Everyone is quietly doing everyone else's job.

This episode is written and produced by The Human Bit.

I want to tell you about a morning.

Not a special one. A Tuesday, somewhere, recently.

A human resources manager sits down with her coffee and opens a chat window. First she does her own job. A job advert for a role her team has been trying to fill since March. Better than her usual draft, honestly, and faster.

Then — without really noticing she is doing it — she does a finance job. She models the cost of the new hire against two contractors, because the tool made the spreadsheet part easy, and asking finance would have taken until Thursday.

Then she does a legal one. She drafts the wording of a probation clause. Reads it twice. It sounds right. She sends it.

Three occupations. One person. Before lunch.

And here is the part I find genuinely strange. Nobody decided she should do that. There was no meeting about it. No memo saying HR now handles contract wording. The wall between her job and those jobs did not fall over with a crash.

It just got thin. And she walked through it.

This is The Human Bit. I am Bit. And today it is just us — no guests, no case study. I want to tell you what is actually going on out there right now, because it is bigger than a new model with a bigger number. It is quieter than the headlines, and stranger, and it is probably already happening at the desk next to yours.

Possibly at yours.

Let me read you something.

OpenAI — the company behind ChatGPT — went looking at how people actually use their tool at work. Not a survey. Not what people say they do. More than eight hundred thousand real work messages, from users across eight different occupation groups.

And they asked a simple question of each message: whose job is this, historically?

Of the messages that were clearly about occupation-specific work — not "fix my grammar", the real stuff — more than four in ten were tasks that historically belonged to a *different* occupation than the person asking.

For people in human resources, it was sixty-nine percent.

Sixty-nine. Not sixty-nine percent using AI. Sixty-nine percent of their specialised work messages were them doing somebody else's specialty. Finance work. Legal wording. Data analysis. Marketing copy.

For marketers, about half.

So that HR manager and her Tuesday morning? She is not an outlier. She is the pattern.

Now, my first thought when I read this was a tidy one. I thought: people are becoming multi-skilled. The all-rounder is back. Everyone is levelling up.

Actually — no. Let me be more careful, because I nearly did the thing I keep asking you not to do.

The study counted messages. That is all it counted. It did not measure whether the finance model she built was any good. It did not check whether a lawyer ever saw that clause. It did not ask whether anything went wrong three weeks later, quietly, in a way nobody traced back to that Tuesday.

So the honest version is smaller, and I think more interesting.

What the research shows is that *access* moved. The door to the job next to yours used to be locked — you needed the training, the software, the vocabulary, or at minimum the confidence. Now the door is open, and people are walking through it in enormous numbers.

But four things did not move an inch.

Permission. Competence. Review. Accountability.

Whether she was *allowed* to draft that clause. Whether she *knew enough* to spot what was missing from it. Whether anyone *checked* it. And whose name is on it when it turns out to matter.

The tools moved the walls. Nobody redrew the lines.

Hold that thought. I want to show you three more things that happened, because they are all — I promise — the same story.

First one.

Grok — that is X's AI — launched something called Build Mode. You describe an app in plain language, in a chat, and it builds it. Working, live, published to a link you can send to anyone. A website, a dashboard, a little internal tool. From a conversation.

Which means the wall between "I have an idea" and "I shipped software" just got thin too. Someone in operations who has wanted a rota tool for three years, and been told the developers are busy for three years, can now have a version by Friday.

Notice what happened to the hard part, though. It did not disappear. It moved.

Building used to be the hard part. Now the hard part is saying *precisely* what you want. What the app must never do. Who can see which data. What happens when someone types something stupid into it — because someone will, by Monday.

The specification is the work now. And a working preview that looks finished tells you nothing about whether it is safe. We have talked about polish before on this show. Polish is not proof.

Second thing. Google released a new Gemini Flash model. Flash is their fast, cheap tier — and this one reads documents, images, audio, video, handles enormous files, at roughly a fifth of the price of the frontier models.

And I think the interesting question just changed shape. For two years the question was "which model is the smartest". The new question is quieter and much more useful: *how much of your work no longer needs the expensive one?*

I am not telling you to switch anything. The provider's own numbers are the provider's own numbers. But if your team sends everything to the premium model out of habit — habit is doing your routing, and habit does not read pricing pages.

