How to Calculate the ROI of an AI Agent Before You Build One

Estimate AI agent ROI per process, not per technology. Start from a measured baseline (volume, handling time, loaded cost, error and delay costs), estimate the share of cases the agent can complete to the required standard, and subtract its full cost: build and integration, model usage at real volume, human review, monitoring and maintenance. Run a pessimistic, expected and optimistic scenario, price in the cost of mistakes, and set kill criteria before you build. If the case only works in the optimistic scenario, choose simpler automation or a smaller scope.

Most agent ROI estimates are written backwards: start from a vendor's productivity claim, multiply by headcount, and call it value. They miss three things. First, agents rarely complete every case; the remainder still needs people, and often more context-switching than before. Second, someone has to review, correct and monitor the agent, permanently. Third, the agent depends on systems (CRM, ERP, ticketing, documents) whose integration and maintenance costs dominate the budget.

The market data reflects this. Gartner predicted in June 2025 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. McKinsey's State of AI 2025 found 62 percent of organizations at least experimenting with agents, but only 23 percent scaling one anywhere, usually in one or two functions. A disciplined estimate before building is the cheapest way to avoid becoming part of the first statistic.

You cannot estimate savings without knowing current cost. Measure, do not guess:

The key number is the completion rate at acceptable quality: the share of cases the agent finishes correctly without human rework. It is not the share it attempts. Estimate it from a small offline test on real historical cases before any build: give the model the same information a person would have and score the output against what actually happened.

Separate three outcomes: completed correctly, escalated to a person (cost: review time), and completed incorrectly (cost: the mistake plus its correction). The third bucket is what makes or breaks the case, because a wrong refund, a misfiled document or an incorrect customer answer can cost more than the agent saves on many correct cases. Our guide to AI agent evaluation explains how to build that test set.

Agent costs fall into one-off and recurring categories. Recurring costs are the ones most often missing from business cases.

Combine the inputs into a simple monthly model and run it three times with pessimistic, expected and optimistic completion rates and costs.

Labour savings are the easiest benefit to count, but often not the largest. Faster response times can raise conversion; shorter cycle times can release cash; consistent checks can reduce compliance risk; 24-hour coverage can capture demand that is lost today. Count these only where you can measure them before and after, and attach them to an owner who will report on them. Unmeasurable "strategic value" is the usual way weak cases survive approval.

Two risks deserve explicit treatment. Error impact: list the worst plausible mistake the agent could make, how often controls would let it through, and what it would cost. If a single error could be severe (money movement, legal commitments, safety), budget for approval steps, which reduce the savings, or keep the agent advisory. Delivery risk: integrations with legacy systems, unclear process ownership and poor data each add uncertainty; widen the pessimistic scenario accordingly.

An illustrative scenario, not a client case: a distributor processes 4,000 supplier invoices a month. Baseline: 6 minutes per routine invoice, 18 percent exceptions. An offline test on 300 historical invoices shows the agent extracts and matches 85 percent of routine invoices correctly, escalates 12 percent and gets 3 percent wrong, all of which are caught by a three-way match rule before posting. The model shows labour savings on roughly 2,800 invoices a month, against model and infrastructure costs, a reviewer for escalations and maintenance. The expected scenario pays back the build within the first year; the pessimistic scenario does not, so the team pilots on one supplier group first and keeps posting behind the rule-based match.

ROI depends heavily on choosing the right process and the right approach. Many processes are better served by deterministic workflow automation, which is cheaper to run and easier to trust; see which processes suit AI agents. If you do need an agent, compare platform and custom options in build vs buy AI agents, and plan for the problems covered in why AI agents fail in production. For model cost control, see LLM cost optimization.

A credible AI agent ROI estimate is a process baseline, a tested completion rate, a full cost model, three scenarios and a stop rule. It takes a few days to build and saves months of building the wrong thing. If the numbers only work when everything goes right, the answer is usually a smaller scope or simpler automation, not a bigger model.