Think

My father was a Turkish Airlines captain. And his machine was an Airbus A310.

He was immensely proud of it. He had moved to the A310 from the Boeing 727, and to him, it felt like the future. His favourite thing to tell me was that the aircraft COULD LAND ITSELF.

But there was another thing he was just as proud of. If anything went wrong, he could take over.

Because while aviation kept automating, it never stopped training pilots for the moments when automation wasn't enough.

Simulator sessions. Recurrent training. Proficiency checks. Practising situations they might rarely, or hopefully never, encounter in real life.

Aviation even has an acronym for one of the problems behind this: CSD. Cognitive Skill Degradation.

When automation keeps doing something you still need to know how to do, eventually you may not be quite as good at doing it yourself.

But AI may give our industry an even stranger version of the problem.

What if people working in advertising production never acquire some of those skills in the first place?

Think about all the junior work we're quite reasonably trying to automate. Versions, adaptations, artwork changes, file checks, basic edits, small production problems. The repetitive stuff.

Good. Let's automate it.

But those thousands of small tasks were also thousands of small decisions to learn from.

And for an art director, producer, editor, developer or VFX artist, a lot of learning never came from a training department. It came from doing the work. From watching someone better do it. From getting something wrong, fixing it, and seeing what happened next.

Until now, the work trained the worker.

Today's seniors already have that experience. We can put them above an AI-enabled workflow and ask them to supervise, judge and step in when something goes wrong.

But where does the next generation get theirs?

I don't think the answer is to preserve inefficient work just so people can learn from it. Aviation didn't stop automating either.

It deliberately created ways to maintain the experience and capability that everyday work was no longer providing on its own.

We may have to do something similar.

If AI takes away some of the work, somebody will have to start designing the experience that work used to provide for free.

People, Operations and Craft Leadership... I think you three need to have a chat.

We know we can’t just leave AI to its own devices. And whenever that conversation comes up, we seem to end up in the same place:

“We need human judgment here.”

Fair enough. But actually, you’re buying time.

Because sooner or later, you’re going to have to explain what you mean by “judgment.”

We use the word as if we all mean the same thing. But put a human into an actual workflow and you have to give them something a little more useful than that.

So far, these are the ones I’ve managed to catch:

“Judge”: Is this good enough?

“Predict”: What’s likely to happen next?

“Value”: Which outcomes matter more, and how much?

“Choose”: Pick one of the options.

“Decide”: What are we actually going to do?

And I’m pretty sure that’s not the whole list.

“Human judgment” is fine. Until the moment comes when you have to design the workflow.

Then it isn’t really a specific enough term anymore.

And if judgment is going to be one of the things we humans hold on to, we should probably get a little better at knowing exactly what we mean by it.

So next time someone says, “human judgment is gonna happen now”…

ask them:

“Well, which judgment exactly do you mean?”

[I’ll leave a few good reads on judgment and decision-making in the comments if you want to dig a little deeper.]

As more of the execution starts moving to agents, some production roles may start sounding less like “maker” roles and more like “manager of makers” roles.

So here’s my completely unofficial preview of the production department of the near future:

Artworker → Artwork Agent Manager

Retoucher → Synthetic Image Supervisor

Motion Designer → Motion Agent Director

VFX Artist → Generative VFX Supervisor

Editor → Autonomous Edit Supervisor

Audio Producer → Synthetic Audio Producer

Broadcast Producer → Synthetic Broadcast Producer

Digital Producer → Agentic Experience Producer

Integrated Producer → Human-Agent Producer

Executive Producer → Executive Agent Producer

Localization Manager → Language Agent Wrangler

Transcreator → Cultural Intelligence Supervisor

Versioning Producer → Variant Orchestration Lead

QC Specialist → Machine Output Supervisor

Traffic Manager → Agent Traffic Controller

Resource Manager → Human-Agent Resource Planner

Project Manager → Agent Workflow Manager

Senior Project Manager → Multi-Agent Delivery Lead

Account Manager → Client-Agent Interface Manager

Account Director → Human-Agent Business Director

Production Manager → Human-Agent Operations Lead

Studio Manager → Synthetic Workforce Manager

Creative Technologist → Agent Systems Architect

DAM Manager → Asset Intelligence Custodian

Rights Manager → Synthetic Rights & Provenance Officer

And then, inevitably:

Head of Handoffs

Chief Exception Officer

Synthetic Talent Producer

Variant Collision Manager

Agent HR Manager

Prompt Janitor

I’m joking about the titles.

Not entirely about the jobs.

Send this to your friend who’ll have one of these job titles in 2028.

Or add the one I missed. 👇

And if you’re curious about the more serious thinking behind the joke, I’ve left a recent California Management Review piece in the comments.

Or at least, that's where we're heading.

It used to be fairly straightforward. You looked at what the manager did. You looked at how the team was doing. And some of that team performance reflected on the manager too.

Makes sense.

