Did your AI project actually save time? How to tell
Everyone reports that it helps. Almost nobody can show it. Here is a measurement that takes an hour and survives scrutiny.
Ask anyone using AI at work whether it saves time and they will say yes. Ask by how much and the answer becomes vague. That vagueness is a problem when you are deciding whether to extend the programme or quietly let it lapse.
Why the usual measurement fails
Companies measure the wrong thing twice. They count how often the tool is used, which measures enthusiasm rather than value. Or they ask people to estimate hours saved, which produces numbers that are consistently too high and impossible to defend.
Both are worse than not measuring, because a number nobody believes ends the conversation as effectively as no number at all.
The measurement that works
Pick the one automated process. Measure three things, before and after, over two weeks each.
Throughput per person-hour. Invoices processed, enquiries answered, documents handled, divided by the hours actually spent on it. This is the number you can defend, because both halves are countable.
Error rate. How many needed correcting afterwards. If this went up, your time saving is borrowed from somewhere else and will be repaid with interest.
Elapsed time. How long between an invoice arriving and it being booked. Often the largest real improvement and the one nobody counts — it shows up in supplier relationships and discount deadlines rather than in hours.
The honest arithmetic
Two hours a week saved in a Swiss SME is roughly a hundred hours a year. At a loaded cost of eighty francs, that is eight thousand francs against software costing perhaps two.
That is a good result and an unglamorous one, and stating it plainly is more persuasive than a claim of transformation. A managing director who has been told three times that something is transformative will believe a number with an arithmetic behind it.
When the honest answer is no
Sometimes the measurement shows nothing. The process was already efficient, the volume was too low, or the exceptions ate the gain.
Say so and stop. A programme that has cancelled one project on evidence has far more credibility for the next one than a programme where everything has been a success.