How to calculate AI agent ROI

01

Define what counts as one completed job

Choose a repeatable business outcome, such as a correctly routed enquiry, a support request resolved without reopening, or an intake pack accepted for review. Count that job once, even when it takes several messages, tool calls or attempts.

Record the current volume, case mix, active handling time, waiting time, corrections and exception rate. Separate staff effort from elapsed delay. An enquiry waiting overnight may have a long elapsed time but only a small amount of manual work to remove.

Keep the comparison fair. A pilot containing only straightforward cases should not be compared with the whole team's workload. Identify eligible cases and retain a comparable manual or simpler-automation baseline. Read the AI buyer’s guide if the workflow boundary is not yet clear.

02

Separate capacity, cash and commercial value

Use an adoption or capture adjustment where eligible work does not consistently reach the new workflow, or freed time cannot all be reassigned. Keep qualitative benefits separate when a monetary value lacks evidence. Do not add capacity savings and avoided staffing costs for the same hours.

For lead qualification, track accepted opportunities, not just captured contact details. For customer support, track resolution and recontact rather than rewarding the agent for keeping people away from staff.

03

Include the cost of the whole service

Separate one-time discovery, implementation, data preparation, integration, evaluation and training from recurring costs. Operating costs can include model calls, messaging, search, storage, hosting, licences, monitoring, support and content maintenance. Confirm actual supplier terms for the proposed design; there is no universal agent price.

Human review and exception handling belong in the model. Count them either by reducing time saved or as a separate incremental cost, not both. Include failed runs and retries in usage estimates, since a chargeable attempt may not produce an accepted result.

Track operating cost per accepted task as well as cost per conversation. Use all relevant operating costs in the numerator and genuinely accepted tasks in the denominator. If no tasks are accepted, report that failure rather than presenting a misleading unit cost.

A custom build and a platform-based agent can distribute costs differently. Compare the same workflow and ownership obligations before choosing the delivery approach.

04

A worked example with visible assumptions

Hypothetical arithmetic, not LCR pricing, a quote or a performance forecast: consider 1,000 eligible tasks per month. Manual handling takes 12 minutes per task; the proposed workflow leaves an average of 5 minutes of human handling, including failed cases and correction. Assume the full year operates at this volume.

• Net capacity released: 1,000 × 7 ÷ 60 = approximately 116.7 hours per month.
• Annual capacity: 1,400 hours.
• Illustrative loaded hourly value: R200; usable capture assumption: 60%.
• Modelled capacity value: 1,400 × R200 × 60% = R168,000.
• Illustrative first-year incremental cost: R120,000, including implementation and operation.
• Net modelled value: R48,000; first-year ROI: 40%.

This is a capacity-value case, not proof of R168,000 in cash savings. The same assumptions at 30% capture give R84,000 of benefit and a negative 30% ROI. If launch is partway through the year, reduce the operating benefit period and model the ramp-up instead of using a full steady-state year.

Use the existing workflow ROI calculator to explore your own assumptions. Enter net minutes saved after review and full costs. It is a planning aid, not evidence that an agent will achieve those inputs.

05

Prove the assumptions before scaling

Start with representative cases and explicit acceptance criteria. Include missing information, failed integrations and requests the agent must decline or hand over. Repeat cases where variability matters; one successful run does not establish dependable behaviour.

Anthropic’s evaluation guidance distinguishes evaluating an agent’s transcript from checking the resulting outcome. For this business case, check that the intended record, handover or resolution actually exists, not only that the answer sounds plausible.

1. Capture a baseline before changing the process.
2. Test the same eligible case categories against the proposed workflow.
3. Record accepted outcomes, failures, human touches and actual usage.
4. Ask operations to confirm whether released time is usable.
5. Compare conservative, expected and upside scenarios without hiding quality failures in averages.
6. Repeat checks after changes to models, knowledge, tools or routing.

06

Agree when to expand, improve or stop

Define release-blocking failures before the pilot. A serious access breach or unauthorised action is not acceptable because the average ROI is positive. Keep risk controls and relevant human approvals as constraints, not optional costs to remove until the spreadsheet works.

Expand when useful completion, cost and ownership are demonstrated. Improve a narrow bottleneck when the evidence identifies one. Stop or simplify when the agent mainly moves work into a review queue, cannot obtain dependable information, or offers no advantage over fixed automation.

The NIST AI Risk Management Framework is voluntary guidance for governing and evaluating AI risk, not a return benchmark or certification. Discuss the first useful release through LCR’s AI-agent offering without assuming a guaranteed saving.

Questions

Frequently asked questions