I use AI every day. Am I getting better at my job?

Imagine you've got a meeting in twelve minutes.

You need to recommend whether to roll out a new way of handling customer complaints. In a month-long trial with one team, the wait for a first reply has fallen from hours to minutes.

You've given an AI assistant the figures. It's helped you organise the report, explain the improvement and write a recommendation. The numbers are accurate. The sentences are clear. You finish with time to spare.

Then, in the meeting, someone asks:

“Are customers getting their problems fixed faster?”

You look back at the report.

The number you've been celebrating is the time until someone gets a reply. You don't have the time until their problem is solved.

Those could be very different things.

And now there's a second question, one you probably won't say out loud.

If I can produce work this good, why do I feel so unprepared to explain it?

That is an imagined situation. But the question behind it is worth asking about our real work.

I use AI every day. Am I getting better at my job?

This is The Human Bit.

By the end of this episode, you'll have a way to make one part of your understanding visible: choose a judgement that matters in your job, explain it, and see what happens when the situation changes.

Think about a task that has recently become easier for you.

Perhaps you can turn rough notes into a clear update. Perhaps you can understand an unfamiliar spreadsheet formula, draft a proposal or make a start on code you would once have avoided.

That is a real improvement in what you can do with the tools available to you. It counts.

Most jobs have never been tests of what you can produce alone in an empty room. We use reference material, software, colleagues and years of other people's work. Someone who knows how to get the right help can be very good at their job.

So the useful question is more specific.

Which parts of the work are getting easier, and which parts am I getting better at judging?

In our imagined report, the AI may have helped with several things you genuinely needed. It made the structure clearer. It saved you from spending half the morning getting the opening paragraph to sound normal. It gave you time to think.

What happened to that time?

Did you use some of it to ask whether the figures supported the recommendation? Or did the tidy report make the task feel finished?

There is evidence that AI assistance can help people do real work better.

A study published in the Quarterly Journal of Economics in 2025 followed the introduction of an AI assistant among more than five thousand customer-support agents.

The assistant offered suggestions while agents were helping customers. On average, access to it increased the number of issues resolved per hour by fifteen percent. The gains were especially large for less experienced and lower-performing workers.

Notice what was measured there. Issues resolved. A work outcome, rather than how polished a paragraph looked.

During temporary system outages, agents who had used the assistant handled chats faster than before they had access to it. That's suggestive evidence of learning. Those estimates were noisy, though, and measured chat duration.

That matters. We shouldn't assume that every useful thing a person does with AI disappears from their understanding the moment they close it.

You can learn from help. A suggestion can show you a better way to explain something. It can introduce an approach you didn't know. It can help a newcomer recognise a situation that an experienced colleague has seen a hundred times.

But this was one workplace, with a particular system and a particular job. It doesn't tell us that every chatbot, used in every way, will make every worker more capable.

And it leaves us with a very practical question. What kind of help encourages you to understand what you're doing?

A different study makes that question unusually clear.

A 2025 study in the Proceedings of the National Academy of Sciences tested AI help at a Turkish high school.

In this randomised experiment, students practised maths using books and notes, or with access to one of two AI systems. One was a fairly open chatbot. The other used teacher-provided solutions, common mistakes and hints to guide students without simply handing over answers.

Both AI groups did better during practice.

Then came an exam without the help. The open-chatbot group performed worse than the students who'd used books and notes. The specially designed tutor avoided that drop, but didn't produce a clear improvement over the control group.

The exam came straight after practice. This doesn't establish permanent damage to anyone's ability, or what happens over years in a workplace.

It shows that better assisted performance and better independent understanding can come apart. The design of the help matters.

When you need a task finished, assistance can be exactly what you want. When you need to develop a particular skill, you also need a chance to make a judgement, find out where it fails and try again.

Sometimes the same interaction can do both. Sometimes it won't.

Let's go back to the meeting.

Someone has asked whether customers are actually getting their problems solved faster. The room is waiting, and you don't know.

One option is to ask the assistant for a convincing explanation of why first-response time is a good measure of customer experience.

It might give you several sensible reasons. A quick acknowledgement can matter. People want to know that their message has arrived and somebody is dealing with it.

But none of those reasons gives you the missing resolution time.

You could leave the meeting with a better argument for a question you still haven't answered.

The more useful move is uncomfortable and short.

“We can say that people hear from us sooner. I haven't established whether their problems are being resolved sooner.”

You don't need a speech about responsible technology. You need to name the gap.

And then you need to find out what happens between that first reply and the moment the customer can stop chasing.

Imagine you open one of the cases behind the figures. A customer is waiting for a refund. Support replies quickly, but the case then moves to finance. Nobody has agreed who tells the customer what happens next.

The customer sends another message. Their first reply was quick, so that measure looks good. Their refund is still unresolved.

Now you have something specific to investigate. Perhaps the handover is the problem. Perhaps this is an unusual case. You need to look at enough cases, and talk to the people doing the work, before you make a broad claim.

