What's worth keeping in your work as AI takes on more of it

"Will AI take my job?" is the question I hear most from people who've spent twenty years getting good at something. It's a fair one. But after thirty years leading technology change, and testing AI tools every week, I think it's the wrong place to start. I've written about a better question to ask.

This piece is the practical follow-up. It's a way to work out which parts of your work are worth keeping, and what to do about the rest. It's also one of the exercises I ask Horizon Plan clients to do, so treat it as a look inside the program.

Your work is a bundle

What you do for clients isn't one thing. Some of it is production: work with a clear input and a clear output. Some of it is work you now do faster with AI beside you. And some of it only works because it's you doing it.

"Will AI take my job?" treats the bundle as one piece, which is why the answers tend to be either vague or alarming.

The useful question is which parts AI is coming for, and which parts get more valuable when it does. You can answer that one yourself, by sitting down with your own week.

Three kinds of work

Take one ordinary week and sort what you did three ways.

Production. AI does it now, or soon will. Drafting, summarising, reconciling, formatting, first-pass analysis. If a large share of your income rests here, that is worth knowing early rather than late.

Assisted. AI and you are better together. You bring the brief, the context and the standard. It brings speed. You may already be here for part of your week without having decided it deliberately.

Judgement. It needs you. Someone has to be trusted, someone has to be accountable, and someone has to know the context well enough to see what the numbers leave out.

An example from my own work. A client asked me for an AI agent to handle their customer care. It was a very complex job, and the agent was meant to integrate with every platform they ran.

I knew we had to start with the simplest version of it. Staff needed to trust the agent, and so did customers. So in the beginning it handed off to a person rather than push on alone, and it took on more of the task slowly, as reliability and trust built up. The client hadn't asked for that approach. It came from my read of what their staff and customers would accept.

The share of your week that sits in that third group is one of the most useful numbers to know about your work. The exercise at the end of this piece is how you measure it.

What to do with each part

Sorting is the start. Next, ask two questions of each block of work:

1. How soon will AI reach it?
2. How hard are you to replace in it?

Two questions with two answers each give four boxes.

Build on. AI is slow to reach it, and you're hard to replace. Put more of your time here, and more of what you offer clients.

Reinvent. AI is reaching it soon, but you're hard to replace because you know the work so well. You understand the old way well enough to design the new one, for your own firm or for your clients.

Watch. AI is slow to reach it, but other people could do it as well as you. It's safe for now, and it won't set you apart. Keep doing it while it pays, check on it when you review the plan, and don't build your next five years on it.

Release. AI is reaching it soon, and you're easy to replace. Hand it to AI deliberately, and plan for whatever it supported: the income, the service line or the type of client.

You'll probably find something in all four. If you look only at the last box, the picture is darker than it is.

What I think is worth keeping

This is what tends to land in "build on" when I do this for myself. You may disagree with parts of it.

Depth. Knowing a field well enough to tell a good answer from a plausible one. AI produces plausible answers all day. Someone still has to know the difference.

Relationships and trust. When a problem is hard, clients take it to someone they trust to care about the result.

Accountability. A tool can recommend. It can't stand in front of a client, a board or a team and own what happens next.

Taste. Knowing what good looks like, and what to leave out.

Hard-won experience. Having been wrong before, in ways that taught you something.

One lesson of mine: neither people nor AI are particularly good at judging the quality, scalability or security of what they've built themselves. I learned the hard way that taking a developer's word, or an AI's word, that something is secure is risky and often wrong when they're the one who built it. My instinct now is to have the important things checked independently.

In my view this is what clients have been paying experienced professionals for all along. The production work was bundled in with it, and the bundle is now coming apart.

Why it's hard to see from the inside

There is a catch. The things that make you hard to replace are often the things you find easy, so you discount them.

Ask a good accountant what clients value and they'll often say accuracy. Ask the clients and you may hear something else: that she tells them the truth, or that she noticed the problem before they did.

So put the question to three to five clients or colleagues who know your work: what do you rely on me for? The answers may not match the services you list.

For me, it's that I tell people what I think, whether they'd like to hear it or not. People ask me for an opinion because they know they'll get my honest one, not the one I think they want to hear.

The plan has to move

I don't know how fast this will go. A sorting exercise done today will be partly wrong in a year, because work will shift from one box to another.

So treat the result as a living plan. Name two or three AI developments that would change your answer. If one of them happens, look again straight away. Otherwise, look again every few months.

How to do it: five minutes a day for a week

You don't need a program to begin. You don't need software, and you don't need to time anything. Five minutes at the end of each working day, for one week, is enough.

Don't try to split the whole day in your head. Start with the tasks. List what you did, one line per task, with a rough time beside each. Your calendar and sent emails will jog your memory.

Then go back down the list and mark each task as production, assisted or judgement. Sorting one task at a time is far easier than judging a whole day. Estimate, don't measure.

Then write one line: a judgement call you made that day.

At the end of the week, add up the time in each group. The share that sits in judgement is your judgement share: how much of your week is work only you can do. Then read your five judgement calls together and ask what they have in common. That is likely to be close to what clients rely on you for.

One week is a snapshot. Do it again in a few months and watch which way the number moves.