The 3 metrics that prove or disprove ROI

After 90 days in production, every agent project gets a financial review. The owner or the CFO asks: "Is this paying off?" There are three numbers that answer the question. Other metrics are interesting but not load-bearing. If you cannot produce these three at the 90-day mark, the project will lose its budget regardless of how well the agent is technically performing.

Metric 1: Hours saved per week

The most basic ROI input. For each workflow the agent runs, calculate: how many runs per week × average human time per run before the agent × (1 - share of runs the human still has to review fully). For a reservation-modification agent at 180 keys: 150 runs/week × 4.5 min/run × (1 - 0.15) = roughly 9.5 hours saved per week, or 41 hours per month.

Translate to euros at the loaded cost of the role being displaced (front-desk reservations: typically €18-28/hour fully loaded in EU markets). 41 hours × €23/hour = €943/month of capacity recovered. Compare to the all-in cost of the agent (model costs + engineering allocation + tooling): if the agent costs €350/month and recovers €943/month of capacity, the ROI is positive and defensible.

Metric 2: Quality delta

Hours saved alone is not enough — if the agent saves hours but quality drops, the owner will rightfully kill the project. Quality delta measures how the agent's output compares to the human baseline for the same workflow. Three sub-measures:

Metric 3: Incident count

Hours and quality are averages. Incidents are the tail — and the tail is where projects die. An incident is any agent action that required human-noticeable correction: a wrong PMS modification reversed, an incorrect email retracted, a guest complaint traceable to agent behavior, a regulatory or financial issue caused by an agent action. Track count and severity weekly.

Target: ≤2 minor incidents per month, 0 severe incidents per month. Severe is defined upfront: anything involving a payment error >€100, anything that reaches a guest review, anything that requires a written apology from the GM. If severe incidents are non-zero in any given month, the project pauses for a root-cause analysis regardless of other metrics.

How to present these to the owner

One page. Three numbers, the trend over the last 90 days, and one paragraph of plain-language interpretation. Owners do not need to see the architecture diagram or the prompt; they need to see whether the project is paying off and what they would lose if they killed it. The page goes to the owner every month, not just at the 90-day mark. Frequent transparent reporting is how you build the institutional trust that lets agentic projects survive their inevitable bad weeks.