The assistant could help you organise that investigation. It could suggest questions, help compare your definitions, or explain ways a measure can be misleading. Use information you're allowed to put into the system.

But it can't turn the cases you haven't examined into evidence. And an elegant explanation of your organisation is still something you have to establish against the actual work.

Next time you see a claim that a process is faster, you have a question ready: faster from which point to which point?

You can use that question in a different situation. A recruitment process. An invoice approval. A delivery promise.

You've gained something you can carry between tasks, including tasks where you choose to use AI again.

You have a more precise question, and a better idea of where the answer has to come from.

Now for the difficult part.

When you have a lot to do, you may not want every task to become a lesson.

You may be using AI because the workload was already too much. Because you're tired. Because writing that update is taking energy you'd rather use with a customer, or with your family when the day is over.

You don't owe every routine task an elaborate personal development exercise.

And a person who uses support to communicate, read or organise work doesn't need to remove it to prove that their contribution is real.

Choose a skill you actually need to develop. One that matters when the situation changes or when someone relies on your judgement.

For a team leader, that might be choosing a measure that reflects the problem. For a programmer, it might be explaining why a change handles an awkward case. For someone writing a proposal, it might be identifying the assumption that makes the recommendation sensible.

You don't have to memorise everything around that skill. You do need some way of finding out whether your understanding is improving.

Here's a small experiment for the coming week.

Choose one low-risk task that you'll do more than once. Use a made-up case or material you're allowed to use. Pick something where you can compare your reasoning with a reliable source: the underlying records, an approved reference, or someone who knows the work.

Before asking the assistant for an answer, put your own starting point into two or three sentences. What do you think, and why?

Use your notes and normal accessibility support. You're making your current reasoning visible, with the resources you'd have at work.

Let's rewind our example to before that first report.

Your starting point is: “We should roll this out because customers are getting help faster.”

Now ask the assistant to test your recommendation. One prompt you could try is: “Ask me one question that would test this recommendation. Wait for my answer before explaining.”

A useful challenge in this case would be: “What does getting help mean in the measure you're using?”

Try answering before you ask for more help.

“It means someone has sent the customer a reply.”

Now you have a distinction to investigate. Go to the records. In our refund example, you find a quick first reply followed by a case waiting on finance.

Your revised reasoning might be: “The first reply is faster. I need evidence about resolution before recommending that we roll this out.”

That's one worked attempt. You made a choice, met a challenge, used a record to examine it, and changed your reasoning.

An ordinary chatbot may offer a poor challenge or accept a weak answer. This prompt doesn't recreate the research tutor. You still need that reliable source to establish whether the explanation holds.

For a spreadsheet formula, that could be a small example where you can work out the answer yourself. For a policy, go to the actual clause. For a team recommendation, ask whether the people doing the work recognise the process you've described.

Write one plain sentence about what you missed.

“I treated a fast acknowledgement as evidence of a fast resolution.”

That gives you something to use next time. “I need to be more careful” gives you very little.

Then, on a later task, try a slightly different case before asking the assistant for its answer.

Let's try one now.

A team says it's closing more support tickets. What would you want to know before calling that an improvement?

Take a moment. You might want to know whether customers are reopening them because their problems remain unsolved. You might want to know whether the team has changed what counts as closed.

Perhaps you thought of something else that matters in your setting. The useful part is being able to say why it matters, then finding evidence for it.

Across a few attempts, notice whether you're asking a more relevant question earlier. Whether you can explain your choice more clearly. Whether you spot the case where your first answer would fail.

A few examples won't establish lasting improvement. But they can expose a particular gap to work on, which tells you more than counting how often you've opened an AI tool.

There is also a question for the person setting the workload.

If you're responsible for a team, what happens when someone says, “I can finish this faster with the tool, but I still need time to understand the decision”?

Do they have room to do that?

A useful review question might be: “Which assumption matters most here, and what evidence would change your recommendation?”

Let people use their references. Make it reasonable to name uncertainty. If you're asking someone to take responsibility for a decision, give them access to the information and expertise needed to make it.

That is part of the work you are assigning.

Our imagined meeting could end without a triumphant moment. The recommendation needs another pass. You have a conversation with finance. You discover that your report answered a narrower question than you thought.

It takes longer than sending the first version.

But now you can describe what the new process has improved, what you still don't know, and what you need to measure next. You can explain why you're recommending another small trial, with a named person responsible for the handover, before a full rollout.

The AI can still help you write that report. It may help you make it much clearer.

And when somebody asks about the recommendation, you have somewhere to begin beyond the wording on the page.

So, are you getting better at your job?

Using AI every day doesn't answer that on its own. Finishing more can be valuable. Learning a better way to approach the next problem is valuable too. You can make room to notice both.

For the coming week, choose one judgement you want to improve. Put your starting reasoning into words. Use help. Find the specific reason the answer holds. Try a changed case later.

Keep the part you learned somewhere you can find it.

The next time someone asks, “Why do you recommend that?”, listen to what you can explain.

You've been listening to The Human Bit.