How to Measure AI ROI Before You Buy Any Tools

Why can't you measure AI ROI after you buy?

Most AI business cases are written backwards. A vendor demo looks good, a team estimates it will 'save hours', the tool is bought, and six months later someone is asked to prove the return. By then the old process is gone, the people who ran it have adapted, and the only evidence left is opinion.

The gap shows up in the data. According to IBM's 2025 CEO Study, which surveyed 2,000 CEOs between February and April 2025, only 25% of AI initiatives have delivered expected ROI over the last few years, and only 16% have scaled enterprise-wide. According to Deloitte's 2025 AI ROI survey of 1,854 executives, most organizations reach satisfactory ROI on a typical AI use case within two to four years, and only 6% report payback in under a year.

Deloitte also names one of the reasons directly: siloed platforms make it hard to track before-and-after impact, and AI usually arrives alongside other changes, so its effect is difficult to isolate. A baseline captured before purchase fixes both problems. It gives you the 'before', and it pins the measurement to one workflow so the effect is attributable.

What should an AI ROI baseline include?

Two rules make the baseline credible. First, observe rather than interview. People describe the process they think they run; they rarely remember the copy-paste between two systems, the side spreadsheet, or the chase message sent three times a day. Second, capture enough time to see variation. A single good week will overstate the opportunity, and a single bad week will overstate the pain.

This is where Mitori starts. The Mitori Audit runs for 30, 60 or 90 days, with consent, observing real work across the tools teams already use: email, spreadsheets, chat and business systems. The output is an observed baseline for each workflow, which becomes the reference point every later ROI figure is measured against.

• Time per task: the hands-on minutes a person spends on one unit of work, measured across real cases rather than estimated in a meeting.
• Volume: how many units arrive per week or month, and how much that swings by season, client or shift.
• Rework: how often work is redone, corrected or sent back, and how long each correction takes.
• Handoff delays: the elapsed time work sits between people, queues and tools. This is often larger than the hands-on time and is where cycle-time gains come from.
• Exception and error rates: the share of cases that fall outside the standard path, and what an error costs when it reaches a customer, a payer or the ledger.

Read next:See how the Mitori audit works →Where should a company start with AI? →

How do you calculate return on AI investment before implementation?

The formula: monthly net value = (volume × addressable share × time per task × loaded hourly cost) + rework and error cost avoided − review time cost − run costs. Risk-adjusted value = monthly net value × confidence. Payback in months = one-time effort ÷ risk-adjusted monthly net value.

Here is a worked example. It is a hypothetical illustration with made-up round numbers, not a Mitori result or a customer outcome. A finance team handles 2,000 invoice exceptions a month at 12 minutes each, which is 400 hours. At a loaded cost of $50 an hour, that work costs $20,000 a month.

Assume an agent can handle 30% of exceptions in the low case, 45% expected and 60% high. In the expected case it handles 900 items, freeing 180 hours, worth $9,000 a month. Each handled item still needs two minutes of human review, which is 30 hours or $1,500. Run costs are $1,500 a month. Net value is $6,000 a month. With $60,000 of one-time effort, simple payback is 10 months.

Now apply a 0.7 confidence factor. Risk-adjusted value falls to $4,200 a month and payback stretches to about 14 months. The low case nets $3,500 a month (about 17 months to payback); the high case nets $8,500 (about 7 months). That spread is the honest answer, and it is far more useful to a CFO than '$108,000 a year in savings', which is the expected case before review time, run costs and confidence are counted.

One more caution: freed hours are only cash if capacity is redeployed, hiring is avoided or overtime falls. Write down which one you expect, because that is what the realized review will test.

1. 1Pick one workflow and define the unit of work: one invoice exception, one referral, one shift fill, one claim.
2. 2Capture the observed baseline: volume, time per task, rework, handoff delay and exception rate.
3. 3Estimate the addressable share the agent can handle as low, expected and high cases, not a single figure.
4. 4Convert time to cost using a fully loaded hourly rate, and add the cost of rework and errors avoided.
5. 5Subtract recurring costs: licenses, usage, monitoring and the human time spent reviewing agent output.
6. 6Apply a confidence factor that reflects data quality, integration readiness and how stable the workflow is.
7. 7Divide one-time effort (build, integration, change management) by risk-adjusted monthly net value to get payback, and set the date you will check it.

What separates a weak AI ROI case from a strong one?

What questions should a CFO ask before approving an AI tool?

These questions also sort the order of investment. When several workflows are scored the same way, the one with high savings, low effort and high confidence goes first, and the rest wait for more evidence.

• Was the baseline observed in our own workflow, or estimated from interviews or a vendor benchmark?
• What are the low, expected and high cases, and which assumption moves the result most?
• What does it cost to run each month, including the human time to review agent output?
• What one-time effort is required for integration, data cleanup and change management?
• How confident is the estimate, and what would raise or lower that confidence?
• Where does the value land: redeployed capacity, an avoided hire, lower overtime or faster cycle time?
• What happens when the agent is wrong, who approves exceptions, and what does an error cost?
• When will we compare realized results to this forecast, and what result would make us stop?

Read next:How CFOs should decide what to automate first →Build a board-ready roadmap in 30, 60 or 90 days →

How do you track forecast vs realized ROI after rollout?

Mitori builds this loop in. In the roadmap, each automation is scored on savings, effort and confidence, then sequenced into deployment waves with governance checkpoints, including approvals and escalation. After Mitori builds the agents, which start from the observed workflows, Control monitors them: it logs every action and tracks drift against the observed baseline, so realized value can be read against the forecast instead of argued about.

• Agree checkpoints in advance, for example 30, 90 and 180 days after go-live.
• Measure the same unit of work the same way, so the comparison is like for like.
• Track review time and exception volume, because rising human correction quietly erodes the return.
• Watch for drift: a workflow that changes after launch can make the original forecast stale.
• Feed what you learn back into the next case, so confidence scores improve with every deployment.

Read next:Why AI pilots fail →How Mitori handles governance →

Where should you start if you have no baseline today?

You do not need a perfect enterprise-wide model to begin. You need one workflow measured well, one honest range and one review date. That first case teaches the organization how to build the next one.

If you want a quick read on whether your data, owners and processes are ready to support that kind of measurement, the readiness assessment is a useful first pass. If you want the observed baseline and a scored roadmap, an audit scoping call is the next step.

Read next:How to deploy AI in your business →Take the readiness assessment →AI consultants vs an AI audit vs DIY →

Frequently asked questions

Sources

• IBM Newsroom, "IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles" (May 2025)
• Deloitte, "AI ROI: The paradox of rising investment and elusive returns" (October 2025)

Related reading