To measure the ROI of AI automation, time one workflow before you change it, then track the same measures afterward: hours spent, turnaround, backlog and rework. Value the hours saved at your fully loaded labor cost, subtract everything the project costs to build and run, and divide by that cost. Without a baseline, there is no return to measure.
How do you measure the ROI of AI automation?
Use the ordinary ROI formula on a single workflow: the value of what changed, minus what it cost, divided by what it cost. The math is easy. The hard part is having honest numbers for "before" and "after".
- Name the workflow. "Client intake" or "checking loan packets for missing documents", not "AI across the back office".
- Record a baseline. Time the work on real examples for two to four weeks before anything changes.
- Change one thing. Automate the preparation and keep a person on the approval, so the before and after are comparable.
- Measure the same things afterward, on the same kind of work, over a similar period.
- Convert the difference to dollars, subtract the full cost, and divide by that cost.
Expect the first numbers to be modest. In McKinsey's 2024 survey, reported in Stanford's 2025 AI Index, 49% of respondents whose organizations use AI in service operations reported cost savings, but most of them reported savings of less than 10%.1 A narrow, well-measured workflow gives you a number you can defend, which a broad rollout rarely does. For a business that feels behind on AI, it is also the fastest way to show a result the next project can build on.
What should you measure before you automate?
Measure the four things the workflow is supposed to improve, on real work, before anyone changes the process. If you skip this step, every later number is a guess.
| Measure | How to record it | Why it matters |
|---|---|---|
| Hours spent | Staff minutes per item, times items per week | The main source of savings you can count |
| Turnaround | Time from request to finished work, waiting included | What clients and colleagues feel |
| Backlog | Items still waiting at the end of each week | Shows whether the team is keeping up |
| Rework rate | Share of items a person has to fix or redo | Catches savings that come back as errors |
Time the waiting, not just the hands-on work. A task that takes 15 minutes of effort can still take three days to finish because it sits in a queue. That is often where the biggest change shows up. At one law firm we worked with, intake that used to take three days now takes about 20 minutes. More on why intake is a common first pick for law firms.
How do you put a dollar value on hours saved?
Multiply the hours saved by the fully loaded cost of the people who did the work: wages plus benefits. Using salary only understates the savings.
For a reference point, the U.S. Bureau of Labor Statistics reports that private industry employers spent an average of $46.89 per hour worked on employee compensation in June 2026. Wages and salaries were $32.82 of that, and benefits were $14.07, or 30.0%.2 Use your own payroll figures for the roles involved; they will differ by role and region.
An illustrative example, using that BLS average: a workflow that handles 200 items a month and saves 30 minutes of staff time on each frees 100 hours a month, worth about $4.7K at $46.89 an hour. Those hours are only a saving if the team uses them for other work, so say what they went to.
Which costs belong in the calculation?
All of them, for the whole period you are measuring. ROI goes wrong most often on the cost side, because the build is visible and everything after it is not.
| Cost | What to include | Why it gets missed |
|---|---|---|
| Build | Engineering time or vendor fees to design and build the workflow | Rarely missed; it is the first invoice |
| Software and compute | Model usage, hosting, licenses for any new tools | Usage grows as more people rely on it |
| Your team's time | Interviews, testing, reviewing early outputs, training staff | It comes out of people's normal week |
| Data and access | Getting system access, organizing documents, cleaning records | Often takes longer than the build |
| Review and oversight | The person who checks and approves the AI's output | It is the point of the design, so it feels free |
| Running and improving | Monitoring, fixing issues, updating as the process changes | Starts after the project is called done |
Keep the review step in the cost. People should still approve the work that matters, and their time is part of what the workflow costs to run.
How long does AI automation take to pay back?
Often longer than other technology projects, if you measure across all AI spending. Deloitte's 2025 survey of 1,854 senior executives in 14 countries across Europe and the Middle East found that most respondents reported satisfactory ROI on a typical AI use case within two to four years. Only 6% reported payback in under a year.3
Those are averages across every kind of AI project, from chat tools rolled out to everyone to custom systems. A single workflow with a clear baseline is easier to judge, because you can see each month whether the hours, turnaround and rework moved. That is why we start every engagement with a first workflow that has return you can attribute directly to the work.
What about benefits that are hard to put a number on?
List them, but keep them out of the ROI figure. Mixing estimates into the number makes the whole calculation harder to trust.
- Faster response to clients or customers, which can help win or keep work.
- Fewer errors reaching clients, where a single miss can be expensive.
- Work people dislike, gone, which matters for retention.
- Capacity to grow without hiring at the same pace.
If one of these is the main reason for the project, find a way to measure it directly, such as response time to new inquiries, and add it to the baseline worksheet.
When is AI automation not worth it?
When the work is rare, when a simple rule already handles it, or when a mistake is costly and hard to catch. In those cases the build and review costs outweigh the hours saved, at least for now.
- It happens a few times a quarter. The savings won't cover the build.
- The steps never change. Ordinary automation is cheaper and easier to audit.
- Nobody can describe the process. Fix the process first, or you'll automate the confusion.
- The data is out of reach. If you can't get access to the systems, the project stalls before it starts.
For a scoring checklist that ranks candidates by return and effort, see our guide to what to automate with AI first.
How does Soren measure the first workflow with a client?
Soren AI is an AI-native engineering firm that helps established businesses catch up on AI quickly and at lower cost: our engineers work with AI agents to build, run and maintain AI inside the systems a business already uses. In the first two weeks, our engineers agree on a scope and a goal with the client, including how success will be measured.
After launch our engineers stay on to monitor and improve the workflow, so the measuring continues. More on how we work, or see what you get from an AI transformation assessment.
Find a workflow worth measuring
Start with a 30-minute call with our team. We'll look at where your back office loses time and help you pick a first workflow with a return you can measure.
Frequently asked questions
Is AI worth it for a small business?
It can be, if you start with one specific workflow instead of a broad rollout. Smaller teams often feel the change sooner because a few people carry most of the repetitive work. The test is the same at any size: can you name the task, time it today and show what changed after the automation went live?
How do you calculate the ROI of an AI project?
Take the dollar value of what improved, usually hours saved at your fully loaded labor cost, and subtract the total cost of building, running and reviewing the system over the same period. Divide the result by that total cost. Measure one workflow at a time, against a baseline recorded before anything changed.
What is a good ROI for AI automation?
There is no standard figure, and surveys show most companies report modest gains so far. A better test is whether the workflow returns more than it costs within a period your leadership accepts, and whether the trend improves month to month. Set that target with your finance team before the project starts, not after.
Do hours saved count if nobody's job changes?
Only if the time goes somewhere useful. Hours saved are a real gain when the team handles more work, clears a backlog, responds faster or takes on tasks that were being skipped. If the hours simply disappear, record the capacity gained, but don't count it as cash savings. Say plainly which one you are reporting.
Sources
- The 2025 AI Index Report, Chapter 4: Economy, Stanford HAI, April 2025 (citing McKinsey & Company's 2024 survey)
- Employer Costs for Employee Compensation, June 2026, U.S. Bureau of Labor Statistics, September 9, 2026
- AI ROI: The paradox of rising investment and elusive returns, Deloitte, October 22, 2025



