A process audit before AI implementation examines how work is actually done in a company: what people do, how long it takes, which data and tools they use, and where delays and errors occur. The findings show which steps are worth automating, where AI would only add work, and where a person is essential.
The first audit question is: how is the work done now, and where is a problem worth solving? The question of which tool to buy comes last. The audit looks at a real working week, because a job description describes it about as well as a menu describes a kitchen. The only thing this has in common with an internal audit or an ISO check is the word “audit”: no one is being evaluated here.
The audit examines:
A process map with problem areas marked. Automation opportunities, with an estimate of the expected effect of each. Constraints and risks: where AI may make mistakes, who will notice, and whose work would stop without a person. A list of ideas without estimates does not count as an audit result.
The audit shows the current picture; the implementation decision comes afterward. The selected task is described in an AI specification and tested in a pilot before it is rolled out to the whole department.
Run the audit before buying a tool and before making any decisions about people. After implementation, it becomes an autopsy: it explains what went wrong after the money has been spent. Before implementation, it answers questions that shape the budget. Where will AI remove routine work? What can tools you already pay for do? Who will it help?
A university classmate of mine has been writing code for more than ten years and now leads a team. One day, his company gave everyone an AI assistant subscription and announced that no one would write code by hand anymore. They gave no instructions and measured no increase in speed. He made it work because he had a lot of experience and the tasks were small. No one knows how people with less experience managed. No one thought to ask.
I made the same mistake myself as a contractor. An acquaintance received 100–200 emails a day and spent about two hours sorting them. We built automatic sorting with Zapier. While we were building it, Google added a Gemini button to Gmail that sorts email into labels on the plan he was already paying for. My automation added little beyond reminders about important emails. No one measured the hours saved.
Both stories skipped one step. No one looked at how the work was organized before changing it.
What to examine: a whole department or a single workflow. Who will receive the findings and make decisions. And how to explain to people why they are being asked. That last point matters more than it may seem: if employees think this is a check before layoffs, they will recite their job descriptions.
Related terms: Process map · Pilot · Failure point · Total cost of ownership
Planning to implement AI in a department? AI audit for a department →