How to measure ROI on AI. Start before you build.
Measure ROI on AI by baselining the workflow before anything is built: how long it takes, what it costs and how often it goes wrong. Set a target, measure the same numbers after launch and report them in figures the business already tracks. Modern Thesis, an AI strategy and implementation firm for non-tech companies, starts every engagement this way.
Why do so many companies struggle to show ROI from AI?
A common reason companies can't show a return on AI is simple: they never measured where they started. In BCG's 2024 survey, 74% of companies had yet to show tangible value from their use of AI, and BCG attributed about 70% of the problems to people and processes rather than technology.1 Among smaller firms, a 2026 Clarify Capital survey of 241 owners found that 54% had invested in AI without setting any defined ROI goal, and 26% had fully recouped what they spent.2 Without a before number, there is nothing to compare the after number to.
Step 1: Which workflows should you measure?
Measure two to four workflows that are frequent, mostly repeatable and tied directly to money. Good candidates share a few traits:
- They happen often, weekly or daily, so small gains add up.
- They follow a pattern most of the time, with a few points where a person has to decide.
- They touch cash, capacity or customers: quoting, invoicing, scheduling, customer requests, reports, proposals.
- Someone complains about them. The work your best people dread is usually where hours go.
Leave out workflows that happen a few times a year, however painful. They rarely pay back the cost of building and maintaining a system.
Step 2: How do you measure the baseline?
Measure three things for each workflow, the same way you will measure them later: time, cost and error rate.
- Time. Record how long one run takes from start to finish, across a sample of 20 to 30 runs. Use timestamps from your systems where they exist. Where they don't, have the people doing the work log start and finish times for two weeks. Multiply the average by monthly volume to get hours per month. Note the elapsed time too, such as days from request to delivery, because customers feel that.
- Cost. Multiply hours by the loaded hourly cost of the people involved: wages plus payroll taxes and benefits. Your accountant can give you that figure. Add software, outside services and overtime tied to the workflow.
- Error rate. Count the runs that needed correction, rework, a second review or a follow-up call, and what each one cost: credits, write-offs, delays or another hour of someone's time.
Write down the date, the sample and the method. Tell the team why you're doing it. You are measuring the work, not grading the people, and they need to hear that from you before you start.
Step 3: How do you set a target metric?
Pick one primary metric per workflow, stated in terms the business already uses, and set a target and a break-even point before you build. "Quotes out within one business day" or "invoices processed per clerk per week" work. "Efficiency up" does not. Then work out what the improvement is worth in a year:
- Hours saved per run, times runs per year, times loaded hourly cost.
- Plus the cost of errors avoided.
- Plus the margin on any extra work the freed capacity lets you take on.
ROI is that annual value minus the annual cost of building and running the system, divided by the annual cost. The break-even point is the improvement at which value equals cost. If the target can't clear break-even comfortably, pick a different workflow.
A hypothetical example, in round numbers. A team prepares 300 quotes a year at 4 hours each: 1,200 hours. At a loaded cost of $60 an hour, that is $72,000 a year. If AI drafts each quote and an estimator reviews it in 1.5 hours, the team gets back 750 hours, worth $45,000 at cost. Whether that is profit depends on what happens to those hours next.
Step 4: How do you measure after launch?
Measure the same workflow with the same method and a similar sample, once the team has used the new system for a few weeks and early fixes are done. Count the review time. If people spend longer checking the output than the system saves, the ROI is negative, and the fix is usually in the design, not the people. Track adoption as well: the share of the work that now goes through the new system. A system that half the team avoids can't return what the baseline promised.
Step 5: How should you report AI ROI?
Report AI ROI in numbers the business already tracks, on one page, on a fixed schedule. Show the baseline, the target, the current figure and the cost to date for each workflow, then the business measures they feed: jobs completed per month, days to invoice, overtime hours, backlog, revenue per employee. Leave out counts of prompts, licenses or tasks automated. Those describe activity. Your board, your banker or a future buyer will ask about results.
Is AI ROI about cutting costs or adding capacity?
AI pays back in two ways: lower costs, or more work done by the same team. Lower costs show up as less overtime, fewer temporary staff, smaller outside service bills, or a role you don't need to backfill when someone leaves. Added capacity shows up as more jobs, more projects, faster turnaround and bids you no longer have to turn down. For most established companies short of skilled people, capacity is worth more, and it keeps your best people doing the work that needs their judgment.
Why doesn't time saved raise profit for firms that bill by the hour?
In a firm that bills by the hour, saved time can lower revenue. If an engineer finishes a 10-hour task in 6 hours and the client is billed for hours worked, the firm bills 6. The client gains and the firm earns less, unless the 4 freed hours go to other billable work. For these firms, measure capacity, not time: projects per person, backlog, utilization and work won that you would previously have declined. Some firms also move routine work to fixed fees, so faster delivery raises margin directly. Time saved is the input. Added capacity is the return.
How Modern Thesis handles this
Every Modern Thesis engagement starts with a strategic assessment that maps how work really moves, measures a baseline of time, cost and error rates on key workflows, ranks opportunities by return and defines the first build. After launch, we report results against that baseline in numbers you already track. If AI won't pay for itself somewhere in your business, we'll tell you.
Questions
How long does it take to see ROI from AI?
It depends on the workflow and on how quickly the team adopts the new system, so measure rather than guess. Plan to take the first after-launch measurement once people have used the system for a few weeks and the early fixes are done. For a sense of pace, MIT's 2025 study found that the top-performing mid-market companies averaged 90 days from pilot to full implementation. Payback comes after that, at a rate your baseline will show.
What if we don't have good data for a baseline?
Most companies don't, and it isn't a reason to wait. A baseline can come from a sample. Ask the people doing the work to log 20 to 30 runs of the workflow over two weeks: start time, finish time, and anything that had to be corrected or redone. Combine that with monthly volumes from your systems. It is rough, but it is a real number, and it beats an estimate nobody can check later.
Should we count time saved as money saved?
Only when the time turns into something the P&L shows. Saved hours become money when you avoid a hire, cut overtime, stop paying for temporary help or outside services, or use the freed capacity for more paid work. If the hours simply disappear into the day, the saving is real for the team but invisible in the accounts. Report time saved as an input, and report what the business did with it as the result.
Which AI metrics should we ignore?
Ignore measures that count activity instead of results: number of licenses, logins, prompts, tasks automated or AI features switched on. They show that AI is being used, which matters for adoption, but they say nothing about whether the business is better off. Keep the report to time, cost, error rates and the business numbers they feed, such as jobs completed, days to invoice or revenue per employee.
Do we need a consultant to measure AI ROI?
No. The baseline method on this page needs a spreadsheet, a few weeks of honest logging and someone with authority to own the numbers. An outside firm helps most when nobody inside has the time, when the workflow crosses several departments or systems, or when you want the people who measure the baseline to also build the system and stay accountable for the result.