Case study · Realize · AI value realization
An AI value measurement system, built after the fact for a Swiss engineering group
How a Swiss engineering group rebuilt baselines, instrumented five workflows and set explicit gates to decide what 18 months of AI spending was worth.
Anonymised case — representative example from real Numezis engagements
Starting situation
Eighteen months after rolling out office copilots and then an internal document assistant, the group had committed roughly CHF 300k in licences and projects — without a single value indicator. Feedback remained anecdotal: some employees swore the assistant saved them hours, others had quietly stopped using it. When the board asked what AI was actually bringing the group, the CEO had nothing to present.
Two camps had formed internally. The CFO, seeing no evidence of return, wanted to cut the budgets at the next review; the teams, convinced by their own practice, were pushing to extend the tools. Neither side was demonstrably wrong — nobody had a baseline. Usage data sat with the vendors in poor export formats, and billable time, the one measure everyone understood, was too sensitive to handle carelessly.
The decision on the table
Scale, adapt or stop each initiative — on what evidence?
Real constraints
No historical baseline
Nothing had been measured before the tools were deployed. Pre-AI processing times existed nowhere — they had to be reconstructed without being invented.
Scattered usage data
Usage statistics lived with the vendors, in poor, heterogeneous and sometimes non-existent exports. No consolidated view was possible as things stood.
Measure without surveilling
The system had to prove the value of workflows without ever becoming a tool for monitoring individuals — a red line set on day one.
Billable time is sensitive
In a group that lives from billing hours, any measure touching billable time is politically delicate. Every indicator was designed with that reality in mind.
What we built
We did not deploy yet another tool: we built the measurement system that was missing. For five high-stakes workflows, a value tree links AI usage to a verifiable economic effect, a reconstructed baseline provides the point of comparison, and gates defined in advance state — before the numbers are known — what justifies scaling, adapting or stopping.
System components
Delivery sequence
Framing & value trees 3 weeks
Inventory of the seven AI initiatives under way, selection of 5 high-stakes workflows, construction of the value trees with business leads, first alignment with CEO and CFO on decision criteria.
Baselines & instrumentation 5 weeks
Reconstruction of baselines through sampling and historical data, briefing of the staff representation, light workflow-level instrumentation, definition of the gates with CEO and CFO.
Pilot measurement cycle 8 weeks
Two full monthly cycles: collection, dashboard, reviews with CEO and CFO, scale / adapt / stop decisions prepared and documented for the 7 initiatives — including two deliberate stops.
Governance & handover 4 weeks
Handover of the system to the client analyst, integration of the value review into the existing monthly governance, documentation of the attribution method and the gate rules.
Measured results
The first result is not a number: it is the ability to answer the board with evidence. The metrics below compare the state of the program at the start and at the end of the five-month engagement.
Figures rounded. Anonymised case: a representative example drawn from real Numezis engagements, not a nameable reference.
What we would do differently
This section is part of our editorial standard: no case study without its lessons.
- 01
Negotiate access to usage data at licence purchase. Vendor exports turned out to be poor, and we lost three weeks reconstructing what a contractual clause would have delivered from day one.
- 02
Involve the staff representation from week 1, rather than only briefing it before instrumentation. Two weeks of clarifications about what the system measured — and did not measure — would have been avoided.
- 03
Freeze the gates — not merely define them — before the first measurement. The temptation to move the target once the numbers are known is real, and we experienced it: only thresholds made non-negotiable up front can resist it.