Before buying a tool or launching a development project, an SME benefits from knowing where artificial intelligence truly pays off for it. This is the purpose of an AI audit, provided it focuses on work rather than technology. It answers three questions: which tasks AI can handle, how much this yields, and which one to start with.
The word “audit” suggests a control, but the exercise is of a different nature. It focuses neither on models nor on vendors, but on the company’s daily operations: quotes, follow-ups, reports, data entry, and checks.
When an AI audit is useful
Four situations justify using one. Teams are already using ChatGPT or Copilot, each in their own way, with no one knowing what it yields. Management has several project ideas but no method to choose between them. A vendor proposes a tool, and it’s necessary to verify it addresses a real need. Finally, a repetitive task occupies several people each week.
On the other hand, the audit is useless if the company is only looking to get to grips with market tools: training is sufficient then.
Five questions to classify tasks
Each candidate task goes through the same filter. The first question concerns the time it takes today and the number of people involved: without this initial measurement, no gain can be proven later. The second concerns frequency, as a task performed ten times a month does not justify a dedicated tool. The third asks if the data exists in a usable form, as AI does not compensate for missing information or data scattered across email inboxes.
The fourth question is about error: what happens if the AI makes a mistake, and who checks? The more costly the error, the more human validation is needed. The fifth concerns the cost of a first prototype and the timeline before it is usable. A significant gain requiring a year of work comes after a moderate gain available in a few weeks.
A use case is selected when the gain is measurable, the data exists, and a domain expert can check the result. Others are discarded, and the audit must explain why.
What to require as the output
The deliverable should not be a report on “the potential of AI.” It takes the form of a short list of use cases, each with an estimated gain, a cost, a timeline, and a clear recommendation on which one to launch first. The selected cases are rarely spectacular: generating a quote from a customer request, processing non-conformities in a factory, preparing the technical specifications for a tender in construction, or qualifying incoming requests for a services firm.
Two requirements are worth setting from the start. The report belongs to the company, which must be able to use it internally or with another provider. It must also be understandable by those who do the work, and not just by IT management.
Conversely, a flawed audit starts with technology, quantifies nothing, or recommends everything. “High potential” is not a result, whereas a number of hours per week is. In an SME, the exercise lasts a few days, starts from specific tasks, and does not assume teams are already using AI. The manager who has a list of quantified tasks chooses their first project based on hours and euros, not on a demo.
