How do you measure the commercial return from AI?

An AI tool can halve the time a task takes without reducing what the business spends. It might still be worth paying for. I want the business case to be clear about what we’re getting.

Quicker customer resolution, better product information and room to deal with an overloaded team’s backlog can all be worthwhile. You don’t need to dress them up as payroll savings. You do need to show what improved and what it cost.

I’d separate the cash return, the improvement in the work, and the demands on the people keeping it running. The example below works through the cash calculation. It is one part of the decision, not a universal test of whether AI is worth using.

Count the job as it happens now

Start with one recurring piece of work. Record its volume, quality and total effort across everyone involved, including preparation, checking, corrections and exceptions. Keep elapsed waiting time separate from staff time: a faster response and less work are different benefits.

Ask whose time changes. Saving a junior employee an hour while adding work to your busiest specialist may still be worthwhile. Use the actual roles, costs and available capacity when assessing that trade. Human review takes attention, and someone needs to make room for it.

Compare AI with a better process or a capability you already pay for, using the same outcome and cost boundaries. Beating a broken process doesn’t establish that AI is the best option.

Where practical, test comparable work with and without the change, including difficult cases and ordinary users. Agree quality measures before seeing the results. A simple before-and-after comparison can be affected by seasonality, product mix, staffing and other changes. If you can’t isolate AI’s contribution, keep that uncertainty in the claim.

Put some numbers around it

Consider a retailer preparing 1,200 product records each month from approved supplier information. Every figure here is synthetic, not a client result, market benchmark or service quote. Amounts are Australian dollars excluding GST.

For this illustration, assume the same volume and acceptable quality, with these whole-process times:

• Current workflow: six minutes per record, or 120 hours monthly.
• Proposed workflow: three minutes per record, or 60 hours monthly, including preparation, human checking, corrections and allocated routine internal support.
• Capacity released: 60 hours monthly. At an assumed loaded employee cost of $50 an hour, that represents $3,000 of capacity value.

That $50 is a valuation assumption, not evidence of a saving. Now allow for:

• $6,000 of initial cash expenditure for setup, integration and external training support.
• 20 hours of internal setup and training time, representing another $1,000 of staff capacity at the assumed rate.
• $600 monthly cash expenditure for software and external support.

The 20 internal hours still need to come from somewhere. Salaries remain unchanged in this example, so I’ve kept that demand separate from additional cash expenditure. Overtime, backfill or work displaced by the setup would change the case.

These are simplified allowances. For a real decision, cost the data preparation, integration, testing, training, model usage, oversight and support the workflow needs. Include replacement or exit costs over its expected life. Count each item once, with cash costs and internal capacity clearly separated.

What happens to the hours?

If employees complete the same workload faster but payroll and other spending stay unchanged, the business has gained capacity. It has not saved $3,000 in cash. With no other benefit established, the cash position includes the $6,000 setup payment and another $600 of monthly expenditure. There is no cash payback to report.

That doesn’t make the capacity worthless. It could mean fewer late product launches or time for neglected work. The National AI Centre’s ROI guidance also recognises quality, capacity and customer and staff outcomes. Measure those benefits in their own terms where a credible dollar value isn’t available.

Now assume the retailer can use the released hours for work it would otherwise pay a contractor $1,800 monthly to complete, and actually reduces that bill. The staff need the skills and usable blocks of time to take it on. Both sets of work must still meet their requirements, and finance needs evidence that the bill would otherwise have continued.

That gives us a cash benefit. Don’t also add the $3,000 capacity valuation: the contractor saving is how those same hours create value. An hour spread across a dozen interruptions may be less useful than the spreadsheet suggests.

Calculate the return, and show what’s in it

Assume the $6,000 setup payment happens at the start, with the full contractor saving and $600 running cost from month one. Monthly net cash benefit is:

$1,800 avoided expenditure − $600 recurring cost = $1,200.

With that steady monthly benefit, simple cash payback is:

$6,000 ÷ $1,200 = five months.

That’s five months on these assumptions. The work still has to fit into those released hours, and the contractor bill still has to fall.

For the first 12 months:

• Cash benefits: $1,800 × 12 = $21,600.
• Cash costs: $6,000 + ($600 × 12) = $13,200.
• Net cash benefit: $21,600 − $13,200 = $8,400.
• First-year cash ROI on total cash cost: $8,400 ÷ $13,200 × 100 = 63.6%, rounded to 64%.

Here, cash ROI means (cash benefits − cash costs) ÷ cash costs × 100, using setup and running costs over the same period. The 20 internal setup hours remain a separate capacity requirement. A broader economic assessment would account for the value of the work those hours displace. Keep the time horizon and cost basis consistent when comparing options.

If costs or savings start at different times, calculate cumulative cash flow month by month. Don’t divide the setup bill by an eventual steady benefit and call that the elapsed payback period. ACCA’s payback guidance explains why timing matters and why payback alone misses later cash flows and the time value of money.

This simple example excludes tax, inflation and discounting. For a material or multi-year commitment, I’d have finance assess the relevant cash flows over the expected life, including net present value: future net cash flows discounted to today’s value, less the initial outlay. A quick payback calculation isn’t a complete investment appraisal.

See what happens when the assumptions change

Suppose only half the records use the new workflow. At the assumed task times, 600 records at six minutes plus 600 at three minutes takes 90 hours, releasing 30 hours. That holds only if the three-minute total, including its support allowance, holds at lower use. Any fixed support effort must stay in the calculation. Keep the $600 monthly cash cost unchanged in this scenario.

If the retailer can then avoid only $900 of contractor expenditure, monthly net cash benefit falls to $300. Simple payback becomes $6,000 ÷ $300 = 20 months, and first-year net cash benefit becomes ($300 × 12) − $6,000 = −$2,400.

The $900 saving is a separate assumption to verify; contractor commitments may not shrink neatly with hours. The 20-month payback also assumes that benefit and cost continue long enough to recover the outlay. A workflow likely to need replacing sooner doesn’t have that payback case.

If use is lower, find out why. The work may not fit, the eligible volume may be smaller, or the team may lack time and support. “People aren’t adopting it” needs a bit more investigation before it becomes a verdict on the people.

Test review effort, errors, usage costs and delays too. Under the original cost assumptions, average avoided expenditure must reach $1,100 a month to cover the $13,200 first-year cash cost. That is a useful threshold to test against evidence.

Go back and check what changed

Give each benefit an owner, a baseline, an evidence source and a review date. Compare actual invoices, completed work and quality with the forecast. For revenue initiatives, count additional contribution after relevant costs, supported by a credible comparison, rather than treating every new sale as an AI benefit. The ecommerce strategy guide works through that distinction. For a cash forecast, also account for when customers pay and when stock, suppliers and other costs need funding.

Agree the quality and risk limits before the trial. Include the cost of correction and incidents, but a positive return doesn’t make an unacceptable privacy, safety or customer outcome acceptable. The governance guide covers those decisions.

The pilot-to-everyday-work guide connects the measures with adoption and handover. Recheck them when the model, workflow or workload changes. If the benefit doesn’t appear, investigate and decide whether to adjust, limit or stop. Money already spent isn’t a reason to keep adding to the bill.

That’s the discussion I’d want before another rollout: what gets better, who can make it happen, what we’re committing to and what evidence would change the decision. An AI opportunity review can help work through it with your team.

The figures are illustrative, not measured results or predicted returns for your business. Method reviewed on 3 November 2026, including UK government guidance on evaluating AI’s impact and ACCA’s investment appraisal guidance. These support the evaluation methods; they do not validate the example’s assumptions.