Why the First 90 Days Are the Critical Window
Deploying an AI agent is not the end of the process. It is the beginning of the most important phase: proving that it works.
In most mid-size companies, financial leadership approves the project with reasonable expectations. But without a clear measurement framework, those expectations become vague. Three months in, no one knows with certainty whether the agent is generating value or simply running.
That information gap is the problem. Not the technology.
The first 90 days are the window in which internal confidence in the initiative is either consolidated or lost. A CFO who lacks concrete data during that period cannot defend the project to the board, nor justify the next phase of investment.
This article describes the metrics that close that gap.
The Metrics That Don't Work (and Why They Get Used Anyway)
Before defining what to measure, it is worth setting aside what provides no useful information at this stage.
Number of interactions processed. An agent can process ten thousand queries and have saved a single euro if those queries were not the actual bottleneck.
Agent response time. The fact that the agent responds in two seconds is irrelevant if the process it automates previously took two days for reasons unrelated to response time.
Internal user satisfaction. This is a useful long-term indicator, but in the first 90 days it fluctuates for reasons that have nothing to do with the agent's value — learning curve, resistance to change, configuration adjustments.
These metrics appear in many reports because they are easy to obtain. Not because they are the ones that matter.
The Five Metrics a CFO Should Require
1. Operational Hours Recovered per Week
This is the baseline metric. Identify the process the agent automates, calculate how many hours the team devoted to that process before deployment, and measure how many it devotes now.
This is not an estimate. It is a concrete measurement that requires a time log prior to deployment — something that must be established before launch, not after.
Example: A distribution company with 80 employees deployed an agent to manage invoice reconciliation with suppliers. The process occupied two people for three days at each month-end close. After deployment, the same process took four hours with minimal supervision. The estimated savings: between 40 and 50 hours per month, equivalent to an operational cost of between 1,200 and 1,800 euros per month depending on the team's profile.
2. Error Rate Before and After
Manual processes carry an inherent error rate. Incorrect reconciliations, duplicate data, empty fields, outdated document versions. That rate has a cost: correction time, rework, decisions made on inaccurate information.
Measure the error rate of the process before deployment and compare it with the rate afterward. A reduction of 20% to 40% in operational errors is a reasonable range for structured data processes in the first 90 days.
This figure is especially relevant for the CFO because errors in financial processes have direct consequences: duplicate payments, incorrect provisions, reports that must be redone.
3. Cost per Automated Task
Divide the total cost of the agent during the period — licenses, infrastructure, configuration time, and maintenance — by the number of tasks processed. Compare that figure with the unit cost of the manual process.
If the agent processes 500 expense-approval requests in a month and the total system cost is 800 euros, the cost per task is 1.60 euros. If the same manual process cost between 8 and 12 euros per task in staff time, the difference is clear.
This metric makes it possible to project ROI at 12 months using real data, not assumptions.
4. Effective Adoption Rate
An agent the team avoids using generates no value. The adoption rate measures what percentage of tasks eligible for automation actually go through the agent versus those the team continues to handle manually.
An adoption rate below 60% in the first 90 days is a warning signal. Not necessarily of a technical problem, but of an integration problem within the team's actual workflow.
This metric is a shared responsibility between the vendor and management. If the agent is available but the team is not using it, there is a process or training issue that must be resolved before scaling.
5. Time to First Human Intervention
In processes where the agent operates autonomously, measure how frequently it needs to escalate a task to a person. This indicator reflects the agent's maturity within your company's specific context.
A well-calibrated agent should require human intervention in fewer than 15% of cases for structured processes. If the percentage is higher, there is a configuration issue or an input data quality issue — both correctable, but ones that must be identified early.
How to Structure the First-90-Day Report
A complex dashboard is not necessary. A monthly report covering these five metrics, compared against the pre-deployment baseline, is sufficient to give the CFO genuine visibility.
The recommended format is straightforward: value before, value after, percentage change, associated cost. No additional narrative. The numbers speak for themselves if they were measured correctly from the outset.
The most common mistake is failing to establish the baseline before launching the agent. Without data on the prior state, any subsequent comparison is an estimate. And estimates do not justify investments.
Conclusion
An AI agent that is not measured is a cost. One that is measured correctly is the case for the next phase of investment.
The first 90 days are not for celebrating the deployment. They are for building the business case with real data. The five metrics described in this article — hours recovered, error rate, cost per task, effective adoption, and human intervention — are the minimum starting point for a CFO to maintain control over what is happening.
If your company is in the process of deployment or evaluating a first agent, and you want to define the measurement framework before launch, we can help you structure it.
Request a free diagnostic and in a brief conversation we will define what to measure, how to measure it, and what to expect in the first three months.
