The short answer
Defensible AI agent ROI counts four things: direct labor hours reclaimed, error and rework costs avoided, cycle-time value (when faster output produces incremental revenue), and opportunity cost on the headcount freed. It subtracts implementation cost, ongoing model and infrastructure cost, evaluation and monitoring cost, and the cost of human review time you still need. The traps: counting fully-loaded labor that nobody actually reallocates, double-counting cycle-time and revenue, and ignoring the eval and ops cost that arrives in month 4.
The four legitimate sources of agentic AI ROI
Most ROI models inflate by mixing apples and oranges. A defensible model separates four distinct sources and only counts each once.
- ›Direct labor: hours per task × number of tasks × loaded hourly cost × % of work actually displaced.
- ›Error and rework: error rate × rework cost × volume — measured against the agent's evaluated error rate.
- ›Cycle-time value: when faster output produces revenue (proposals out the door faster close at a higher rate; faster invoicing reduces DSO).
- ›Opportunity cost: the dollar value of work the freed humans now do — only counted if leadership commits to the reallocation in writing.
The four costs every honest model subtracts
Implementation cost (one-time), ongoing model and infrastructure cost (per-month), evaluation and monitoring cost (per-month, often understated), and the human-review time you still need (per-task — never zero for a well-governed agent).
The three traps that inflate vendor spreadsheets
First, counting fully-loaded labor when nobody actually reallocates the freed hours — they evaporate into Slack. Second, double-counting cycle-time and revenue: a faster proposal does not produce both a cycle-time saving and a revenue lift independently. Third, ignoring the operational cost that arrives in month 4 — model versioning, prompt regression testing, drift monitoring, and on-call rotation.
A simple defensible model
For a Tier-1 support deflection pilot: 600 tickets/week × 12 min × $35/hr × 50% genuinely displaced = $109K/yr in direct labor. Error reduction from inconsistent macros: 4% × $200/error × 31K tickets/yr = $248K/yr. Subtract: $40K build, $36K/yr inference, $24K/yr evaluation and on-call, $14K/yr residual human review of the 50% non-displaced tickets. Net year-one: ($109K + $248K) − $40K − $36K − $24K − $14K = $243K. Year-two recurring: $283K. That is a defensible number you can put in a board pack.
Key takeaways
- Count four sources of value and four costs — separately, never mixed.
- Only count opportunity cost if leadership has committed in writing to the reallocation.
- Budget for evaluation, monitoring, and on-call from day one — they are not optional.
- If the model still shows positive ROI after honest accounting, the pilot is real.
Frequently asked questions
What's a typical payback period for a first agent pilot?+
Most well-scoped first pilots pay back in 6–14 months on the conservative model. Pilots that promise sub-3-month payback are usually counting opportunity cost that nobody reallocates.
Should we count headcount reduction as savings?+
Only if it is actually planned. Counting headcount savings that never happen is the fastest way to lose credibility with finance after pilot two.
How do we measure error costs we don't currently track?+
Sample 50 historical outputs, score them against ground truth, and extrapolate. It is not perfect but it is defensible — and far better than the zero that most pilots assume.