable result contains, and who accepts it. A pilot that mixes several jobs makes it difficult to explain which change produced the result.
Write the decision in one sentence: “Should we spend this amount to run this workflow for this group, subject to these checks?” Keep the current process as a comparison. Include a simpler alternative, such as fixing the intake form or using an existing template, if it could address the same bottleneck.
Microsoft’s account of its internal AI value framework distinguishes several forms of business value and asks how recovered capacity will be used. That is a useful distinction for a small pilot too: a time estimate needs an operational destination before it becomes a business outcome.[3]
Measure completed work, including the cleanup
Select examples spanning the work you expect to handle: routine cases, incomplete inputs, and difficult exceptions. Record task type, operator, active working time, review time, corrections, and whether the result passed the acceptance check. Keep the same definition of “finished” in the manual and assisted groups.
When possible, compare different but similar tasks across the same people, or alternate which group starts with the assisted process. Repeating an identical task can make the second attempt faster because the person already knows the answer. Separate setup and training from repeat use, while including both in the investment cost.
NIST AI RMF 1.0 recommends documenting test sets and metrics and evaluating performance under conditions similar to deployment. Use that principle to test the workflow with the review steps and exceptions it will face in practice; this worksheet is not a certification or a substitute for a broader risk assessment.[1]
| Field | What to record | Why it changes the decision |
|---|---|---|
| Task and difficulty | An anonymous ID and a consistent task category | Shows whether one group received easier work |
| Active time | Minutes spent preparing, operating, reviewing, and correcting | Captures the complete labor requirement |
| Quality result | Pass, correction required, or unusable; reason recorded | Prevents fast but incomplete work from counting as success |
| Usage | Eligible tasks and tasks actually attempted with AI | Separates adoption from the size of the team |
| Exception | Failure mode and who resolved it | Reveals work shifted to another person |
Keep recovered capacity and cash benefits in separate columns
Recovered capacity is time available for other work. Multiplying those hours by a loaded hourly labor cost gives a planning value for that capacity. It does not automatically reduce payroll. Name the work that will use the released hours: clearing a backlog, reviewing more cases, or reducing a service delay.
A cash benefit needs a budget-level mechanism and an owner who can validate it. Examples include an approved reduction in paid overtime or a lower contractor invoice. Avoided future hiring is a separate scenario until there is a credible hiring plan that changes. For revenue improvements, use attributable contribution after incremental delivery costs, rather than counting the entire sale.
Do not add the full labor value to cash savings if both describe the same recovered hours. The UK Digital and Data Benefits framework explicitly calls for avoiding benefits counted in more than one place. In this calculator, the cash calculation uses the cash-benefit input only; capacity value is displayed separately.[2]
Worked example: a support-ticket summary pilot
This is a fictional planning example, not a Kona customer result or a product benchmark. Ten staff each complete 20 eligible summaries a week. The manual task takes 12 minutes. The assisted task takes 7 minutes plus 2 minutes of review. At 75% usage and 4.33 weeks per month, 649.5 tasks use the assisted process, releasing 32.475 hours.
At an assumed loaded labor cost of $50 per hour, that capacity has a planning value of $1,623.75 a month. The example budget contains $600 of monthly software, support, and infrastructure cost, plus $2,400 of initial setup. The budget owner can substantiate only $400 per month in reduced paid overtime. That $400 is the cash benefit used below.
| Measure | Calculation | Result |
|---|---|---|
| Time recovered per task | 12 − 7 − 2 minutes | 3 minutes |
| Monthly recovered hours | 10 × 20 × 4.33 × 75% × 3 ÷ 60 | 32.475 hours |
| Monthly capacity value | 32.475 × $50 | $1,623.75; not added to cash benefits |
| Monthly cash change | $400 − $600 | −$200 |
| First-year total cost | $2,400 + 12 × $600 | $9,600 |
| First-year net cash benefit | 12 × $400 − $9,600 | −$4,800 |
| First-year cash ROI | −$4,800 ÷ $9,600 | −50% |
| Cash payback | Monthly net cash is negative | No payback under these assumptions |
Test the assumptions most likely to break
Change one input at a time before trying a combined downside case. If review rises from two to five minutes in the example, the time benefit disappears. At six review minutes, the assisted process consumes additional labor. Keep that negative result in the analysis; replacing it with zero conceals the cost of the workflow.
Usage and cash benefits are deliberately separate inputs. A low-usage case will usually require a fresh estimate of cash benefits as well. Adjust both when the benefit depends on adoption. Likewise, changing the team size does not automatically change subscription costs here: update the monthly budget if seat or usage charges move.
The calculator assumes a full year at a constant monthly run rate. It does not model a gradual rollout, taxes, discounting, or uncertain benefit distributions. For a phased launch, put each month’s costs and benefits into a separate schedule and check the cumulative cash balance. Preserve the simple model as the version you can reconcile to the detailed one.