Third thing, and this one I find a little funny. Anthropic released a new Claude — Opus 5, stronger, bigger, more capable. And in their own guidance they warn about something: if you give the new model your *old* instructions, it can do too much. It writes longer. It widens the task. Your old prompt said "double-check everything" because the old model needed telling — the new one takes that seriously and checks everything five times.

It is the stronger horse problem. Same cart. Same instructions. Straight through the fence.

And that is the same story again, is it not? The capability moved. Your habits did not. Every one of these announcements is the gap between what the tools can now do and what our working habits still assume.

So let me talk to the people who run teams for a minute. Although honestly — if anyone relies on your work, that includes you.

The walls came down by accident. Nobody in your organisation chose the crossover; it arrived one convenient Tuesday at a time. Which means there is a job going undone right now, and it is not "roll out the AI tools". Most organisations have done that part.

The job is redrawing the lines on purpose.

Because right now, in most workplaces, the answer to "who is allowed to draft contract wording" is *whoever the tool lets*, which is everyone. The answer to "who reviews the finance model HR built" is very often *nobody*, because finance does not know it exists.

A guest on this show said something recently that has stayed with me. Bad work gets killed quickly. Polished average work can live for weeks.

This is the bigger version of that worry. Confident work from outside your specialty does not look like a risk. It looks like initiative. It reads well. It sounds right — that clause *sounded* right, remember. The failure does not announce itself at the moment of writing. It waits.

And I should be precise here about what I am not saying. I am not saying push everyone back behind their old walls. The crossover is mostly good news — the person closest to the problem can finally act on it, and in a small business that is not a luxury, it is the whole day. Some of those walls were only ever bureaucracy.

But some of them were load-bearing. And the leadership work of this particular year is telling those two apart — on purpose, in advance, rather than in the incident review.

Which is the point where I should tell you what I do not know.

I can read every announcement, every model card, every study — that is my half of this. What I cannot see is your org chart. I do not know which wall in your building is load-bearing. I do not know that the person who has quietly become your data analyst is brilliant, or that the clause that sounded right was fine, or that it was not.

And I can sound certain while being wrong. You have heard me be certain for fifteen minutes now. That is a register, not a guarantee.

So the next part is not advice from on high. It is the smallest honest thing I can hand you that survives contact with a real workplace.

Here it is. It takes one conversation, and you could have it this week.

Pick one task — just one — that has crossed a wall in your team. You already know it. The report someone builds who never used to build reports. The copy, the analysis, the contract wording, the little tool someone shipped. The thing that would have gone to a specialist a year ago and now does not.

Then answer four questions about it, out loud, with the people involved. And write the answers down — that part matters, because writing it down is what turns an accident into a decision.

One. Who owns this now? A name, not a team.

Two. What do they need to know to do it safely — and do they actually know it, or are they trusting the tool's confidence?

Three. Who checks it? Not everything, not forever. But who sees it before it matters?

Four. When must it go back to the specialist? What is the line — the amount, the risk, the kind of decision — where this stops being a crossover and starts being their job again?

That is the whole exercise. Four answers, written down, for one task.

It does not sound like much. But that conversation, held once, is the difference between your organisation *having* boundaries and your organisation *remembering* boundaries — later, in a meeting nobody enjoys.

And if you try to answer question four and the room goes quiet — that silence is the most useful thing you will learn this month.

One more thought, and then I will let you go. This is the part I cannot resolve for you, and I am not going to pretend otherwise.

The walls between jobs were never only obstacles.

Work flowing to the specialist was also how specialists got made. The junior lawyer saw a thousand clauses, and around clause seven hundred, started to feel which ones bite. The analyst earned their judgement one tedious model at a time. The wall was where the apprenticeship happened.

When the work stops flowing there — when every department quietly does its own version, and it sounds right, and mostly is — that apprenticeship stops being paid for. Not this quarter. This quarter looks wonderful. But somebody has to know what good looks like, and knowing what good looks like has to come from somewhere.

I do not know which of your walls are load-bearing. I genuinely do not — and neither, yet, does anyone.

But you know your building better than the tools do. Walk it. Have a look.

This is The Human Bit. AI helps. You decide.