Calculating AI project ROI: costs, benefits and decision criteria

The ROI of an AI project is not simply the cost of a model compared with staff hours. Start from a measured baseline, define the useful outcome, include development and operating costs, then verify what actually changes. The calculation should support a decision to continue, adjust or stop. It should not turn uncertain assumptions into a financial promise.

1. Describe the baseline

Choose a precise scope: one request category, team, channel and representative period. Measure the current process before assuming what AI will change.

Consider a fictional team that prepares 800 cases each month. Each case takes an average of twelve minutes of research and four minutes of rework. Those averages need context: some simple cases take three minutes, while exceptions involve several people.

Baseline element	Measure to collect	Possible source
Volume	Cases actually completed per period	Business application
Active time	Research, entry, review and correction	Observation or timed sample
End-to-end time	Receipt to approved outcome	Process timestamps
Quality	Rework, errors and abandoned cases	Quality review or tickets
Exceptions	Cases handed to another person	Work queue

An incomplete baseline can still be useful when its limitations are recorded. It simply cannot support conclusions beyond what was observed.

2. Define one unit of value

Choose an outcome that can be counted consistently before and after: a correctly prepared case, an approved response, a document found with its source or a request classified without rework. Generated text and model calls measure technical activity, not business value.

Connect the desired effects to that unit: less active time, shorter turnaround, freed capacity, fewer errors or more consistent service. Not every effect becomes a financial saving. Freed time may allow the team to handle more requests or improve service without reducing expenditure.

3. Include the full cost

Initial costs include discovery, integrations, interface work, tests, deployment and change support. Recurring costs include infrastructure, AI processing, external services, human oversight, support and improvements.

Use three scenarios rather than one figure: cautious, central and high-volume. The AI agent budget guide explains these cost categories and the assumptions to document.

Cost	Assumption	Frequency	Evidence to obtain
Development	Scope and integrations included	One-off	Proposal and acceptance criteria
Processing	Volume, input size and retries	Variable	Pilot measurement
Human review	Review rate × average duration	Recurring	Timed sample
Operation	Monitoring, incidents and maintenance	Recurring	Responsibilities and history
Improvement	Business and technical changes	Periodic	Prioritised roadmap

4. Build the calculation without hiding assumptions

For a chosen period, use a calculation that others can follow:

Net value = observed benefits − full costs

ROI = net value ÷ full costs

As an illustrative example, suppose a pilot saves four minutes of active time on 600 approved cases each month. That creates 40 hours of monthly capacity. Before converting those hours into money, establish how that capacity is actually used: additional volume, shorter lead time or expenditure genuinely avoided. Then subtract review time, corrections, operation and the allocated initial cost. This is neither a forecast nor a customer result.

Record the source and sensitivity of each assumption. If the proportion of usable cases falls from 75% to 40%, does the outcome change enough to reverse the decision? This is often more useful than showing ROI to two decimal places.

5. Decide with more than one criterion

ROI is one dimension. A solution can appear profitable while introducing an access risk, an operational dependency or a poorer experience. Conversely, an initial stage may be justified by the learning needed before any reliable calculation is possible.

Decision	Expected signal
Continue	Observed value, acceptable quality and sustainable costs
Adjust	Useful outcome but volume, review or scope assumptions need correction
Pause	Data or access is insufficient for a fair test
Stop	Outcome is worse than the baseline or risk cannot be controlled within scope

Use the AI product discovery playbook to define the first test. When several delivery paths are possible, also compare whether to build or buy an AI solution.

Frequently asked questions

Should every benefit receive a monetary value?

No. Separate monetisable effects, operational gains and qualitative benefits. Record each category instead of assigning an arbitrary amount to everything.

What period should the calculation cover?

Choose a period aligned with the decision cycle and the likely life of the scope. Also show the break-even point and the costs beyond that period.

Can ROI be calculated before a pilot?

You can create a forecast model, but its inputs remain assumptions. The pilot should gradually replace those assumptions with measurements that are comparable with the baseline.