Skip to main content
K4M2 AI

Practical guides

How to calculate ROI for an AI initiative

Most AI business cases are built on hours saved multiplied by a loaded hourly rate. That number is almost always wrong, and it is wrong in a predictable direction.

By Subhash Trivedi9 min read

Every AI business case eventually reduces to one arithmetic sentence: this much value, against this much cost, arriving by this date. The difficulty is that both halves are routinely mis-stated, and the error compounds.

The value side is usually inflated by counting time saved as money earned. The cost side is usually understated by counting only the build. What follows is a way to do the calculation so that the answer survives a finance review.

Start with a baseline you already have

An initiative without a baseline cannot produce a return, only an opinion. Before estimating anything, establish what the workflow costs today using data that already exists in your systems.

Volume

How many times a month this decision or task happens. Take it from a system of record, not from a manager's estimate.

Unit effort

Time per instance, and by whom. Include the rework, the chasing, and the second pair of eyes.

Unit error

How often it goes wrong today and what a wrong one costs. This is the line most business cases omit entirely.

Cycle time

Elapsed time from request to resolution, including waiting. Effort and cycle time are different numbers and they pay differently.

If none of those four can be produced within a week, that is the finding. The initiative is not ready to be sized, and the first piece of work is instrumentation rather than AI.

Four ways value actually arrives

Be explicit about which of these you are claiming, because they convert to money at very different rates.

01

Capacity released

The same team handles more volume without growing. This is real, but only if the volume exists and the capacity is redeployed. Hours saved with no redeployment plan is not a return, it is a slack increase.

02

Cycle time compressed

Decisions arrive faster. This pays when speed changes an outcome: a quote won, a claim settled, a shipment unblocked. Otherwise it is comfort.

03

Error rate reduced

Fewer wrong decisions, each with a known cost. This is the cleanest line in a business case because both halves are measurable, and it is the most frequently overlooked.

04

Revenue or margin changed

Higher conversion, better pricing, retained customers. The hardest to attribute and the only one that is unambiguous when it lands.

Hours saved is an input. Only capacity redeployed, cycle time that changes an outcome, errors avoided, or margin moved are outputs.

The UK government's AI adoption research found businesses reporting improved workforce productivity alongside no change in revenue. That is this distinction, observed at national scale.

The cost lines that get left out

Build cost is the line everyone estimates and the smallest of the four over three years.

Inference at real volume

Consumption pricing is cheap at pilot scale. Model cost per decision at expected volume and at three times it, because usage expands as people find new uses.

Evaluation and monitoring

Building and maintaining a held-out evaluation set of real cases, plus the dashboards that detect quality drift. Ongoing, not one-off.

Human review

The escalation path for cases the system should not decide alone. Budget it as a permanent percentage of volume, not a transitional cost.

Process change

Retraining, documentation, policy updates, and the productivity dip while people adjust. This is where most overruns live, and it is not a technology cost at all.

Two structural notes. Run the calculation over three years, because the recurring lines dominate. And model a version where the quality target is missed by ten per cent, because that scenario decides whether the initiative is fragile.

Put it in one line

Annual return equals volume, times the value per instance you are actually claiming, times the share of volume the system can handle unaided, minus annual inference, evaluation, review, and process cost. Divide the build cost by that figure to get payback in years.

Two parameters carry most of the sensitivity. The share of volume the system can handle unaided is usually assumed to be higher than it is, and the value per instance is usually the loaded hourly rate when it should be the marginal one. Vary both first.

The result is not meant to be precise. It is meant to be honest enough that a threshold can be agreed before launch: the figure that would justify scaling, and the figure below which you stop. Both should be written down while nobody is defending anything.

Four tests a number has to pass

Before a business case leaves the room, run it against these.

01

Would finance recognise the currency

If the return is expressed in a unit that does not appear in a management account, it will not be believed later. Translate it now.

02

Is there a named owner for the number

Not the innovation team. The person already accountable for that line, who can also change the process around it.

03

Does it survive the pessimistic case

Ten per cent below target quality, twice the expected inference cost, three months late. If it still clears the threshold, it is a decision. If not, it is a hope.

04

Is the counterfactual honest

Some of the improvement would have happened anyway through process change alone. Estimate that share and subtract it, or someone else will.

We have written about the organisational half of this at more length in AI ROI is an organisational problem, and about the specific places money escapes between pilot and return in the value gap. The measurement discipline itself is part of how we scope work.

A business case that passes these four is usually smaller than the one that walked in. That is the point of doing it.

Further reading

UK Government, AI adoption research ↗Forbes on MIT NANDA, The GenAI Divide: State of AI in Business 2025 ↗McKinsey, The State of AI ↗Stanford HAI, AI Index Report ↗

More from us

AI ROI is an organisational problem, not a technical oneThe real cost of AI isn't the budget. It's the value gap.Why enterprise AI implementations failDiscovery: how we scope work before buildingAI opportunity discovery

Bring us the workflow. We will help you size it.

Discovery produces the baseline, the threshold, and the stopping condition. If the number does not clear, we would rather tell you that than build.