How to Measure AI Agent ROI

Start with a definition, not a spreadsheet

Return on investment (ROI) for an AI agent is the value the agent creates, minus everything it costs, divided by what it costs. The formula takes ten seconds; the honest part is deciding what counts as "value" and what counts as "cost." Most vendor ROI claims inflate the first and shrink the second. This guide shows you how to measure both for yourself.

An AI agent, for ROI purposes, is best thought of as a system with running costs: software that plans and executes multi-step work with tools. Unlike a static app, its costs move with usage — and its value moves with how well it's tuned. If you're new to agents, our SI agent explainer covers the basics.

The ROI formula

The arithmetic is simple:

ROI = (Annual value created − Annual total cost) ÷ Annual total cost

A result of 1.0 means the agent returned its cost plus an equal amount — a 100% return. A result of 0 means it broke even. Below zero, it's losing money. The interesting question is never the formula; it's what's inside each term. Track both monthly for the first quarter, because the first month's numbers are almost never the steady-state numbers.

Costs to include (all of them)

• Platform subscription. The plan price — the only cost most calculators include.
• Usage and API costs. Credits, executions, tasks, tokens. Pull real figures from the usage dashboard, not the pricing page.
• Build and setup labor. Every hour spent designing, connecting, and testing the workflow, valued at the person's real hourly cost. This is usually the largest line item.
• Ongoing monitoring. Reviewing logs, approving gated actions, retuning instructions. Budget it weekly for the first two months.
• Error remediation. The cost of fixing what the agent got wrong — refunds, corrections, customer appeasement. Track incidents as they happen.
• Training. Time spent teaching the team to work with and around the agent.

Our pricing models guide helps you forecast the usage side accurately.

Benefits to measure

• Labor hours saved. The direct one: hours the task took before, minus hours of supervision now, valued at real labor cost.
• Speed gains. Faster response times — speed-to-lead, support first response — often matter more than hours saved. Measure the before-and-after in hours, not feelings.
• Error reduction. If the agent is more consistent than the manual process, count the avoided rework.
• Capacity unlocked. Work that simply didn't get done before — follow-ups that were skipped, data that was never cleaned — now happens. Assign it a conservative value.
• Employee satisfaction. Real but hard to quantify; note it qualitatively rather than inventing a dollar figure for it.

A worked example (illustrative)

This is a hypothetical illustration of the method, not a claim about any product or industry — plug in your own numbers:

• A 12-person company builds a support-triage agent. Build labor: 25 hours at a loaded cost that totals $2,500.
• Year 1 costs: $1,200 subscription + $900 usage + $2,500 build + $1,400 monitoring and fixes = $6,000.
• Year 1 value: 8 hours/week of triage saved × 50 weeks = 400 hours; valued at $35/hour = $14,000, plus faster first responses.
• ROI = ($14,000 − $6,000) ÷ $6,000 = 1.33, or a 133% return.

Notice what happens if you count only the subscription: the same agent looks like it returned 10x — a number that impresses in a slide deck and misleads in a budget meeting. The honest version includes the labor. That's the version that predicts whether the next agent is worth building, too. For small teams, our small business guide suggests starting with one such measured workflow.

Metrics worth tracking monthly

• Cost per run — total monthly cost ÷ runs. If this drifts up, investigate.
• Success rate — runs completed without human rescue. The driver of both value and cost.
• Approval rate — share of gated actions approved vs. rejected. Falling approvals mean the agent is drifting.
• Time to value — hours from build start to the first unsupervised run.