How to measure the ROI of AI agents.

01What ROI means for an AI agent

Return on investment for an AI agent is the net value it produces in a period, as a share of what it cost to build and run in that period. The unit that makes it measurable is the task: one invoice matched, one ticket resolved, one contract reviewed. Cost per task against value per task, times the number of tasks, is the whole calculation.

Measuring per task keeps the conversation honest. A claim that an agent "saves 2,000 hours a year" is hard to check. A cost per invoice that fell from six dollars to two, across 4,000 invoices a month, is a number finance can audit.

02Why agent ROI is hard to prove today

In June 2025, Gartner predicted that over 40 percent of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. Two of those three reasons are measurement problems.

Headcount is the shortcut many organisations reach for. In a survey of 350 executives at companies with at least $1 billion in revenue, fielded in the third quarter of 2025 and published in May 2026, Gartner found about 80 percent of organisations piloting or deploying autonomous capabilities reported workforce reductions. The rate of reductions was nearly the same among those reporting higher ROI and those reporting modest or negative results. A smaller team is a budget decision. It does not show what the agent returned.

The cost side moves, too. In August 2026, Gartner predicted that AI inference costs per agentic workflow will increase more than fivefold through 2028, and that routing a task to an agentic reasoning model costs at least five times a basic chatbot interaction. An ROI calculated on last quarter's model prices can be wrong by the next review.

03The cost side: what an agent really costs

• Run cost. Model calls per task times price, including retries, tool calls and the context that grows as a task goes on.
• Platform cost. Hosting, the gateway or control layer, storage, monitoring and licences, allocated to each agent.
• Build and change cost. The hours to design, integrate, test and update the agent, spread over the months you expect it to run.
• Human review. The time people spend checking, correcting or approving its output. It is the cost most often left out, and for many agents it is the largest.
• Rework. The cost of putting right what the agent got wrong, measured against the error rate of the old process.

04The value side: what counts

• Hours moved to other work. Time saved counts when the people it frees take on work you can name. Time that disappears into the week never reaches the accounts.
• Throughput. More tasks completed by the same team: claims processed, tickets closed, invoices matched.
• Cycle time. Work that finishes sooner, where speed has a price: faster quotes, a shorter month-end close, quicker onboarding.
• Quality. Errors avoided, measured as rework, write-offs or penalties, against the baseline error rate.
• Revenue. Conversion or retention that moved, where a control group lets you attribute it.

05How to measure it: six steps

1. Pick one workflow and define the task. Write down what one completed task is and how you know it is done.
2. Measure the baseline before launch. Record tasks per week, time per task, cost per task and error rate for at least four weeks before the agent starts. Without a baseline, every later number is an estimate.
3. Record the cost of every run. Tag each model call with the agent, the team and the workflow at the moment it runs, so cost per task comes from a ledger rather than a split of the monthly invoice.
4. Measure the human time in the loop. Log how long reviewers spend per task. If review takes as long as the original work, the agent has moved the work rather than removed it.
5. Compare against the baseline, with a control where you can. Run the agent on part of the volume and the old process on the rest for a few weeks. The difference is the agent's contribution.
6. Report monthly, per task. Cost per task, value per task, tasks completed, net value and payback. Look first at agents with negative net value: narrow their scope, move them to a cheaper model, or retire them.

06The formula, with a worked example

AGENT ROI, PER MONTH
Agent [name] · Owner [name] · Workflow [name]

Tasks completed            [n]
Value per task             [$]  cost of the task the old way,
                                from the baseline, plus any
                                quality or revenue effect
Gross value                [tasks x value per task]

Run cost                   [$]  model calls, from the cost ledger
Platform cost              [$]  allocated to this agent
Human review               [$]  review hours x loaded rate
Build cost, amortised      [$]  build cost / months in service
Total cost                 [$]

Net value                  [gross value - total cost]
ROI                        [net value / total cost]
Payback, months            [build cost / (gross value - run,
                                platform and review cost)]

An illustration with round numbers, not a customer result. An accounts payable agent matches 4,000 invoices a month. Before it started, a person took six minutes per invoice at a loaded rate of $60 an hour, so each invoice cost $6.00 and the month cost $24,000. The agent's run cost is $0.40 an invoice, or $1,600. Platform cost is $1,000. People review the 15 percent it flags, 600 invoices at six minutes, which is 60 hours or $3,600. It cost $36,000 to build, spread over 24 months at $1,500 a month.

Total cost is $7,700 a month. Gross value is $24,000, so net value is $16,300 and ROI is 212 percent. The build pays back in about two months.

Now test the assumption that matters most. If only half of the freed hours go to other work, gross value falls to $12,000 and net value to $4,300. It is the same agent with about a quarter of the return. Check where the hours went before you report the bigger number.

07Signals to watch before the ROI is in

• Cost per task, trending down as prompts, routing and model choice improve.
• The share of tasks completed without a human correction.
• Review minutes per task.
• Tasks completed per week, against the baseline.
• Spend against the agent's cap, so a cost surprise shows up as a request rather than an invoice.