How to measure AI ROI: formulas, costs, and a worked example

AI ROI is the net financial benefit of an AI initiative divided by its total cost. The arithmetic is simple. The hard part is building a baseline, counting costs that vendors leave out, and separating genuine economic value from activity that merely looks useful.

1. Measure the workflow before AI changes it

Start with a period that represents normal operations. Record transaction volume, labor time, loaded labor cost, error and rework cost, cycle time, conversion or completion rate, and any losses tied to delay. A baseline based on guesses will produce precise but unreliable ROI.

Use medians when a few unusual cases distort the average. Segment work by complexity if the proposed system will handle only the simpler cases. If AI can assist 60% of tickets, do not apply the expected saving to all tickets.

2. Count benefits that can reach the financial statements

Labor capacity

Time saved becomes value only when the business can use it. The team may absorb more volume, avoid planned hiring, reduce overtime, improve service, or redirect staff to higher-value work. State which mechanism applies. Avoid calling every saved minute a cash saving.

Quality and rework

Calculate the number of errors avoided and the average cost of correction, refund, escalation, write-off, or delay. Include the cost of AI errors as a negative benefit. Human review may reduce that exposure, but review time belongs in the cost model.

Revenue and retention

Use contribution margin, not gross revenue, when possible. If faster lead response is expected to improve conversion, model the incremental customers, expected margin, and confidence range. Do not credit the AI project for growth that the sales team cannot plausibly attribute to the changed workflow.

Risk reduction

Risk value can be estimated as the change in probability multiplied by the financial impact. Keep it separate from realized savings because it is expected value, not cash already captured. For high-stakes use cases, document the assumptions and have the relevant risk owner approve them.

3. Build a complete AI cost model

1. 1.Model, platform, software, and per-use fees.
2. 2.Internal and external discovery, design, engineering, and integration labor.
3. 3.Data cleaning, retrieval, labeling, evaluation sets, and ongoing data access.
4. 4.Security, privacy, legal, governance, red-team, and compliance work.
5. 5.Training, change management, support, human review, and workflow documentation.
6. 6.Monitoring, maintenance, incident handling, model changes, and vendor management.
7. 7.Expected cost of inaccurate output, downtime, failed actions, and rework.

Separate one-time implementation costs from recurring operating costs. Then calculate first-year ROI and steady-state annual ROI. This prevents a mature use case from looking unattractive because of its initial build cost, while still showing the cash required to reach production.

Worked example: AI-assisted support triage

Consider a support team handling 50,000 requests per year. The loaded cost is $7.50 per request, so the current annual handling cost is $375,000. A pilot suggests AI can reduce average handling effort by 18% across 70% of requests. Conservative annual labor capacity value is therefore $47,250.

Better routing is also expected to prevent $12,000 in annual rework. Total quantified benefit is $59,250. First-year implementation, integration, software, review, training, and monitoring cost $38,000.

That result is useful only if the organization can turn reduced handling effort into capacity and if service quality holds. The pilot should therefore measure handle time, routing accuracy, reopen rate, escalation rate, adoption, and cost per assisted request.

Run a downside case before asking for budget

Reduce the expected benefit, increase usage and review costs, and delay adoption. If the project still clears the organization's hurdle rate, the business case is robust. If a 10% change in one assumption destroys the return, present the initiative as an experiment rather than a committed savings program.

External benchmarks can help set a range, but they do not replace local measurement. Case Ledger's AI ROI benchmark library provides source and methodology context so you can see whether a comparison is relevant. Use the free AI ROI calculator to test your own volume, cost, improvement, and implementation assumptions.

Move from a spreadsheet to a measured pilot

Approve the smallest pilot that can test the value hypothesis. Keep a comparison group or pre-pilot baseline, set a minimum sample size, and decide in advance what result would justify scaling. The 90-day AI implementation roadmap gives this measurement work a practical delivery sequence.

If you are still deciding what to measure, start with how to create business value with AI and use the AI use-case prioritization guide to rank candidate workflows.