1. Measure the situation before the agent
During a representative period, count the number of completed tasks, their actual duration, resumptions, deadlines and exceptions. Don’t start from a general memory. For emails, sample multiple days; for quotes, several files of different complexity.
The reference must use the same definition as the driver. If the starting point measures a finalized estimate, the pilot cannot count a simple draft as an equivalent result.
2. Calculate the net time, not the theoretical time
The gross time saved corresponds to the initial time minus the assisted preparation time. We must then subtract validation, corrections, exceptions and the time spent supervising the device. The hourly value used must be that chosen by the company and remain visible in the calculation.
Net time = initial time − time with the agent − validation − corrections. Value of time = net time × documented internal hourly cost.
3. Include the full cost
The cost is not limited to the subscription. Add possible start-up, integrations, training, monitoring, variable consumption and internal time spent on the pilot. Distinguish one-off costs from monthly costs so as not to artificially penalize subsequent months.
Errors also have a cost. Faster but less reliable automation can move work instead of removing it. Track recovery rate and incidents by category.
4. Decision formulas
Monthly net profit= value of net time + value of avoided errors + attributable additional income − full monthly costs.
ROI over the period= (observed benefits − full costs) ÷ full costs. The result must always specify the period and the assumptions. When the attribution of income is uncertain, present it separately rather than adding it as a fact.
5. The minimal dashboard
- Volume of eligible and actually processed tasks.
- Initial time, assisted time, validation and corrections.
- Result rate accepted without major modification.
- Number and type of exceptions.
- Fixed, variable costs and internal time.
- Adoption: active people and real frequency.
- Decision documented at the end of the pilot.
A zero result is not a measurement failure. It helps avoid extending automation that does not create value. A positive result on a small sample does not allow immediate generalization either.
Swiss Agent Network does not publish a generic percentage of time saved. Public results must indicate the authorized client or scope, period, volume and method.