How to Measure AI ROI at the Employee Level: Six Outcomes Beyond License Counts

Last updated November 6, 2026

In a survey of 2,000 CEOs across 33 countries, run by the IBM Institute for Business Value's 2025 CEO study with Oxford Economics, chief executives reported that about 1 in 4 of their AI initiatives from the previous three years had delivered the return on investment they expected.

AI ROI at the employee level is the verified change in how work gets done after people start using AI, measured against a baseline or a comparable group still working without it. At The Decision Lab, we measure it the way we measure any behavior change, across six outcomes: workflow performance, decision quality, errors and rework, employee experience, customer outcomes and business impact. License counts and logins show who has access, and neither shows whether the work improved. We used this approach in a 20,000-person Fortune 500 AI adoption program, where we piloted eight programs against a control group.

For each workflow, define one unit of output, one speed measure, one quality counter-metric, one downstream outcome and a baseline or comparison group before rollout. We ask clients to keep the full template on a single page that finance and operations both sign off on, so the method is agreed before any results come in.

The template's fields look different in each function. The examples below are starting points, to be replaced with whatever each team already tracks.

Most AI dashboards report what is easy to collect, such as seats assigned and weekly active users, and those figures describe access and activity. A heavy user may be producing more finished work or may be spending much of the day correcting drafts, and the dashboard looks the same either way.

Self-reported time savings usually fill the gap, and controlled research shows how far they can drift from measured time. In a randomized trial of AI coding tools run by the research nonprofit METR in early 2025, 16 experienced open-source developers completed 246 real tasks on codebases they knew well, with AI allowed on some tasks and blocked on others. Tasks where AI was allowed took 19% longer to finish than tasks without it. After the study, the same developers estimated that AI had cut their completion time by about 20%.

METR describes the result as a snapshot of one setting, experienced developers working on large and mature projects, and it does not show that AI slows everyone down. It does show that a survey asking these developers how much time AI saved them would have recorded a gain where the task timings recorded a loss.

Even where time savings are real, some of them get spent again. In Workday's 2026 study of AI rework, fielded by Hanover Research in November 2025, 3,200 full-time employees who actively use AI at organizations with more than $100 million in annual revenue were surveyed. More than 4 in 5 said AI saves them between one and seven hours a week, and Workday estimates that close to 4 in 10 of those saved hours are lost to rework, mainly correcting errors and rewriting generic output. Both figures rest on employees' own estimates, so we read them as a reason to measure time saved and value created separately.

The studies cited in this guide used different workers, tasks, tools and research designs, so each percentage describes its own setting and should not be compared with the others or used as a target. In our own projects, the most useful finding is often the gap between what employees report and what their work shows, because it points to the tasks where people believe AI helps and the results say otherwise.

Workflow performance is the most direct measure of AI value at the employee level: how much finished work a defined process produces per person, and how long each unit takes. Start by naming the unit, such as a resolved support ticket, a drafted contract clause, a reconciled account or a qualified sales lead. Then record how long each unit takes and how many each person completes before AI is introduced.

A study of generative AI in customer support by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in The Quarterly Journal of Economics in 2025, followed more than 5,000 support agents as an AI assistant was rolled out in stages. Agents with the tool resolved about 14% more customer issues per hour, on average, than agents without it. Among the newest and least-skilled agents the gain was about a third (34%), while the most experienced and skilled agents saw little change.

For measurement, this means reporting workflow gains by role and by experience level. An organization-wide average can blend a team with large gains and a team with none into a modest number that describes neither.

In our Fortune 500 pilot, we measured how easy employees found AI to use and how long it took them to use AI on a task, before and after each program. Measuring at the level of a single program let us trace changes in behavior to a specific intervention instead of to the rollout as a whole.

Many of the most valuable uses of AI support a judgment, such as which claim to investigate or which candidate to interview, and speed says little about whether the judgment was right, so decision quality needs a measure of its own.

Fabrizio Dell'Acqua and colleagues ran a field experiment on AI and knowledge-worker quality with 758 Boston Consulting Group consultants, published in Organization Science in 2026. On 18 realistic tasks that fell within what the authors call AI's capability frontier, consultants with access to GPT-4 completed about 12% more tasks and finished them about a quarter faster than consultants without it. On a business problem chosen to fall outside that frontier, consultants using AI were 19% less likely to reach the correct answer than those working without AI.

