The problem with "measuring AI impact"
When an AI project reaches the board, the question is not whether the technology works. The question is: what do we have to show, and how soon?
Most projects fail at this point not because the implementation is poor, but because no one defined what would be measured before the work began. At 90 days, the technology team talks about models and pipelines. The CFO needs to talk about hours, euros, and risk.
This article proposes a measurement framework for the first 90 days that holds up in a board meeting: no inflated projections, no vanity metrics, no promises that cannot be verified.
Before you start: a baseline is non-negotiable
There is no impact metric without a point of comparison. Before activating any agent or automation, the team must document the current state of the process being addressed.
This includes, at a minimum:
- Time invested per person in the process (weekly or monthly hours)
- Frequency of errors or rework (number of incidents per cycle)
- Full cycle time (from process initiation to output delivery)
- Estimated direct cost (hours × hourly cost of the profile involved)
Without this baseline, any number presented to the board will be open to challenge. With it, the CFO can defend the delta using the company's own data—not third-party benchmarks.
Weeks 1 to 4: activation metrics
In the first four weeks, the goal is not to demonstrate ROI. It is to demonstrate that the system works under real conditions.
The relevant metrics at this stage are:
- Team adoption rate: how many people are using the agent or workflow in their daily work? A system no one uses generates no value, regardless of what it does in theory.
- Task completion rate: what percentage of tasks assigned to the agent are completed without human intervention?
- Response or processing time: how long does the system take to produce an output compared with the manual process?
These metrics do not go to the board yet. They are internal signals for adjustment before the project is formally evaluated.
Weeks 4 to 8: operational metrics
From week four onward, the project must begin producing data comparable to the baseline. The metrics a CFO can use are:
Hours recovered per process
If the invoice reconciliation process took 12 monthly hours per person and now takes 3, the delta is 9 hours. Multiplied by the hourly cost of the profile and by the number of people involved, this yields a concrete monthly avoided cost.
Reduction in errors or rework
In processes with high manual load—data entry, order validation, report generation—errors are frequent and carry real cost: correction time, delays in financial close, risk of decisions made on incorrect data. Measuring the reduction in incidents per cycle is direct and verifiable.
Cycle acceleration
If the monthly close moved from 5 days to 3, that has financial value: faster decisions, reduced exposure to stale data, the ability to react sooner. This type of metric is intelligible to every board member, not just the CFO.
An example with range-based assumptions
A mid-size manufacturing company—between 80 and 200 employees—had an operational reporting process that involved three people from the operations team for approximately 15 monthly hours each. The process included manual consolidation of production data, cross-validation with the ERP, and generation of management reports.
After deploying a consolidation and report-generation agent in the first six weeks, time per person dropped to between 3 and 5 monthly hours. The delta was 10 to 12 hours per person, equivalent to between 30 and 36 monthly hours recovered across the team.
Assuming an average hourly cost of 25 to 35 euros for that profile, the monthly avoided cost falls between 750 and 1,260 euros. Over 12 months, between 9,000 and 15,000 euros in direct cost—before accounting for improvements in data quality or the reduction in consolidation errors.
These figures are not guarantees. They are range-based assumptions built on the company's actual baseline. That is what makes them defensible.
Weeks 8 to 12: cumulative value metrics
By day 90, the board needs a consolidated read. The metrics that work at this point are:
- Cumulative avoided cost (sum of the deltas from prior weeks, expressed in euros)
- Freed capacity (recovered hours that the team has redirected to higher-value tasks)
- System stability (how many times did it require unplanned manual intervention?)
- Operational team satisfaction (a simple qualitative metric: does the team use it because they want to, or because they are told to?)
This last point matters more than it appears. A system the team adopts naturally has a far longer useful life than one that requires constant pressure to be used.
What not to measure in the first 90 days
Some metrics are premature or outright misleading at this stage:
- Total project ROI: at 90 days, the system is still being refined. Presenting a definitive ROI at this point is premature and can set incorrect expectations.
- Industry benchmark comparisons: every company has its own baseline. Comparing against external averages distorts the evaluation.
- Technical metrics without a business translation: model accuracy or system latency are internal data points. The board needs hours, euros, and risk.
Conclusion
The first 90 days of an AI project are not for demonstrating that the technology is sophisticated. They are for demonstrating that it produces measurable results under real conditions, using the company's own data.
A CFO who arrives at the board with a documented baseline, a verifiable delta, and a conservative 12-month projection holds a strong position. One who arrives with generic promises does not.
If you want to review which processes in your operation have the conditions to produce that kind of metric in 90 days, you can request a free diagnostic. No commitment, no prior call required: complete the form and we will respond with an initial assessment within 48 hours.
