- Home
- AI agents
- Enterprise AI agent examples
- Measuring AI agent ROI
Plan · Enterprise AI agent examples
How do you measure the ROI of an AI agent?
AI agent ROI compares the value an agent creates on one unit of work, such as a ticket, an invoice or a quote, with its full cost per successful outcome, against a baseline measured before launch. Value is time, rework and loss avoided or revenue gained; cost includes running, building, reviewing and the errors the agent causes.
In this guide: 83 pages
Key takeaways
- Measure the baseline before the agent goes live. Without one, the only numbers you will have are the vendor’s.
- Count cost per successful outcome. The FinOps Foundation warns that "the cheapest model per token is not always the cheapest per outcome".
- Most published ROI figures are estimates or program totals. Amazon’s USD 260 million a year from Java upgrades is an estimate (AWS DevOps Blog, 2024), and Salesforce attributes Wiley’s 213% return to its wider Service Cloud implementation.
- Hours saved turn into money only when the capacity is used, a hire is avoided or a cost stops. Report everything else as capacity.
How do you measure the return on an AI agent?
Pick one unit of work, measure it before the agent, measure it again with the agent on the same definitions, and set the difference against the agent’s full cost per successful outcome. The arithmetic is short; most of the work is in the baseline and the definitions.
- Choose the unit of work: one ticket, one invoice, one quote, one contract review. Value and cost are both counted per unit.
- Measure the baseline for at least a month before launch: volume, handling time, waiting time, error and rework rate, and the hours of the people involved at their full loaded cost.
- Write down what success means for one unit, for example resolved without reopening within seven days, posted without correction, or quote accepted. Only successful units count toward value.
- Measure the same things after launch on the same definitions, against a control group or a comparable period without the agent where you can.
- Divide the agent’s full cost by its successful units and compare that with the baseline cost per unit.
Worked example with illustrative numbers: an accounts payable team handling 10,000 invoices a month. None of these figures comes from a real company.
| Line | Before the agent | With the agent | How to read it |
|---|---|---|---|
| Invoices a month | 10,000 | 10,000 | Same volume, so the comparison is fair |
| People time per invoice | 6 minutes | 1.5 minutes on average | The agent posts 70% untouched; people spend 5 minutes on each of the other 30% |
| People hours a month | 1,000 | 250 | 750 hours of capacity released |
| Invoices needing rework | 2% | 3% in the first quarter, 1.5% after | Failure cost rises before it falls; count it |
| Successful invoices | 9,800 | 9,700 in the first quarter | Value is counted on these only |
| Agent run cost | None | Tokens, tools and runtime per run, from monitoring | Multiply by all runs, including failed ones |
| Build, evaluation and review | None | Spread over the expected life, with one model change a year | Retesting after model changes is a recurring cost |
In this example the agent releases 750 hours a month and adds rework in its first quarter. Whether that is a return depends on what happens to the hours, which is the subject of the last section, and on the run and build cost, which you price at your own rates.
What costs belong in an AI agent business case?
Every cost that exists because the agent exists, over its expected life: running it, building and changing it, the people who review and own it, and the errors it causes. A business case that counts only tokens understates the cost.
- Run cost per unit. Model tokens, including cached and reasoning tokens, plus tool and gateway calls and runtime, multiplied by the steps each request takes. The agent cost monitoring page shows how to attribute this per agent.
- Platform metering. Some platforms bill in their own units. Copilot Studio meters in Copilot Credits: a generative answer is 2 credits, an agent action 5 and agent flow actions 13 per 100, and employee-facing use by Microsoft 365 Copilot-licensed users carries no charge within fair-use limits (Microsoft Learn, 30 October 2026).
- Build and change. The first build, the evaluation set and the retest each time the model, the prompt or a connected system changes. Models retire on the vendor’s schedule, so budget for at least one model change a year.
- People. The reviewers who approve the agent’s output and the owner who answers for it, plus the exceptions that still go to the team.
- Failure cost. Rework, wrong postings, reopened tickets and the occasional incident. Count them from production monitoring; a pilot rarely shows them.
The FinOps Foundation recommends tracking "cost per inference, per user, and per successful outcome for each use case" as distributions, and showing costs back to teams before charging them. Averages hide the long runs that make up much of an agent’s bill.
How should you read the ROI figures companies publish?
Read them as claims to test. Most published figures are calculated, estimated or attributed to a wider program, and few give a baseline or a method. The table shows how seven well-known figures were built.
Published AI return figures and how each was built, as the sources showed on 2 November 2026. None of these companies is a QueryNow client.
