AI Agent ROI: Building the Business Case (with Numbers)

An AI agent ROI case has three parts: your current process's real cost (the baseline), the agentic system's build-plus-run-plus-oversight cost, and the payback timeline that connects them. Most business cases fail not because the math is wrong, but because they skip cost categories that show up later — evaluation time, monitoring, and model spend variability, specifically.

Below is one worked example, end to end, with every assumption visible — copy the shape, not the numbers, since your own baseline will differ.

Worked example: a 40-person ops team's invoice-exception process

Baseline (before): 3 FTEs spend roughly 40% of their time resolving invoice exceptions — mismatched PO numbers, unusual line items, missing approvals. At a fully-loaded cost of $75,000/year per FTE, that's roughly $90,000/year of labor cost tied up in this one process.

Agentic cost (build + run + oversight):

| Cost category | One-time / Year 1 | Ongoing (Year 2+) | |---|---|---| | Build (discovery, integration, testing) | $35,000–$55,000 | — | | Model/inference spend | $6,000–$14,000/year | $6,000–$14,000/year | | Monitoring & evals | $8,000/year (0.15 FTE) | $8,000/year | | Human-in-the-loop oversight | $12,000/year (0.25 FTE) | $12,000/year | | Total Year 1 | $61,000–$89,000 | — | | Total Year 2+ | — | $26,000–$34,000 |

Payback timeline: against a $90,000/year baseline, Year 1 net savings land between roughly $1,000 and $29,000 depending on build cost — a thin first-year case on its own. Year 2 onward, at $26,000–$34,000 in ongoing cost against the same $90,000 baseline, net savings run $56,000–$64,000/year, meaning full payback of Year 1's build cost typically lands between month 10 and month 16.

What cost categories do most business cases forget?

• Evaluation time — someone has to build and run test cases before go-live and periodically after; this isn't free and isn't optional.
• Monitoring — an agent in production needs someone watching for drift, not a "set it and forget it" deployment.
• Model spend variability — inference pricing changes; a business case that assumes today's per-token cost forever is fragile.
• Change management — the 3 FTEs whose work changes need training and a real transition plan, not just a memo.

Skipping any of these doesn't make the cost disappear — it just makes it a surprise instead of a line item.

Where do our 35–50% and 20–40% figures come from, and where do they hold?

These are AIGist24's own benchmark ranges from client engagements, not third-party research. 35–50% efficiency gains show up on exception-heavy, judgment-light processes — the invoice-exception example above is a representative case. 20–40% is the range for processes that are more deterministic but still benefit from removing manual triage. They don't hold universally: a process that's already highly optimized, or one where the "automation" would just move the bottleneck somewhere else, won't see numbers in this range — treat these as planning ranges for the process shapes described in our AI readiness checklist, not a guarantee for any given process.

What should a one-page board memo include?

A board-readable ROI memo needs exactly five sections, in this order:

1. The problem, in one sentence — what process, what's it costing today.
2. The baseline, in numbers — current cost and cycle time, sourced (not estimated on the spot).
3. The proposed solution and its full cost — build + run + oversight, not just build.
4. The payback timeline — when net savings turn positive, with the assumptions stated.
5. The risk section — what could make this wrong, and what the rollback plan is.

A memo missing section 5 reads as sales copy, not a business case — boards notice the difference.

Want the bespoke version of this model?

The worked example above uses illustrative numbers. Our AI readiness assessment builds this exact ROI model against your real baseline data as one of its core deliverables — the fixed-price, two-week version of the exercise above, with your actual numbers instead of a placeholder scenario.

How should you sanity-check a vendor's ROI claims?

Any ROI figure a vendor hands you (ours included) should survive three checks before it goes in front of your board: is the baseline sourced, not estimated? — ask where the "before" number came from, and be suspicious of a round number nobody can trace to a system of record. Does the cost side include run-rate, not just build cost? — a business case that only shows the build invoice is showing you half the picture; ongoing inference, monitoring, and oversight costs don't disappear after go-live. Is there a stated assumption for what could go wrong? — a business case with no risk section is marketing copy wearing a spreadsheet's clothes. A vendor confident in their numbers will welcome all three questions; one who can't answer them plainly is a signal worth taking seriously before you commit budget.