Which work should your team hand to AI first?

Nov 7, 2026
Harry Siggins

Most of the leaders I talk to are under pressure to get their team using AI, and sooner or later they bring me the same practical question. Which work should we hand to AI first? And when I talk to the people on those teams before a workshop, a lot of them tell me some version of "I don't know what I don't know."

That's why I'd start with what they do know. Pick one piece of work your team does regularly and already does well, and use AI to make it faster. It sounds like a modest place to begin. But speed is usually what impresses people first, it gets a team going with AI quickly, and it sets up everything that comes after. To see why, it helps to know one thing about how AI behaves.

Why start with work you already know

AI is uneven, and you can't see in advance where it will fall short. A 2023 study led by researchers at Harvard Business School showed how this plays out. Boston Consulting Group consultants who used AI on tasks it handled well got more done, faster and to a higher standard. But on one task the researchers had picked because it was beyond what the AI could do well, the consultants using AI were less likely to reach the right answer than those working without it. And from the outside, the tasks looked about equally hard.

So you can't judge from the task itself whether the AI will handle it. That gives you two reasons to start with work your team already knows.

Your team can check it. The only way to find out is to check the result, and your team can only check well on work they already know.
It shows the AI your standard. The AI doesn't know what good means to your team until you show it. You show it by walking it through the work, correcting it as you go, and keeping those corrections as written instructions it reads every time it does the job. On work your team knows, those corrections come easily, because they can see straight away what's off. And the instructions will matter later, when you ask the AI to do more than speed things up.

How to choose the first piece

Knowing why to start with familiar work narrows the field, but it doesn't pick the piece for you. A good first candidate looks like this.

It happens regularly, every week or every month.
It's manual, or it takes several steps.
It ends in the same kind of result each time.
The people who do it can describe exactly what a good version looks like, so they can check what the AI gives them.

Those people are also the right ones to choose it, once they've seen what today's AI tools can do. Before that, they choose from their own assumptions about AI, and those are often months out of date. That's why, in my workshops, the first of three sessions covers what the tools can do now, and only the second turns to the team's own work.

To get ready for that second session, each person sends me one workflow they run, with answers to four questions.

What is it for?
What does a good result look like?
How is it done today?
Which tools does it touch?

You don't need a workshop to ask them. And once they're answered, you also have a first brief for the AI.

Then pick one, and only one. Going deep on a single piece of work is how a team finds out how far the AI can go. Spread the effort across five, and each gets done at a basic level, and basic results don't make anyone want to go further.

Here's what that first piece can look like. In a recent session with a finance and operations team, someone who handles collections brought their start-of-week routine. They pull a report from the accounting system, then check each invoice against the billing tool and their email to see whether it's still outstanding.

I suggested starting small. In the AI tool the team already used, they could set up an agent, an AI that works through a series of steps on its own, to run early on Monday, gather all of that into one sheet and take no action. The person would still act on each invoice themselves. But they'd start the week with everything in one place, and because it's a routine they know well, they'd spot it straight away if the agent got something wrong.

What comes after faster

Faster, though, is only where it starts. Any piece of AI work should earn its place in one of three ways, and the three tend to arrive in this order.

It makes the work faster.
It lets you do the work more often.
It makes the result better.

If it does none of these, it probably isn't worth keeping.

Faster has to be real, which is less obvious than it sounds. An agent that copies every step a person takes, downloading the same files and pasting them into the same spreadsheet tabs, can take as long as the person did and add a bill for AI usage on top.

When the time saving is real, it makes the next step possible, because work that takes less time can be done more often. Take the quarterly business review a customer success team prepares for each client. Once it takes a fraction of the time to put together, there's little reason to hold it only four times a year.

Better comes last, and it's the one people miss, because it starts with a different question. Instead of asking how to do the work faster, you ask what the work is for.

Would it be better with more data, or different data?
With more context?
Or does it need a different process altogether?

A people team I worked with found out what that question can do. They were setting up a dashboard to show managers who on their team had time off coming up. Then someone asked why managers should have to open a dashboard at all, when an agent could send each of them a Slack message with that information ahead of time. The team ultimately dropped the dashboard. Having AI build it would have saved the people team some effort, but the managers would still have had to go and look. The Slack message was better for them, because the information came to them directly and proactively.

That's why it helps to describe the outcome you want, rather than the thing you've always made. Ask the AI for a time-off dashboard and you'll get a time-off dashboard. Ask how managers could keep track of their team's time off, and you may get something better you hadn't thought about before.

Rethinking the work like this is harder at the start than speeding it up, and it takes more back-and-forth with the AI. Begin there, before your team has seen what the AI does well, and they can miss what it's good at and get frustrated when the first results fall short. But by the time a team has made a few pieces of work faster, they know what the AI is good at, and they've already written down their standard. They're ready for it.

What stays with your people

All of this hands work to AI, but none of it hands over the whole job. Your team was hired for more than the hard skills on their job descriptions. They were hired for how they think and the calls they make. And most of what goes into those calls, like knowing what matters, why things run the way they do and when something looks wrong, lives in people's heads. Very little of it is written down, and AI models learn from text, so what was never written down is out of their reach.

You can show an AI your standard for a piece of work, but deciding what that standard should be is judgment, and judgment stays with your people. So let AI do the gathering and analysis a decision rests on, and keep the decision itself with the people who own it.

For example, employee engagement surveys show where that line is. AI does a good job of pulling the responses together, finding patterns and spotting answers that don't fit. Deciding what to do about them is a different job. A general AI model doesn't know:

the history behind the results
your people strategy
your culture
how your people team does their best work

So the response should come from the team themselves, working with the people involved.

Where to start this week

Which brings it back to that first question, and a first step you can take this week.

Ask each person on your team for one piece of work they do regularly and know well.
Have them answer the four questions.
Choose one.
Use AI to make it faster, and check every result against what they already know good looks like.