How to run an automation audit on your own team

You don’t need a consultant to find your first automation. You need one week, a shared document and a willingness to look at what your team actually does rather than what the process diagram says they do.

Here’s the exercise we run at the start of most engagements. It works just as well when you run it yourself.

Day one — capture, don’t analyse

Ask everyone to keep a rough log for one week. Not a timesheet — a log. Every time they do something that felt repetitive, mechanical, or like it should already be handled, they add a line: what it was, roughly how long, roughly how often.

Two rules make or break this. It has to take under ten seconds per entry, or people stop by Wednesday. And it must be explicitly not a productivity measurement, or people log what makes them look busy instead of what’s actually eating their week.

Say plainly: we’re hunting for work worth deleting, and the person who finds the most gets the most time back.

Day five — cluster

You’ll have a messy list. Group it by the shape of the work, not the department. Common clusters:

• Moving data between systems. Anything described as “then I copy it into…”
• Assembling documents. Reports, quotes, packs, summaries built by hand each time.
• Chasing. Following up on approvals, invoices, information, signatures.
• Classifying and routing. Reading something to decide where it goes.
• Looking things up. Hunting for the precedent, the clause, the previous answer.

Each cluster maps to a different kind of solution. Data movement is usually a plain integration and doesn’t need AI at all. Classification and routing is where models earn their keep. Looking things up is a retrieval problem. Assembly is templating plus a bit of judgement.

Day six — do the arithmetic

For each cluster: frequency per week, minutes per instance, number of people. Multiply out to hours per year, then apply a loaded hourly cost.

Do this honestly and the results are usually uncomfortable. A fifteen-minute task done twice daily by four people is around 500 hours a year. That’s a quarter of a full-time salary spent on something nobody chose to spend it on.

Then add the invisible costs the log won’t show you. Transcription errors and the time spent finding them. Delays while something sits in someone’s inbox. The opportunity cost of your best people doing mechanical work. These are frequently larger than the hours themselves and almost always left out.

Day seven — score before you build

Rank the clusters on four axes, one to five:

Value. The annual cost you just calculated.

Feasibility. Is the data accessible via an API, or trapped in a PDF someone scanned in 2019?

Verifiability. Can you tell quickly whether the output is right? Rules-based checks score high; subjective judgement scores low.

Change friction. How many people have to work differently, and how much do they want to? This is the axis that kills projects and the one nearly everyone omits.

Build the highest total first. It will rarely be the most exciting item on the list, and that’s the point.

What you’ll probably find

Three things come up in almost every audit we’ve run.

The biggest single win is usually in finance or operations, not in the customer-facing team where the attention is.

At least one cluster turns out not to need AI at all — it needs a fixed integration, or a form that captures the right fields the first time. Don’t reach for a model when a webhook will do.

And there’s always one task that everybody assumed was necessary, which on inspection nobody actually needs. Deleting it is the cheapest automation available and it costs nothing to ship.