What is an AI audit, and what should it cut?

Guides to AI automation tend to share a shape. List your repetitive tasks, pick the ones with the most volume, find a tool for each. The result is a long list of things to build, and the assumption that every one of them is worth building.

In practice, one of the most useful things an audit does is shorten that list. A lot of what looks like a candidate for automation turns out not to be, once you look at what it would cost to build, run and maintain against what it would save.

01Why does the list of ideas shrink?

Every automation has a running cost that is easy to forget at the start. Someone has to watch it, update it when a form changes, handle the cases it gets wrong and explain it to new people. For work that is rare or already quick, that cost can be higher than the time it gives back.

• The task happens too rarely for a system to pay back.
• The inputs change so often that the automation would break every month.
• The judgement involved is the valuable part, and taking it away would make the work worse.
• A small change to how the team works removes the task entirely.

02Which problems are process problems, not software problems?

Some of the slowest work we see is slow because of how it is organized, not because nobody has automated it. Two people do the same check. A form asks for information that is never used. An approval goes to someone who always says yes. Automating these makes a bad process faster. Fixing them is often free.

We point these out even though there is nothing for us to build. A roadmap that is honest about what not to do is worth more than one padded with projects.

03Can the tools you already pay for do it?

The other thing that comes up in most audits is how much the existing tools can already do. CRMs, help desks, accounting software and project boards have added automation and AI features quickly, and most teams use a small part of what they pay for. Turning on and configuring what is already there is often the fastest first win, with no new login for anyone.

04How do we decide, case by case?

Every idea on the list goes through the same short set of questions. None of them is a formula, but together they make it hard to argue a weak project onto the roadmap.

1. Step 01Is it worth the time?Minutes per occurrence times how often it happens, plus the cost of getting it wrong. If the total is small, it stops here.
2. Step 02Is it stable?If the inputs, the rules or the people involved change every few weeks, an automation will spend more time being fixed than working.
3. Step 03Is the judgement the point?Some work is slow because it needs care. Speeding up the routine parts around it can help. Removing the judgement usually does not.
4. Step 04Is there a simpler fix?A changed form, a removed step or a feature already in your tools. If one exists, it goes first.

Ideas that fail these questions are not thrown away. They are written down with the reason, so when something changes, such as volume growing or a process settling down, they can be picked up again without starting from scratch.

05What is left worth building?

After the list has been cut down, what remains is usually a handful of workflows where AI clearly earns its place: high volume or high value, stable enough to automate, and painful enough that the team will notice the difference. That short list is the one worth spending money on.

It also makes the first project easier to judge. When you only build the things with a clear before and after, you know quickly whether it worked, and the next decision is based on evidence rather than enthusiasm.