How to measure AI return without inventing numbers
Most AI return figures are fiction constructed after the fact. Avoiding that is simple: you pick two or three measures before you start, using data the company already has.
Measure beforehand, on the task, not on the company
You cannot measure a company's overall return on AI — you measure one task. How long it takes today, how often it happens per week, how many errors it produces. One week of honest recording is enough for a credible baseline.
If you don't have the starting number, you don't start: you record for a week first.
Use data that already exists
Email response times, number of documents processed, deadlines met, repeat calls. This data is already in the company's tools and doesn't depend on anyone filling in new spreadsheets.
Always prefer an imperfect measure that already exists to a perfect one that requires new work.
Count the whole cost, including your own time
Licences, build hours, training hours and supervision hours. If the calculation only counts the licence, the return is always spectacular — and false the moment the project needs maintenance.
Include maintenance in the calculation from the start: one to two hours a month, per workflow.
The frame I use with clients
- One task, two numbers: time per occurrence and occurrences per week.
- One total cost: licences, implementation, training and annual maintenance.
- A review date at 30 and at 90 days, set before starting.
- An explicit decision at each review: keep, extend or switch off.
Want to know whether your case can be automated? Let's talk for 30 minutes.
Book a call