Why the First 90 Days Are the Real Decision Point
Most AI projects at mid-size companies don't fail at the technical implementation stage. They fail in the transition between the pilot and phase two.
The technical team delivers something that works. The vendor presents a demo. And the CFO has to decide whether to approve additional budget without a clear sense of what to measure or what to expect.
The typical outcome: the decision is deferred, the project loses momentum, and the team reverts to its previous processes.
This article proposes five metrics a CFO can require before that decision. These are not technical metrics. They are business metrics that any finance leader can interpret without needing to understand how the underlying language model works.
Metric 1: Actual Agent Adoption Rate
An agent in production that no one uses generates no value. The first question is not whether the system works, but whether the team is actually using it.
What to measure: the percentage of transactions or queries in the target process that go through the agent, versus those still being resolved manually or through alternative channels.
Reasonable threshold at 90 days: between 60% and 80% adoption within the specific process where the agent was deployed. If adoption is below 50%, there is an integration problem, a trust issue within the team, or a workflow design flaw.
Why it matters to the CFO: low adoption is the earliest indicator that the projected ROI will not materialize — not because the system is failing, but because the team is working around it.
Metric 2: Reduction in Manual Hours on the Target Process
This is the most direct metric for justifying the initial investment.
What to measure: weekly hours dedicated to the process before implementation versus after. Include both execution time and the time spent reviewing and correcting outputs.
Expected range: for reporting, reconciliation, or internal query-handling processes, a reduction of between 30% and 60% in manual time within the first 90 days is reasonable, depending on process volume and complexity.
Concrete example: a distribution company operating across three countries deployed an agent to consolidate its weekly inventory report. The process previously required between 6 and 8 weekly hours from a financial analyst. By week 10, the time spent on manual review and adjustment had dropped to under 2 hours. The analyst shifted from building outputs to validating them.
Why it matters to the CFO: it converts the project cost into a comparable number. If the agent frees up 20 monthly hours from a role costing 3,000 euros per month, the payback calculation stops being abstract.
Metric 3: Output Accuracy Rate
A system that automates but introduces errors creates more work, not less.
What to measure: the percentage of agent outputs that require manual correction before use. This includes data errors, incorrect formatting, omissions, and misinterpretations.
Reasonable threshold at 90 days: in structured processes, an error rate above 10% signals that the system needs adjustment before scaling. In critical processes such as financial reporting, the threshold should be stricter.
Why it matters to the CFO: accuracy determines whether the system reduces or merely redistributes the workload. An agent with a high error rate doesn't free up time — it shifts it toward review and correction.
Metric 4: Operational Cost per Automated Transaction
This metric connects AI infrastructure spending to the volume of work processed.
What to measure: total monthly system cost (licenses, APIs, governance time) divided by the number of transactions or tasks the agent completed during that period.
What it's used for: it allows you to compare the cost of automating a task against the cost of executing it manually. It also allows you to project what happens to unit cost as volume scales.
Warning signal: if the cost per automated transaction is similar to or higher than the equivalent manual cost, the business case doesn't hold at current volume. It may hold if volume grows, but that assumption must be explicit in the projection.
Why it matters to the CFO: it is the metric that lets you assess whether the model is scalable, or whether the savings depend on a minimum volume that has not yet been reached.
Metric 5: Estimated Payback Period on the Initial Investment
This is not a 90-day metric in itself — it is the result of combining the previous four.
How to calculate it: initial project investment divided by the net monthly savings generated (hours freed, valued at actual cost, minus the operational cost of the system).
Reasonable range for mid-size projects: between 4 and 10 months for medium-volume processes. If the calculation exceeds 18 months, the project requires a review of scope or underlying assumptions.
Why it matters to the CFO: it is the number that shifts the conversation from technology to profitability. And it is the number that justifies — or doesn't — approving phase two.
How to Use These Metrics Before the Review Meeting
The practical recommendation is straightforward: before the 90-day review meeting, ask the team responsible for the project to produce a report that answers these five questions with real data, not estimates.
If the team cannot answer them, that is not necessarily a sign that the project has failed. It may indicate that success criteria were not defined at the outset — a methodology problem worth correcting before scaling.
If the team can answer them and the numbers are reasonable, the decision to continue has a solid foundation. If the numbers are not reasonable, the conversation about adjustments can be concrete rather than speculative.
Conclusion
Approving phase two of an AI project without clear metrics means accepting unnecessary risk. The five metrics described here require no technical knowledge to interpret. They do require that the project was designed from the start with measurable business criteria.
If you are evaluating an AI project at your company and want to assess whether the current indicators are sufficient to support that decision, we can work through that analysis together in a brief call.