The same tool raised output on one kind of task and lowered accuracy on another, and a dashboard tracking speed alone would have counted both as wins. Practical ways to measure decision quality include:

• auditing a random sample of AI-assisted decisions against a known correct answer
• tracking how often AI-assisted decisions are later reversed
• having a second reviewer score a sample without knowing which cases used AI
• comparing downstream results, such as loan defaults or new-hire retention, for decisions made with and without AI

Decision quality sits close to the center of our work as a behavioral science firm, since most of what we study is how people make choices under uncertainty. We usually build a measured starting point before testing any change, as in the decision-quality diagnostic we ran with 500 shoppers for Pinterest, where the brief was to measure shopping decisions directly instead of relying on what shoppers said about themselves.

Rework is where reported time savings most often disappear, so measure it at the point where AI-assisted output gets checked. Track the share of AI-assisted drafts that need correction before use and the time spent reviewing and fixing them, then follow the errors that get through, especially those that reach a customer or another team.

Rework often lands on someone other than the person who used AI. A summary that leaves out a key figure costs its reader time, so measure rework on the receiving end of a workflow as well as at the desk where the output was produced.

We prefer first-pass yield, the share of work completed correctly the first time, as the headline quality measure for AI-assisted workflows, because it captures speed and accuracy in a single number. Our standing rule for AI-assisted workflows is never to report a speed improvement without its paired quality measure, so tickets that close faster and reopen more often count as no gain at all.

Employee experience shows whether people can keep working this way once the launch period is over. Useful indicators include confidence using AI, perceived effort, after-hours work and voluntary turnover, each compared between teams using AI and teams that are not.

In a six-month randomized field experiment on generative AI at 66 large firms, run with Microsoft by Eleanor Dillon, Sonia Jaffe, Nicole Immorlica and Christopher Stanton, workers who used Microsoft 365 Copilot reduced their time working outside regular hours compared with colleagues who did not get access. The customer support study found that AI assistance also increased employee retention.

In our Fortune 500 pilot, confidence in exploring AI tools was measured before and after the one-month pilot. It rose 41% among participants and fell 26% in the control group over the same period. These are results from a single pilot with more than 100 employees at one company, and they should not be read as a benchmark for other organizations. The published case study does not report statistical significance.

We read confidence next to data on actual use, since a rise in confidence with no change in use can point to a barrier elsewhere, such as unclear rules about what AI may be used for.

For any workflow that touches customers, AI value should show up in what customers experience. Measure satisfaction scores, time to resolution, escalations, complaints and repeat contacts on the same issue, and compare them with the pre-AI baseline or a comparison team.

In the customer support study, the authors report that AI assistance improved customer sentiment toward agents.

Customer measures also catch problems that internal metrics miss. A drop in average handling time that arrives with a rise in repeat contacts usually means issues are being closed before they are solved, and only the customer data shows it.

Business impact converts the first five outcomes into money, and it is the figure most executives will be asked to defend. The biggest measurement risk is counting freed-up time as savings.

The Microsoft field experiment shows the gap between the two. Across 7,137 knowledge workers, employees who used Copilot spent about two fewer hours a week on email than the control group in the second half of the six-month study. The researchers did not detect a matching change in the amount or mix of tasks those workers took on, so the study gives no sign that the recovered hours went into new work.

For a finance audience, we recommend reporting two figures side by side. Realized value is the change that shows up in a budget or revenue line, measured against a baseline or comparison group. Capacity value is the time freed but not yet put to a specific use, and it should never be added into the first figure.

When a global vehicle manufacturer asked us to evaluate a new sales model and measure its ROI, we used data science methods for the evaluation and then built a dashboard so the sales team could keep tracking the indicators that mattered after the project ended. AI investments benefit from the same setup, because a value figure measured once at launch is hard to defend a year later.

Once the six outcomes are measured, the calculation itself is simple:

AI ROI = (verified value gained − full cost of AI) ÷ full cost of AI

Verified value gained is the realized value from the business impact section: cost reductions and additional revenue measured against a baseline or comparison group, after subtracting the time lost to rework. The full cost of AI covers licenses and usage fees, integration work, training time and the time employees spend reviewing AI output. Review time is the easiest of these to leave out, and Workday's rework estimates suggest it can be large.

