AI ROI: How to Measure Whether AI Is Paying Off

AI pays off when the work it touches gets cheaper, faster, better or larger by more than the AI costs to run. Showing that takes two measurements of the same work, one before rollout and one after. The first has to be taken before the tools arrive, because nobody can reconstruct it accurately later.

Take the baseline before rollout

Pick the one to three workflows AI will change first. For each, record these for two to four weeks before anyone gets the tool:

• Volume. Items handled per week: tickets, invoices, proposals, intake forms.
• Time. Minutes per item, from start to finished.
• Cycle time. How long an item waits from arrival to done.
• Quality. The share of items that need rework or get corrected later.
• Cost. Hours times a loaded hourly rate, plus any outside spend on the work.

Use system records where you can: ticket timestamps, time tracking, CRM stage dates, invoice logs. They hold up better than memory. Self-reported time is useful but weaker evidence, and careful researchers label it as such. The St. Louis Fed's estimates of time saved with generative AI are built on workers' own survey answers, and the authors describe them as self-reported time savings.

Time

Measure time per finished item. A general "hours saved" figure is hard to check and easy to inflate.

A large field study shows why the unit matters. Brynjolfsson, Li and Raymond studied the staggered introduction of a generative AI assistant across 5,179 customer support agents. Access to the tool increased productivity, "as measured by issues resolved per hour, by 14% on average." The average covered a wide spread: a 34% improvement for novice and low-skilled workers, and minimal impact on experienced and highly skilled workers (NBER Working Paper 31161).

Two lessons carry over to any business. Measure output per unit of time for a defined task, and break results out by experience level, because a team average can hide where the gains are.

Time saved becomes money only when it goes somewhere. Write down where it went: more items handled with the same staff, faster turnaround, less overtime, a hire that was not needed, or hours moved to work that earns more. Hours that went nowhere are spare capacity. Count them, but separately from cash.

Cost

Count every running cost, including the ones that are not on the vendor's invoice:

• Seats and usage. Some plans combine the two. Anthropic's Enterprise plan is priced as "Seat price + usage at API rates," and "Admins set user and org spend limits."
• Setup and integration time
• Training hours for staff
• Review time. If someone checks every AI draft, that time belongs in the cost.
• Admin time to manage accounts, settings and renewals

On the savings side, look past staff hours. The MIT NANDA report on enterprise AI found that some of the most dramatic cost savings it documented came from back-office automation, and that real gains came from replacing outsourced processing and external agencies rather than from cutting internal staff. Outside spend that stops is the easiest saving to prove, because it shows on an invoice.

Quality and error rates

A faster process with more errors can cost more than it saves. Track the same quality measures you took in the baseline: rework rate, errors caught in review, corrections customers ask for, compliance exceptions. Keep the review standard the same before and after, or the two numbers cannot be compared.

Revenue effects

Revenue is the hardest effect to pin on AI, because so many other things affect it at the same time. The MIT report lists the same problem among its own limits: ROI measurements "complicated by concurrent operational improvements and external economic factors." Measure the steps closest to the work instead: time to respond to a new lead, proposals sent per week, quote turnaround, win rate on proposals, repeat orders. Where you can, compare a group using AI with a group not yet using it over the same weeks.

Adoption

Value depends on use, and low use pulls the average down. In the St. Louis Fed's 2025 surveys, workers reported generative AI time savings equal to 1.6% of all work hours, a figure that counts people who don't use it at all (November 13, 2025). The number a business sees depends on how many of its people actually use the tool.

Track adoption with the vendor's admin reports. Microsoft's Copilot Chat usage report, for example, lists active users, average daily active users, total prompts submitted and active days per user, over the last 7, 28, 90 or 180 days. Compare active users with paid seats every month.

Usage figures are a leading sign. They show people are trying the tool. The value still has to appear in the time, cost and quality numbers.

What not to count

• Vendor estimates of time saved that come from the vendor's own formula instead of your measurement
• Prompt counts or logins presented as value
• Hours saved that were never put to other work, counted as cash
• Gains from something else that changed at the same time, such as a new hire, a process change or a slow season
• Results from the keenest early users, multiplied across everyone
• Model benchmark scores

A simple worksheet

Fill in one sheet per workflow. Measure the "before" column before rollout and the "after" column the same way after at least a full month of normal use.

Then work out three figures:

• Monthly value is minutes saved per item, times items per month, divided by 60, times the loaded hourly rate, plus outside spend avoided.
• Monthly net is monthly value minus AI seats and usage, minus the cost of added review time.
• Payback period is the one-time setup and training cost divided by the monthly net, which gives a number of months.

How long to measure

Measure after at least one full month of normal work once people have learned the tool, and longer if the work is seasonal. Research windows vary. The MIT NANDA report measured ROI six months after the pilot and noted that six months "may be insufficient to fully assess" success for complex enterprise systems. Its headline failure figure also rests on a narrow definition of success, explained in what the MIT 95 percent figure measured.

There is no trustworthy typical ROI multiple for AI. Published figures mix different methods, populations and definitions of a return. A business's own figure comes from its own baseline and its own measurement.