What I refuse to automate

Every time I map a company, I end up with two lists: what we're going to automate, and what we're going to leave alone.

The second list is usually longer.

That can feel strange when you've brought someone in to help you use AI. But once you look at how the work actually gets done, you find plenty of things that need a different fix.

Take the weekly report that three people put together and nobody reads. You could get AI to write it. I'd first want to know what happens if you stop sending it.

If nobody can name a consequence, that's a good candidate for deletion.

The same goes for work created by a problem earlier in the process. If incoming documents constantly need correcting, look at how they arrive. A clearer intake process might remove most of the corrections before you build anything.

Then there's the work I deliberately keep with a person.

A complaint. Bad news for a client. The final decision on a big deal. These are moments where a mistake can cost you a relationship, or more than the automation would save in a month.

AI can prepare the file, flag something that needs attention, or draft a response. I'd still want a person to make the call and have the conversation.

The biggest category in my process maps is usually "not yet."

Every company has someone who knows all the exceptions. Which customer pays late but always pays. Which supplier needs a phone call. When it's fine to bend a rule.

If all of that lives in one person's head, you have some work to do before you can automate their job. Sit down with them and write it out, including the exceptions. You'll need those rules to explain the work to an agent and check whether it's doing it properly.

Sometimes the problem is measurement. You want AI to flag customers who might leave, but you don't track why customers leave today. You need a way to check its predictions.

And sometimes there simply isn't enough work to justify maintaining an agent. My rule of thumb: if a task happens less than weekly and takes under an hour, give the person a chat assistant and move on.

After those decisions, the automation list gets much clearer.

Chasing missing information, copying data between systems, routine status updates. Work that happens often, follows rules you can write down, and produces an output you can check. The mistakes also need to be affordable.

Even then, I start with a person approving every output for the first 30 days.

If you've got a process in mind, run it through these five questions in order:

1. What breaks if we stop doing this?

2. What does one wrong answer cost?

3. Can we write down how it's done, including the exceptions?

4. Can we check the output with a number?

5. Does it happen often enough to pay for the upkeep?

Your answers should tell you whether to delete it, keep it human, prepare it for later, or automate it now.

Revisit the list each quarter. A process that's stuck in someone's head today might be ready once you've written down the rules. "Not yet" should come with something to work on.

Robin

PS: If you'd like to work through this for your company, book a call. We start with one conversation per department to map the work and decide where AI belongs.