You’ve Been Promoting Your AI Too Fast: A Guide to Managing the Office Junior

A few years back, an apprentice was handed a simple job: set up a mailbox to collect audience questions for a Q&A, following a presentation by one of the board directors. Let’s call the director John Smith. Sensible lad, the apprentice — he checked before acting. He went to the director’s PA and asked what the mailbox should be called.

“Ask John Smith,” she said.

So he did. He opened a fresh email to John Smith and asked what he’d like his mailbox called. What came back was not a mailbox name. It was, by all accounts, a right bollocking.

You can see both readings. The PA meant name the inbox “Ask John Smith” — the obvious name for a Q&A inbox. The apprentice heard go and ask John Smith. He did exactly what the words said and precisely not what anyone meant, and a board director’s morning was the poorer for it.

I think about that apprentice a lot now, because most organisations are busy hiring several thousand of him, handing him the login credentials, and calling it an AI strategy.

The mistake isn’t buying the tools. The tools are genuinely good. The mistake is misjudging which member of staff you’ve just hired — treating a keen work-experience kid as if he were a chartered accountant, or worse, treating a chartered accountant as if he were a stapler. Automation has been joining your team for decades. It’s just been getting promoted, and nobody updated the org chart.

Meet the workforce you didn’t know you’d hired

Here’s the reframe. Stop thinking of your automation as software and start thinking of it as staff. Every tool you’ve bought sits on the same career ladder a human climbs, from the kid doing the filing to the professional signing the accounts. Managing it well — and the reason so many “AI pilots” quietly die — comes down to knowing which rung you’re standing on.

We’ll start further down the ladder than you’d expect, because the junior end is where the useful lessons hide.

The spanner (who is not, in fact, a colleague)

At the very bottom sits the automation that isn’t a person at all. The spreadsheet formula. The thermostat. The overnight mail-merge. This is a spanner: one motion, performed identically, forever, with no opinion about any of it.

Don’t be rude about the spanner. It’s reliable precisely because it can’t improvise. Nobody’s login credentials were phished by a SUM function. The failure mode here is boredom, not judgement — you’ll under-use it long before it lets you down.

The Saturday-job kid (brilliant on-script, hopeless off it)

Promote one level and you get scripted automation: the macro, the batch job, classic robotic process automation. This is the Saturday-job kid. Give him a precise, repetitive task and he’ll do it faultlessly for eight hours. Move the shelves, though, and he freezes — because his instructions said aisle four and aisle four is now aisle five, and nobody told him what to do about that.

This is most RPA in one sentence. It’s cheap, it’s fast, and it collapses the instant the underlying form or data schema changes shape. When someone tells you their automation “keeps breaking,” this is almost always the rung they’re on. The kid didn’t get worse. Someone moved his shelves.

The school-leaver apprentice (your first rung of actual learning)

Now it gets interesting. The apprentice — a school-leaver, keen, a bit green, the sort who’ll email John Smith — is where generative AI on a short lead lives. He can tidy your data, key it in, do the supervised prep and even spot the odd obvious error. He learns a little as he goes.

What he can’t do is define “good” for himself. Hand him a messy dataset and he’ll clean it beautifully to a standard you set; forget to set that standard and he’ll cheerfully clean it to one he’s invented. He can hold the tool. He can’t yet choose which tool the job needs. That’s your call, and it stays your call.

The university intern (impressive, and quietly overconfident)

Then there’s the intern, and this is the rung most people mistake for the top. This is capable generative AI, edging into agentic territory. He doesn’t just tidy the data — he interprets it, decides what matters, and drafts you the board pack, recommendation and all.

It’s genuinely good work. It’s also the rung where things go wrong in a way that’s much harder to spot, because the output looks finished. Give the apprentice a duff instruction and he’ll email John Smith: daft, visible, sorted before lunch. Give the intern a subtly wrong dataset and he’ll hand you an immaculate board pack built on top of it — clean charts, a confident narrative, a bold recommendation on the final slide, and one number four slides deep that came from somewhere he can’t quite place. The apprentice fails visibly and safely. The intern fails invisibly and persuasively. Guess which one costs you more.

What ITIL (Version 5) quietly worked out

Here’s where the framework people did everyone a favour. Inside ITIL (Version 5)'s Information and Technology dimension sits the AI Capability Model, better known as the 6C model: six functions an AI can perform — Creation, Curation, Clarification, Cognition, Communication and Coordination. Drafting, tidying, turning a fat incident ticket into plain English, spotting the pattern a human missed, fronting a chatbot, and, at the top, acting across your systems unprompted.

ITIL doesn’t sell those six as a seniority ladder; it’s a list of what the machine does, not how far to trust it. But read them in order and they climb the way our junior does. Creation and Curation are apprentice work. Clarification and Cognition are the intern interpreting. Coordination is the newly-qualified professional reaching for the chequebook. Which is why ITIL (Version 5) pairs the capability model with a companion AI Governance Model and a four-step assess, design, implement and maintain cycle: the promotion criteria and the disciplinary procedure in one publication. The higher up the C’s you climb, the more governance the role needs, not less.

The promotion you must never sign off

That top rung is the newly-qualified professional: agentic AI running a whole outcome end to end, correcting its own course as it goes. The pull to promote it is enormous, because on a good day it’s genuinely excellent, and every excellent day makes the next one feel safer.

Here’s my actual position, and it’s not a comfortable one for the vendors: you promote the capability as far as it’ll go, and you never, ever automate the accountability. The machine can climb every rung on this ladder, right up to Coordination. The name on the audit trail stays human. When the regulator or the board asks “who decided this?”, “the model did” has never once been an acceptable answer, and it isn’t about to become one.

Warning: you might have over-promoted your AI if…

You’re acting on its output without anyone checking the numbers, because the formatting looked authoritative

You bought a “Coordination”-grade tool and quietly gave it the permissions of a trusted senior manager on day one

Your automation “keeps breaking” every time a source system changes, and you’ve concluded the tool is faulty rather than junior

Nobody in the room can answer “who’s accountable if this is wrong?” without pointing at a screen

If two or more of those landed, you haven’t got an AI problem. You’ve got a line-management problem, and the good news is that management is a solved discipline.

Here’s the fix:

✓ Grade each tool honestly against the ladder — spanner, Saturday kid, apprentice, intern — before you trust its output ✓ Match the supervision to the rung: the intern’s work gets read before it leaves the building, every time ✓ Keep a human name against every decision the automation feeds, and make that person’s accountability real, not ceremonial ✓ Use the ITIL (Version 5) 6C model and its AI Governance Model to assess maturity properly, so promotions are earned rather than assumed

Manage your automation like the workforce it is, and it’ll do more for you than any of the brochures promise. Mistake the intern for the board director, and you’ll find out about it one board meeting too late.