AI doesn't fix a process nobody understands

I’ve noticed that almost every conversation about AI at work eventually lands on the same question.

“What can we automate?”

It’s a reasonable question. Most teams have no shortage of repetitive work. There are forms to fill in, messages to send, reports to update, and enough follow-ups to make you wonder what anybody did before calendars had reminders.

But I’ve been trying to ask a different question first.

Why are we doing this in the first place?

Not why the company needs onboarding, hiring, or performance conversations. I mean why that specific step exists, in that particular sequence, with those particular people involved.

Sometimes the answer is good. There is a real decision to make. A risk to manage. A person who needs an answer.

Sometimes the answer is, “We’ve always done it that way.”

And occasionally the answer is that someone added an extra form five years ago because another form wasn’t being completed. Now everyone completes both.

AI can absolutely help with that kind of work. But if we automate a bad handoff, we still have a bad handoff. It just happens faster and with a very confident summary at the end.

The automation trap

I’ve seen a version of this in our own onboarding.

There were times a new hire was about to start, but their manager hadn’t properly connected with them yet, or nobody was quite clear on which project they’d be joining.

The easy answer would have been more reminders. Another Slack message. Another checklist item.

But a reminder doesn’t help much when nobody is clear on who’s responsible.

So when I built our onboarding system, I started with a simple rule: every new hire needs a named manager and a first project before they start. If either is missing, that’s something to flag before Day 1, not something to discover in Week 2.

Then the automation becomes useful. It helps the right people prepare, prompts the conversations that need to happen, and makes gaps visible early.

The important change wasn’t that we automated onboarding. It was that we made the handoff clear first. That’s the distinction I care about: technology should support a sound process, not protect it from being questioned.

Five questions I like to ask first

1. Who is this step actually for?

Sometimes work exists for the person completing it. Sometimes it exists to protect someone else from being surprised. Either can be valid, but it helps to know which.

2. What decision is this step helping us make?

If nothing changes based on the answer, it might not need to be a step at all.

3. Where does the information already exist?

People teams are very good at asking employees and managers to re-enter information we’ve already collected. Before building a new input, look for the existing one.

4. What happens if this goes wrong?

A badly formatted internal summary is an inconvenience. A wrong answer about someone’s leave entitlement or a candidate decision is a different kind of problem. The level of automation should reflect that difference.

5. Who owns the outcome?

If the answer is “the agent,” the workflow isn’t finished. A real person still needs to know when to review, intervene, or correct something.

Better isn’t always more automated

I’m enthusiastic about AI. I use it to help me build things that I wouldn’t have had the time or technical skills to build the same way a few years ago.

But the best outcome isn’t always a more sophisticated workflow.

Sometimes it’s one fewer approval.

Sometimes it’s a clearer owner.

Sometimes it’s a decent template that everyone can find.

And sometimes it’s simply making sure two people have a conversation before a problem grows legs.

That’s why I’d rather start with the work than the tool.

Once the process makes sense, automate away.