A baseline is the measure of how a process works today, before any change, measured as time per task, error rate and time to response. Without it you cannot prove that an AI system saved anything, or decide whether it is worth building. The way to measure a process before automating it is to spend two to four weeks on the records the process already leaves and a sample of real cases.
Why AI only pays off where work can be measured
The AI Index 2026 from Stanford University brings together the most solid studies on productivity with AI. The measured gains show up in work broken into repeatable tasks, with quality that is easy to check. In customer support, cases resolved per hour rose 14% to 15%. Output per person in ad creation rose 50%, and in an accounting study clients served per week rose 55%.
In judgment work, with no clear way of knowing whether the result is right, the gain is small or negative. A METR study found that experienced developers were 19% slower with AI on projects they knew well, and believed they were faster.
The lesson for anyone looking to automate a process is that, before choosing a tool, it pays to make the process measurable. If you cannot tell within a few seconds whether a task’s result is right, the first job is to define that measure.
The three measures that count
Each process has its own, but they almost always come down to three rates, defined before starting and identical before and after.
- Time per task. How long it takes to handle a case, from the moment it arrives to the moment it is resolved. Include the time of whoever does it, whoever reviews it and whoever corrects it.
- Error rate. What share of cases has to be redone, corrected or sent back. An error counts when someone had to touch the case again, even if it looks minor.
- Time to response. How long the customer, supplier or next team waits. Often a case takes ten minutes to handle and sits in an inbox for three days.
All three are measured as rates (minutes per case, errors per hundred cases, hours to response) so they can be compared across months with different volumes.
How to measure without stopping the team
The most reliable way is to use what the process already records: emails with arrival and reply dates, ERP or CRM logs, the change history of a shared spreadsheet. Behavioural data is worth more than what people say they do, because everyone underestimates the time repetitive tasks take and forgets the interruptions.
Where there are no records, take a sample. For two weeks, whoever runs the process notes the start and end of each case on a simple sheet, with a column for what had to be corrected. Two weeks are enough for most processes, and four are better when volume varies a lot from week to week.
At the same time, keep a sample of real cases with the correct result. In an invoice process, for example, 50 to 100 invoices with the correct entry. This sample is the system’s test set: it is used to measure the system before go-live and after every change.
Compare with a simple rule
There is a question many AI proposals avoid: would a simple rule do almost as well? Before building, the baseline is compared with what a simple rule would do on the same cases (a format check, a lookup table, a formula in the spreadsheet that already exists).
If the simple rule handles 80% of cases, it pays to start there, and that is often what we recommend. AI comes in where the rule falls short, for example to read documents whose format changes from supplier to supplier. The split between what is a rule and what needs judgment is in rule-based tasks and judgment tasks.
What to do with the numbers
With the baseline measured, the rest of the project has a point of comparison.
- The decision to go ahead. Hours per month, multiplied by the fully loaded hourly cost, plus the cost of errors, give the most it makes sense to pay for the system. The full calculation is in hire, outsource or automate.
- The project target. When part of the price depends on the result, the target is written from the baseline, as we explain in paying for results.
- Proof for the team. Showing that invoice collection went from 30 hours a month to under 2 convinces more than any presentation (see the case). The baseline is one of the four adoption pieces described in why your team isn’t using the AI system.
Frequently asked questions
What is a baseline in an AI project?
The measure of how the process works before any change, usually with three rates: time per task, error rate and time to response. It proves the gain, decides whether to go ahead and sets the project target.
How long does it take to measure?
For most processes, two to four weeks. With stable volume two are enough, and when volume varies a lot from week to week four are better.
What if the process leaves no records?
Take a sample: for two weeks, whoever runs the process notes the start and end of each case and what had to be corrected. It is simple and enough to compare with the after.
Why compare with a simple rule?
Because if a check or a formula handles almost every case, it is cheaper, easier to maintain and easier to explain than an AI system. AI makes sense where the simple rule falls short.
Want to know what your process costs today?
It is the first thing we measure. Book a call and we will define the three measures with you and how to collect them.