A worked example

The figures below are hypothetical and chosen to show the method, not drawn from a client or a study. Take a support team of 40 agents, each working about 1,600 hours a year at a loaded cost of $60 an hour. After AI is introduced, the team resolves 10% more tickets per agent-hour than a comparison team that starts later, and its reopen rate stays level with the comparison team's, so the quality guardrail holds. Because the team now clears demand faster, it uses 1,500 fewer overtime hours over the year.

A 10% gain in tickets per hour frees about 9% of the team's 64,000 hours, or roughly 5,800 hours. Only the 1,500 that replaced overtime count as realized value. If the remaining hours were counted as savings too, this near break-even result would be reported as a return of about 300%, and finance would be right to reject it. The honest report is a 3% realized return plus about 4,300 hours of capacity that someone now has to decide how to use.

Two rules keep the figure credible. If saved time already shows up as more output, count the output once and leave the hours out. And divide the result by the employees in a role or team, not by named individuals, because individual AI scores invite people to inflate their usage and make them less willing to report where AI is getting things wrong.

1. Choose a small number of workflows where AI is expected to change the work, and agree before rollout on what success looks like for each. Name a manager who owns the results of each workflow.
2. Fill in the measurement template and record the baseline before anyone gets access, since a baseline cannot be rebuilt reliably once AI is in use.
3. Set up a comparison. A randomized pilot gives the clearest answer, and a staggered rollout, where some teams start later than others, is often more practical. Controlled experiments are the strongest evidence, but a careful before-and-after comparison can be enough for a small, low-cost tool.
4. Agree with finance how each outcome converts to money before results come in, so the method cannot be fitted to the numbers afterward.
5. Log review and correction time from the first week, because it is hard to reconstruct later and can decide whether the return is positive.
6. Review results at fixed points, such as 90 days and six months after rollout, and report them by team and role. A single early reading can mislead, since how people use a new tool often changes once the novelty wears off.

If outcomes stay flat, we diagnose the conditions behind AI use before anyone changes the tool. Our SPROUT framework measures five of them, and each points to a different fix:

• Social readiness: whether team norms make AI use acceptable. A typical signal is employees who use AI privately but avoid it in shared work.
• Psychological readiness: how open people feel toward AI, including any fear of being replaced. Low use concentrated in the roles that see themselves as most exposed is a common sign.
• Role fit: whether AI is relevant to the tasks people actually do. Early use that fades after a few weeks often means AI was pointed at the wrong workflow.
• Organizational enablement: whether access, policy, incentives and manager support back AI use. Employees who are unsure what they are allowed to use AI for point here.
• Understanding and technical capability: whether people have the skills to use AI well and to check its output. Steady use combined with high rework is the usual signal.

Our guide to AI adoption readiness assessment covers how each dimension is measured.

How do you measure AI ROI beyond license counts?

Use metrics tied to the work itself, such as output per person, decision accuracy, first-pass yield and customer satisfaction, compared against a pre-AI baseline. Usage still matters as a diagnostic. The most useful usage measure is depth: the share of eligible tasks that employees actually complete with AI, which says more than logins or prompt counts.

How do you separate capacity gains from realized savings?

Count a saving as realized only when it appears in a budget line, such as lower overtime or a planned hire that was no longer needed. Log every other freed hour as capacity, and ask managers each quarter where that time went. Capacity becomes realized value only when someone decides what the hours are for and the result shows up in cost or revenue.

How long does it take to measure the ROI of an AI rollout?

Plan for at least one full quarter of data after rollout, and longer for workflows with slow feedback, such as sales cycles or hiring, where outcomes take months to appear. The Microsoft field experiment ran for six months and measured its email effects in the second half, after workers had time to fold the tool into their routines.

Can you trust employees' self-reported AI time savings?

Treat them as a starting point. Self-reports are useful for finding the tasks where employees think AI helps most, and those are the tasks to time first. Then check the estimates against measured task times on a sample of real work, since the developers in METR's trial believed AI had sped them up when it had slowed them down.

Should AI ROI be measured for individual employees or for teams?

Collect data at the individual level where the analysis needs it, but report and act on it by team or role. Individual rankings encourage people to inflate their usage and hide errors. They also blur the effect of the tool with differences between the people using it, since early adopters may differ from colleagues in ways unrelated to AI.

Do you need a control group to measure AI ROI?

Not always. Randomized pilots give the clearest estimate of what AI changed, but a staggered rollout or a well-matched comparison team is often enough to guide a budget decision. In our projects, we set the evidence standard by how costly a wrong call would be, so a company-wide license renewal gets a stricter test than a single-team pilot.