- Count licenses, usage fees, integration work, training, maintenance, and human review once each.
- Include a low-usage and high-review scenario with revised cash-benefit assumptions.
- Attach the source, date, and owner of every budget or productivity assumption.
- Test output quality separately; a positive financial estimate does not demonstrate acceptable quality.
Finish with a pilot decision and a review date
Use the template below to name the approving owner, scope, budget ceiling, acceptance measure, and next review. Set the threshold before looking at the results. A document extraction pilot might require every accepted record to match the source, with failed records routed for manual completion. The appropriate check depends on the task and the consequence of an error.
Conclude with expand, revise, stop, or collect more evidence. Explain which finding supports that choice and what would reverse it. A small sample with inconsistent results may justify a narrower test instead of either a company-wide rollout or a claim that AI cannot help. Save the baseline log with the recommendation so another person can reproduce the reasoning.
Yours to use
AI pilot business case template
Copy or download the Markdown worksheet. Replace each bracketed field with your evidence, budget, and decision criteria.
Markdown · Opens in any text editor · No signup required
# AI pilot business case
Decision requested: [Approve / revise / stop / extend a bounded pilot]
Owner: [Name and role]
Decision date: [YYYY-MM-DD]
Review date: [YYYY-MM-DD]
## Workflow and alternatives
Task, starting input, and accepted output: [Define]
Participants and exclusions: [Define]
Current process: [Describe]
Simpler non-AI alternative: [Describe]
## Measurement
Baseline sample and task mix: [IDs and selection method]
Manual active minutes per accepted task: [Value; source]
Assisted active minutes per task: [Value; source]
Review and correction minutes: [Value; source]
Eligible tasks per person per week: [Value; source]
People and expected usage rate: [Values; evidence]
Quality measure and threshold: [Definition agreed before the pilot]
Exception handling and human owner: [Define]
## Capacity case
Monthly assisted tasks = people × weekly tasks × 4.33 × usage rate
Net minutes = manual minutes − assisted minutes − review minutes
Monthly recovered hours = monthly assisted tasks × net minutes / 60
Planning value of capacity = recovered hours × loaded hourly labor cost
Work that will use this capacity: [Named backlog or service outcome]
## Cash case
Initial implementation and training cost: [Amount; source; owner]
Monthly total running cost: [Licenses + usage + support + other costs]
Monthly evidenced cash benefit: [Amount; mechanism; budget owner]
Monthly net cash = monthly cash benefit − monthly running cost
First-year total cost = initial cost + 12 × monthly running cost
First-year cash ROI = (12 × cash benefit − first-year total cost) / first-year total cost
Cash payback = initial cost / monthly net cash, only when monthly net cash is positive
Do not add capacity value to cash benefits for the same hours.
## Scenarios and decision
Downside scenario: [Usage, review, cost, and cash benefit assumptions]
Base scenario: [Inputs and evidence]
Quality result and unresolved issues: [Record]
Recommendation and reason: [Record]
Budget ceiling: [Amount]
Stop or reversal trigger: [Observable condition]
Next action, owner, and date: [Record]
Sources
Sources and further reading
[1]
Artificial Intelligence Risk Management Framework (AI RMF 1.0), Measure functionNational Institute of Standards and Technology · 2023-01
[2]
Digital and Data Benefits frameworkGovernment Digital Service and Department for Science, Innovation and Technology · 2026-04-07
[3]
Measuring the impact of our AI investments in IT at MicrosoftMicrosoft Inside Track · 2026-06-04
Editorial method
This guide combines the linked sources with a worked example and a reusable template. Replace example inputs with your own evidence and check the result before making a business decision.
About Kona’s editorial standardsPut it into practice
Adapt this to your business in Workspace
Copy the brief below, open Workspace, and paste it with the inputs you want to use. Review the assumptions and calculations before relying on the output.
Help me prepare an AI pilot business case from the baseline measurements and cost assumptions I provide. Separate recovered capacity from evidenced cash benefits, include review and correction time, identify double counting, show downside and base cases, and finish with a quality threshold, budget ceiling, owner, and review date. Mark missing inputs as unknown; do not invent results or customer evidence.
Keep working on the decision
A decision memo template that makes the trade-off clear
Give a decision-maker a recommendation they can inspect: the alternatives, the evidence, what could go wrong, and when the decision should be revisited.
Vendor evaluation scorecard: a template with a worked example
Put evidence beside every score, separate requirements from preferences, and see whether your preferred vendor survives a change in assumptions.
Standard operating procedure template: a complete onboarding example
A procedure earns its place when another person can follow it, handle the common exceptions, and tell whether the work is complete.