| Figure | Source and date | How it was built | How to use it |
|---|---|---|---|
| Amazon: 4,500 years of development work and USD 260 million a year saved on Java upgrades (AWS DevOps Blog, 1 August 2024) | AWS DevOps Blog, 1 August 2024 | Estimates: time from the number of dependencies migrated, cost from hosts removed | Evidence that large migrations pay; no cost of the agent itself |
| Wiley: 213% ROI and USD 230,000 a year saved (Salesforce customer story, read 2 November 2026) | Salesforce customer story, undated | Attributed to Wiley’s wider Service Cloud implementation; the Agentforce pilot reports over 40% better case resolution than the old chatbot | A platform return; the agent’s share is not separated |
| Klarna: work equal to 700 full-time agents in the first month; resolution in under 2 minutes against 11 | Klarna press release, 27 February 2024 | Company figures from month one; the profit figure in the same release is an estimate | Older source; a volume signal, kept apart from the estimate |
| Highmark Health: a "calculated" USD 27.9 million of AI-enabled value in 2025 (Google Cloud use-case list, 2026) | Google Cloud use-case list, updated 22 April 2026 | Highmark’s own calculation across 74 use cases of its assistant; no method published | A program value; nothing per agent |
| Acentra Health: 11,000 nursing hours and nearly USD 800,000 saved (Microsoft customer story, 2024) | Microsoft customer story, 3 December 2024 | Time per appeal letter fell from about six minutes to about three; the totals give no method | Use the per-letter time; treat the totals as a claim |
| Markerstudy Group: 56,000 hours a year | Microsoft UK Stories, 26 November 2024 | About four minutes saved per call multiplied by 840,000 calls | Capacity from a per-call saving; check the per-call measure |
| Duolingo: "5x ROI" and more than 500 hours saved a month | Glean customer story, undated | No method on the page | A vendor multiple without its inputs |
Microsoft’s own IT shows the difference between a goal and a result. Its May 2026 article states a 40% ticket reduction as an ambition; its October 2026 article reports a 30% reduction in tickets created. Only the second belongs in a business case.
When do hours saved turn into money?
Hours saved turn into money when the released capacity handles more volume without new hires, removes overtime or contract spend, or changes revenue. When none of that happens, report the hours as capacity with no cash figure attached.
We split every business case into separate lines. Cash savings are costs that stop: contractor hours, overtime, outsourced processing, duplicate payments caught. Capacity is hours released that the team uses for other work, reported with the work it went to. Revenue and risk cover faster quotes or lead response that change conversion and errors or penalties avoided; these need a comparison group, because many things move revenue at once.
Siemens’ lead agents show the revenue line done well: Salesforce reports leads answered, engagement and qualification rates, so a sales leader can follow the funnel into pipeline. C.H. Robinson’s quotes falling from days to minutes is a speed figure; its revenue effect depends on how many more quotes became shipments, which the release does not say.
How do Microsoft, AWS and Google implement it?
What each cloud provides for the cost side of an agent business case, with statuses as the vendor pages showed on 2 November 2026. The value side comes from your own systems of record.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Attribute run cost to an agent | API Management llm-emit-token-metric sends token metrics to Application Insights with custom dimensions | Application inference profiles with cost allocation tags; per-request metadata in invocation logs | No agent-level cost attribution feature in our research set |
| Cap or fix spend | API Management llm-token-limit quotas per key; Azure budgets alert without stopping consumption | AWS Budgets actions, with budget data updated up to three times a day | Provisioned Throughput, a fixed-cost, fixed-term subscription for steady traffic |
| Lower cost per outcome | Foundry model router; prompt caching on by default for supported models | Bedrock Intelligent Prompt Routing (GA 22 April 2025); prompt caching (GA 7 April 2025) | Implicit context caching with a 90% discount on cached tokens |
None of the clouds computes an agent’s return for you. Success per unit and the baseline come from the ticketing, ERP or CRM data you already hold.
Frequently asked questions
Is there a standard ROI calculator for AI agents?
We know of no standard one, and the FinOps Foundation says there is "no accepted method for multi-agent workloads" for allocating cost. A spreadsheet with your baseline per unit, the agent’s full cost per successful unit and a capacity line does the job better than a generic calculator with someone else’s assumptions.
How long should you measure before judging an agent’s ROI?
At least one full business cycle after launch, and longer where the volume is seasonal or the outcome is slow, such as a sales cycle. Plan for extra rework in the first quarter while the agent and the team adjust, as the worked example on this page assumes.
Should the ROI include the time people spend reviewing the agent?
Yes. Review time is a cost the agent creates, and in drafting agents it can be most of the remaining people time. Acentra Health’s nurses review every letter the agent drafts, and the published figures count letters reviewed per nurse per day, up from 12 to 14 to between 20 and 30.
Can you trust vendor ROI figures?
Treat them as claims. Check whether the figure is measured, estimated or calculated, whether it covers the agent or a wider program, and whether a baseline is given. Then measure your own; the vendor’s process, volume and definitions differ from yours.
More in Enterprise AI agent examples
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- FinOps for AI Overview
- GenAI FinOps: How token pricing really works
- Copilot Studio billing: messages and capacity management
- Amazon Q Developer just reached a 260 million dollar milestone
- Wiley sees 213% return on investment with Salesforce
- Klarna AI assistant handles two-thirds of customer service chats in its first month
- 1,302 real-world gen AI use cases from the world’s leading organizations
- Acentra Health boosts employee productivity with generative AI and Azure OpenAI Service
- Markerstudy Group embraces AI to drive innovation and excellence
- Duolingo customer story
- Transforming IT support across Microsoft with the Employee Self-Service Agent
- Building trust in AI-powered support at Microsoft: our Employee Self-Service agent story
- Siemens takes lead response from days to minutes with Agentforce
- C.H. Robinson debuts Lean AI for ocean shipping, air freight, and customs
- Emit token metrics for LLM APIs (llm-emit-token-metric policy)
- Limit LLM API token usage (llm-token-limit policy)
- Create and manage budgets in Cost Management
- Model router for Microsoft Foundry
- Prompt caching in Azure OpenAI
- Use application inference profiles to track costs
- Include metadata in requests for cost attribution
- Managing your costs with AWS Budgets
- Amazon Bedrock Intelligent Prompt Routing is now generally available
- Amazon Bedrock prompt caching is now generally available
- Context caching overview
- Provisioned Throughput overview
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner