Quite a while back - during my early initial explorations with AI - I dropped a client email into an AI thread for a quick second opinion. The email had two things in it: a question about a system we'd deployed for them, and a separate question about an advertising campaign.

The reply came back sensibly enough on the first part. Then, on the second:

You may want to forward this portion to whoever handles business development at your organisation.

That's me. I'm the person handling business development. I am, in fact, titled Head of Business Development, and I had just been advised - politely, helpfully, with what I can only describe as a light air of delegation - to escalate my own email to myself.

I sat there for a moment enjoying it. Then I went and found the earlier thread where it had told me to send an agreement to my legal review team, which is also, in the sense that matters here, me.

It wasn't wrong, exactly

This is the part that took me a while to appreciate. Neither of those answers was a hallucination. They were both entirely correct, given what the model had been told.

Somewhere in its picture of me was a person with a job. Jobs have edges. If someone is running digital products, then legal review happens somewhere else, in a room with lawyers in it, and advertising sits with a commercial team, and the sensible advice is to pass things along to the relevant function. That's not a stupid inference. It's the shape almost every organisation has, and it's the shape of nearly everything ever written about how organisations work.

It just isn't the shape of this one. Or, I'd wager, of quite a lot of others.

The model had built a normal company in its head and put me in it. And an awful lot of business advice, tooling, templates, best practice and confident LinkedIn wisdom does exactly the same thing - assumes a function behind every task, a department behind every function, and somebody else to hand the awkward bit to.

I've spent twelve years being the somebody else.

The other failure mode, which is worse and funnier

Once I understood what was happening, I did the obvious thing and told it. Here's my actual scope. Here's what sits with me. Here's what my week involves.

And the pendulum swung directly into the other ditch.

Because if you tell an AI that you handle commercial, digital, legal first-pass, ad operations, product architecture, data privacy and SOPs, it will believe you. Utterly. Without a flicker of doubt. And then it will start planning your Tuesday accordingly.

Sure, you can restructure the pricing model this week - and while you're at it, why not draft the governance framework, rebuild the onboarding flow, and prepare the board narrative?

It never says that's four people's work. It never says you have three meetings and a school run. You've told it you're a superhero, so it plans for a superhero, and hands you back a schedule that a superhero would find ambitious.

The first version thought I couldn't do anything. The second thought I could do everything. Both were reflections of what I'd told it, and neither had any independent way of knowing better.

The unglamorous bit at the end

So the whole thing comes back around to context. What it knows is what you told it. Not what's true - what you said. The gap between those two things is where all the strange advice lives, in both directions.

Which means the work is describing yourself accurately, including the unflattering parts. Not the org chart version. The real one, with the odd overlaps and the things you own that nobody would guess you own and the things you're genuinely not good at and the fact that there is a limit to how many hours exist.

And then - because I've written about this before and I'm afraid it's still true - write it down and keep it somewhere safe. Not in the thread. Somewhere that survives the thread.

Otherwise you'll be doing this again in three months, from scratch, with the same patient tone.

I'd suggest raising this with your line manager.

And you'll laugh as you imagine writing the email addressed to yourself.