AI Agent ROI: How to Calculate It, With a Worked Example

roi · ai-agents · strategy

AI Agent ROI: How to Calculate It, With a Worked Example

How to calculate the ROI of an AI agent: the formula, the inputs to measure first, a worked example with review time and model costs, payback period and what to track after launch.

Remolda Team10 min read

AI agent ROI = (annual benefit − annual cost) ÷ annual cost. The benefit is net minutes saved per case after human review, times cases per year, times the loaded hourly cost, plus errors avoided. The cost is the build spread over its life plus model calls, hosting, review time and maintenance. Payback = one-time cost ÷ monthly net benefit.

The free AI ROI calculator gives a first estimate in a few minutes. A business case built on your own measured baseline is part of the two-week AI Opportunity Audit ($2,900 CAD plus applicable taxes, fixed). Send a written request or call +1 (613) 867‑7767.

What goes into the AI agent ROI formula?

Four inputs drive most of the result. Measure them on the current process before anyone builds anything.

InputHow to get itCommon mistake
Cases per yearTicket system, ERP or mailbox counts for the last 12 monthsUsing the forecast instead of history
Minutes per case todayTime 20–30 real cases, including look-ups and reworkTiming only the easy cases
Minutes per case with the agentAgent run time plus the person's review and approvalLeaving out review time
Loaded hourly costSalary plus benefits, payroll taxes and overhead, from financeUsing base salary

Two more inputs matter when they apply: the error rate and cost per error (rework, write-offs, penalties), and the value of speed (earlier billing, faster onboarding). Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, and unclear business value is one of the three reasons it gives. A measured baseline is the cheapest protection.

The four steps of an AI agent business case. Review time belongs in both the benefit and the cost.

Step 1: Annual benefit from time saved

Annual time benefit = (minutes today − minutes with agent and review) ÷ 60 × cases per year × loaded hourly cost

The agent's own run time rarely matters; the person's review time does. If a 25-minute task becomes 2 minutes of agent work plus 6 minutes of review, the saving is 17 minutes, and the 6 minutes stay on the cost side of the team's week.

Step 2: Errors avoided and speed

Error benefit = (error rate today − error rate with agent) × cases per year × cost per error

Take the error rate from rework logs, credit notes or audit findings. Estimate the agent's error rate from a test on your real cases; vendor figures describe someone else's documents. Speed benefits, such as invoicing a few days earlier, count only when they change cash or revenue; otherwise list them as non-financial.

Step 3: Total cost of ownership

CostOne-time or runningWhat it covers
Discovery and designOne-timeMapping the workflow and its exceptions, the baseline, approval rules
Build and testingOne-timeAgent, tools, test set of real cases, logs
IntegrationOne-timeAccess to the mailbox, CRM, ERP or document store; older systems cost more
Security and privacy reviewOne-timePrivacy assessment, data location, permissions
Training and changeOne-timeTeaching the team to review, updating procedures
Model callsRunningTokens per case × cases × price per token
Hosting and monitoringRunningServers, logs, alerts
Human reviewRunningMinutes per case spent approving or correcting
MaintenanceRunningUpdating instructions and tests when documents, rules or models change

Model calls are usually small. At Claude Haiku 4.5 list prices of US$1 per million input tokens and US$5 per million output tokens, a case that reads 20,000 tokens and writes 2,000 costs about US$0.03, or roughly US$120 for 4,000 cases a year. Larger models and long multi-step runs cost more; measure tokens per case in the pilot. Review time and maintenance usually outweigh the model bill.

Worked example: contract intake agent

Illustrative numbers. Replace each with your own baseline.

InputValue
Contracts per year4,000
Minutes per contract today25
Minutes with agent (2 agent + 6 review)8
Net minutes saved17
Loaded hourly cost$75
Error rate today → with agent4% → 1.5%
Cost per error$800

Benefits per year

  • Time: 17 ÷ 60 × 4,000 × $75 = $85,000
  • Errors: (4% − 1.5%) × 4,000 × $800 = $80,000
  • Total: $165,000

Costs

  • One-time build, integration, review and training: $60,000 (assumed), spread over 3 years = $20,000 a year
  • Running: models, hosting, monitoring and maintenance: $18,000 a year (assumed)
  • Annual cost: $38,000

Result

  • ROI = (165,000 − 38,000) ÷ 38,000 ≈ 334% a year
  • Monthly net benefit = (165,000 − 18,000) ÷ 12 ≈ $12,250; payback = 60,000 ÷ 12,250 ≈ 5 months

Now the conservative case.

Step 4: Conservative and expected scenarios

Present the conservative case; decisions are held to it.

ExpectedConservative
Time benefit100%70%
Error benefit100%50%
One-time cost100%130%
Running cost100%120%
Launch dateAs planned30% later

In the example, the conservative case gives $99,500 of benefit against $47,600 of annual cost: ROI ≈ 109%, payback ≈ 12 months. If a business case only works in the expected column, the risk is too high for a full build; run a pilot first.

Which input moves the result most? Usually review minutes. In the example, review rising from 6 to 12 minutes cuts the time benefit from $85,000 to $55,000. Test the agent on real cases before trusting any review estimate.

What to track after launch

Agree the metrics before the pilot and report them monthly.

  • Volume: cases handled by the agent, cases sent to a person.
  • Review: minutes per approval; share of drafts the reviewer changed.
  • Quality: errors that reached a client or another system; exceptions by type.
  • Cost: tokens and model cost per case, hosting, maintenance hours.
  • Outcome: minutes per case and error rate against the baseline.

With these in place you report actual ROI at 3, 6 and 12 months and widen the agent's scope on evidence.

When an AI agent is the wrong investment

Low volume, steps that never vary, or work where every output needs a full expert rewrite rarely pay back with an agent. Fixed, structured steps are cheaper with plain rules or RPA; our comparison of AI agents and traditional automation shows where each fits. Broader AI projects, such as analytics or training, follow the general AI ROI calculation guide.

How Remolda builds the business case

  • AI Opportunity Audit ($2,900, two weeks): we measure the baseline on your workflows, rank candidates, and deliver the ROI model with expected and conservative cases. The AI business case service covers larger approvals.
  • AI Pilot Sprint ($9,800, six weeks): one AI agent on one workflow, measured against the baseline, so the full investment rests on numbers from your own cases.

Want the numbers for your workflow? Send a written request with the workflow, monthly volume and who does it today. We reply within one business day with the inputs we would need.

Sources

  • Gartner press release, June 25, 2025: over 40% of agentic AI projects will be canceled by end of 2027
  • Anthropic: Claude API pricing (Claude Haiku 4.5)
Share

Related insights

Frequently asked questions

Talk to an AI transformation consultant

A 30-minute consultation (€100): you describe the situation, we tell you what to do first and what it would cost.

30 minutes, €100. English or French.