Measuring AI ROI: A Worked Example and Worksheet

The finance director puts the renewal invoice on the table: "We paid for 40 AI licences last year. Show me what we got back, in dinars, or we cancel." If you are measuring AI ROI for that meeting, a count of hours saved will not be enough. Saved time becomes value only when the hours go somewhere the business can see: less overtime, a cleared backlog, more cases closed by the same team. Below is a worksheet you can copy, a worked example with illustrative numbers in Jordanian dinars (JOD), and the rule we recommend for every figure in it: say whether it was measured, estimated or assumed.

Why "we saved 2,000 hours" does not answer the finance director

Hours saved say people finished tasks faster, not what happened next. If a claims handler finishes each letter 9 minutes sooner and spends the time on email, the organisation pays for the licence and books nothing.

The UK Treasury's guidance on appraising spending, the Green Book (2026 edition, updated 5 February 2026), draws a line any finance director will recognise. It defines a cash-releasing benefit as one that affects the income or spending of a public body, with efficiency savings that reduce labour costs as the example, and it separates these from benefits that are real but release no cash. The same split works for a company in Amman or Dubai.

So when measuring AI ROI, we recommend two lines in every report: cash value, which goes into the payback calculation, and capacity value, which you report and track but leave out of payback. The trade-off: this understates the return early, and a sound project can look weak at its first review. We accept that, because a finance director who finds one inflated number stops trusting the rest.

Measuring AI ROI starts before the licences arrive

A baseline is a measurement of how the work runs today, taken before the change. Without it, every later figure is a guess. The Green Book asks for a "business as usual" option in every appraisal: the outcome expected if current arrangements continue, used as the benchmark for every other option. For an AI rollout, business as usual is the same team doing the same work without the tool.

For each workflow you plan to measure, record four things over two to four weeks:

• Volume: items per month (files, tickets, reports), taken from the system of record, not from memory.
• Time per item: timed on a sample of real items by the people who do the work. A sample of 30 to 60 items per workflow is a practical start; write down how you picked them.
• Quality: the error or rework rate, from a log that already exists (returned letters, reopened tickets, audit findings).
• Where extra hours go today: overtime, temporary staff, outsourced work or a growing backlog. This is the line that later turns time into cash.

If the licences are already live, time the same work in a group without the tool, or rebuild the baseline from system timestamps for the months before rollout, and label it estimated. Our article on the AI readiness assessment explains how to choose which workflows deserve this effort first.

The worksheet, filled in for a claims team (illustrative)

Illustrative example: a mid-sized insurer in Amman gives an AI assistant to a claims team of 12. The assistant summarises each claim file and drafts the letter to the customer; a handler checks and sends every letter. All numbers are invented. The label column shows how each figure would be obtained in a real worksheet; the hourly costs (JOD 7 loaded cost, JOD 8.75 for an overtime hour) are assumptions.

Sixteen lines, from baseline to payback

The version that goes wrong

Multiply all 165 gross hours by JOD 7 and you get JOD 1,155 a month and a payback of about 3 months. It is wrong twice: it ignores the added review time, and it books as cash 76 hours that went to the backlog or cannot be traced. Report the backlog work as capacity with its own measure, such as claims older than 30 days, and keep it out of payback.

Add one sensitivity line

Run one sensitivity line before you present. The Green Book asks for sensitivity analysis to test what happens if key assumptions do not hold, and warns against optimism bias, the tendency to be too hopeful about them. Here, if the time saved per file is 6 minutes instead of 9, gross time falls to 110 hours and net time to 101. The 80 overtime hours still fit inside that, so payback does not move; the backlog line shrinks instead. Tying value to redeployment makes the case depend less on the stopwatch.

Label each figure: measured, estimated or assumed

• Measured: counted from a system or timed on real work, with the method and dates written down.
• Estimated: calculated from measured inputs plus a stated rule, such as a spot-check rate.
• Assumed: a number you chose, such as a price before the contract is signed. Each assumption gets a named person and a date by which a measurement replaces it.

Vendor dashboards need the same labels. Microsoft's Copilot Business Impact report in Viva Insights (Microsoft Learn page dated August 2026) calculates "Copilot assisted hours" from users' Copilot actions and multipliers taken from Microsoft's research, then converts them to money at a default average wage of USD 72 an hour, which analysts can change. Microsoft calls the hours a general estimate and notes that differences between user groups have not been tested for statistical significance. In your worksheet, that figure is estimated, and the wage is assumed until you replace it with your own.

When the first number to fix is unused licences

Illustrative example: a family-owned retailer in Riyadh bought 40 Microsoft 365 Copilot licences for its head office. The finance director's first question is simpler: who uses them?

The Microsoft Copilot usage report in the Microsoft 365 admin center answers that. According to Microsoft Learn (page dated July 2026), it shows enabled users (people holding a licence), active users (enabled users who tried a user-initiated Copilot feature) and the active users rate, over the last 7, 28, 90 or 180 days, with a per-user last activity date you can export. Suppose it shows 14 active users out of 40 over 28 days. Then 26 seats cost money and showed no use in that period, and no time study changes that. Note that "active" means someone tried a feature: a high rate shows use, not value.

In this case we recommend fixing use before measuring value. Find out why 26 people stopped (no clear task, no training on their own work, documents the tool cannot reach), move or cancel the seats with no workflow behind them, and only then baseline the workflows of the people who use the tool. Our Microsoft Copilot rollout plan shows how to tie seats to named workflows from the start, and our page on AI licences bought but not used covers the recovery.

What to show the finance director each quarter

1. The worksheet for each workflow, labels filled in.
2. Cash value and capacity value on separate lines; only cash in payback.
3. Full costs: licences, training and support time, added review time.
4. Quality before and after. Faster work with more errors is not a saving.
5. Active users against seats paid for.
6. One sensitivity line, and which assumptions get measured next quarter.
7. At renewal, look forward. Setup money already spent is a sunk cost, and the Green Book says sunk costs should not affect decisions.

How to start

If the licences are bought and use is low, start with the 30-Day Adoption Rescue: it measures real use, fixes what blocks it, and sets champions and adoption metrics. If you are still deciding where AI should go, the AI Readiness & Workflow Diagnostic ranks use cases and ends in an investment case; baseline the chosen workflows before rollout. See also our Enablement Lab. If you are unsure which fits, the free AI Exposure Scan is a 45-minute conversation that ends in a written next step.

Related lab: Enablement Lab ↗