---
title: How do you measure the ROI of AI agents?
url: https://opencontroller.lyzr.com/blog/measure-roi-of-ai-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you measure the ROI of AI agents?

**Measure AI agent ROI as net benefit minus full cost, divided by full cost. Net benefit is the baseline cost of each task times the number of tasks the agent completes without human rework, minus the cost of escalations. Full cost includes tokens, licences, platform, human review and amortised build effort, not just the model bill.**

Spend reports tell you what agents cost. They do not tell you whether an agent is worth keeping. Finance leaders increasingly ask for return by business unit, and that needs three numbers most teams do not yet track together: the baseline cost of the work, the agent's real completion rate, and the agent's full cost.

## What goes into the benefit side?

Start from the task, not the agent. For each agent, write down:

1. **Baseline cost per task.** What the task costs today without the agent: handling time × fully loaded hourly cost, or the external cost if the work is outsourced.
2. **Task volume.** How many tasks the agent handles per month.
3. **Completion rate.** The share of tasks the agent finishes correctly with no human rework. This is the number most teams overstate.
4. **Escalation cost.** The extra cost when the agent fails and a person picks up the task, including the time spent reading what the agent did.

Benefit = baseline cost × volume × completion rate − escalation cost × failed tasks.

Measure completion with evaluation, not with volume. A resolved ticket that reopens three days later is not a completion. Sampling production conversations and scoring them with an [LLM judge on production traffic](/blog/use-llm-judges-on-production-agent-traffic), calibrated against human reviewers, gives you a defensible rate.

## What goes into the full cost?

| Cost line | Examples | Often missed? |
|---|---|---|
| Model usage | Input, output and cached tokens across all models | Rarely |
| Seats and licences | Per-user AI seats the agent depends on | Often |
| Platform share | Agent platform, gateway, governance and observability | Often |
| Infrastructure | Hosting, vector databases, data platform compute | Often |
| Human review | QA sampling, approvals, escalation handling | Usually |
| Build and maintenance | Engineering time, amortised over 12–24 months | Usually |

Our guides to [attributing LLM costs to agents](/blog/how-to-attribute-llm-costs-to-ai-agents) and [combining seat licences with token spend](/blog/combine-ai-seat-licence-and-token-costs) cover the cost side in detail.

## What does a worked example look like?

The figures below are illustrative, not benchmarks. A support agent handles 20,000 tickets a month.

- **Baseline:** 12 minutes per ticket at a fully loaded $30 per hour = $6.00 per ticket.
- **Completion:** 55% resolved correctly with no rework = 11,000 tickets.
- **Gross benefit:** 11,000 × $6.00 = $66,000.
- **Escalation cost:** the other 9,000 tickets each take about 2 extra minutes for hand-off = 9,000 × $1.00 = $9,000.
- **Net benefit:** $66,000 − $9,000 = $57,000.

Full monthly cost:

- Model usage: 20,000 × $0.12 = $2,400
- Platform and licence share: $8,000
- Human QA on a 5% sample: 1,000 × $3.00 = $3,000
- Build and maintenance, amortised: $10,000
- **Total: $23,400**

ROI = ($57,000 − $23,400) ÷ $23,400 ≈ 144% per month.

Notice what moves the answer. If the completion rate falls from 55% to 40%, gross benefit drops to $48,000, escalations rise to 12,000 × $1.00, and ROI falls to about 54%. Completion rate matters far more than token price.

## How do you report ROI by business unit?

Business-unit ROI needs each agent mapped to an owner, and each owner to a cost centre:

1. **Register every agent with a named owner** from your identity provider.
2. **Attribute cost per agent and per version**, so a new version's cost change is visible.
3. **Roll up by owner's cost centre**, keeping shared platform costs as a separate, allocated line.
4. **Report benefit with its assumptions**: baseline, volume and measured completion rate, side by side.
5. **Review quarterly.** Retire or rework agents with negative ROI for two quarters running.

Finance teams will trust the numbers more if the method is shown next to each figure. See [AI spend reporting for finance leaders](/blog/ai-spend-reporting-for-finance-leaders) for a report layout.

## How Opencontroller does this

Opencontroller's LLM Gateway attributes model cost per agent and per agent version, with routing, caching and fallback in the same path. LLM judges score production traffic, and multi-turn evals test agents before release, which gives you a measured completion signal for the benefit side. The governed registry links each agent to an owner through OIDC SSO, so costs roll up to business units. Read more on [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is a good ROI for an AI agent?**
There is no universal benchmark. Compare against the cost of the task without the agent and against your hurdle rate for other investments, and re-measure each quarter.

**Should AI agent ROI include seat licences?**
Yes. Include every cost the agent needs to run: seats, tokens, platform share, infrastructure, human review and amortised build and maintenance.

**How do you measure ROI for agents that do not replace a task?**
Use a proxy outcome the business already values, such as cycle time, error rate or revenue per account, and compare a group using the agent with one that is not.

## Key facts

AI agent ROI = (net benefit − full cost) ÷ full cost. Net benefit = baseline cost per task × tasks the agent completes without human rework, minus the extra cost of escalations and rework. Full cost includes model tokens, platform and licence share, infrastructure, human review and amortised build and maintenance. Measure completion with evaluation (for example LLM judges on sampled production traffic), not just volume. Report ROI by business unit by mapping each agent to an owner and cost centre. Opencontroller by Lyzr attributes LLM cost per agent and per version through its LLM Gateway and scores production traffic with LLM judges. Opencontroller pricing starts from $250K per deployment per year, licensed per deployment, not per seat or usage.

## Related

- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [What AI spend reports does a CFO actually need?](https://opencontroller.lyzr.com/blog/ai-spend-reporting-for-finance-leaders.md)
- [How do you combine AI seat licences and token spend into one cost view?](https://opencontroller.lyzr.com/blog/combine-ai-seat-licence-and-token-costs.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