But we're starting to hand over actual work to AI agents. We check what they do, decide when to step in, and we're still the ones responsible for the final result.

Of course we are.

We're getting pretty good at measuring how well those agents perform.

But what about the person managing them?

If managing AI agents becomes part of my job, at what point does the agent's performance become part of mine?

So maybe that's where we need to look.

Did the agent do a good job?

And did I do a good job managing it?

Did I give it the right work?

Did I know when to trust it, when to challenge it, or when to step in?

I don't have a good answer yet. But I suspect we'll have to find one soon.

The Washington Post asked if the em dash had become a sign of AI writing, the Guardian wrote about AI stealing it from writers, AP Stylebook looked into where this whole thing came from, researchers started studying it, and Pew counted em dashes across nearly half a million web pages.

AI stole the em dash. And got busted.

Here's why: there is no em dash key on a standard keyboard. To type the one above, I had to press Option + Shift + *. Three keys, every single time.

A normal person doesn't bother. AI didn't care.

Come on, AI. You could have been smarter. Pick the semicolon instead, it's right there on your keyboard, and nobody would've noticed a thing.

[Links in the first comment]

I’m reading Power and Prediction by Ajay Agrawal, Joshua Gans and Avi Goldfarb at the moment. They describe three types of AI solutions: point, application and system.

Let me try three examples from ad production.

1. Point solution

You were already producing different sizes of an existing artwork. Maybe the resizing was manual, maybe there was already a script. Now AI is used for that same step.

2. Application solution

You add a capability you didn’t really have before, without changing the production system around it.

Say you’re working from a multilingual copy matrix. AI keeps an eye on it as the work progresses, scoring which adaptations look straightforward and which are likely to cause trouble, and updating the scores with every revision. Useful for planning and cost forecasting.

I want to call it “dynamic predictive production complexity scoring”. If we have room for one more technical term on a Friday.

3. System solution

The client may still be asking for the same thing. Your agency may get there in a completely different way.

Different steps. Different roles.

AI has changed what’s possible, so you design the entire system differently. Sometimes it’s about what system you can now build that you couldn’t afford to build before.

There. Expert.

And while we're here, which type of AI solution are you seeing around you? Point, application or system?

She was the proofreader at the advertising agency where I'd just started as a junior copywriter.

One of my first lessons from her was about punctuation. I'd put a full stop inside the quotation marks when it belonged outside.

Her job was to catch things like that before the work went any further.

But there was another part of the process I remember even more vividly.

Before work went out, a large sign-off stamp was pressed onto the physical proof. We called it the stampa. It left what looked like a little form, with different roles listed on it and a box for each person to sign. The proof travelled around the agency. You checked what you were responsible for and signed your box.

Copywriters, art directors, proofreaders, account teams and creative directors all had their own part to look after.

By the end, you could see who had checked what.

Which is probably why there's something very familiar to me about the idea of “human in the loop”. (Got the full stop right this time.)

Then I read an article by Selena Cameron about what happens when AI production volume begins to outscale human review.

It made me think about that old stampa differently.

A signature used to mean two things at once:

I've seen this. And I'm taking responsibility for my part of it.

At AI scale, those two things may no longer always travel together.

There may be too much work for any one person to see. Automated checks, sampling, exception handling and risk-based escalation are already becoming part of how that problem is managed.

But that creates another difference.

Back then, your professional role largely told you what your role was in the approval process. The relationship was obvious.

In an AI-driven workflow, it may not be.

At one moment you might be reviewing an output. At another, making a judgement on an exception. Somewhere else, monitoring how a system is performing rather than checking the individual work at all.

Which makes me think “human in the loop” may be becoming too broad a description.

The human is there, yes. But what are they there to do each time, exactly? Reviewing the work? Making a judgement on an exception? Monitoring the system? And what does their approval mean in each case?

If we’re going to put a human in the loop, we should make their role explicit each time they enter it, and make sure they understand the responsibility they’re taking on when they act.

In a strange way, the old stampa already did that.

It didn't just ask for your signature. It made pretty clear what on earth you were doing there.

Selena Cameron’s piece that got me thinking about this:

https://lnkd.in/p/dDkCN9Kq

What got me thinking about this was an article I came across today about job candidates learning how enthusiastic they're supposed to sound about AI in interviews.

Which made me smile.

But the article was actually more interesting than the headline.

Candidates are preparing stories about how they've used AI to solve problems or make their work more efficient. Some are apparently trying to work out how enthusiastic they should sound. Others are playing up their enthusiasm even when they have reservations.

Basically, they're learning to read the room.

Which is understandable. If employers are asking different questions, candidates will learn how to answer them.

But after reading it, I started wondering about the other side of the table.

What should we actually be looking for when we hire people now?

I got curious, so I started looking at current job descriptions across WPP, Publicis and Dentsu. Then I found a few older ones too to compare.

At company level, the shift is hard to miss. AI-powered. AI-enabled. AI-driven. The language has already changed.

Which, at that level, is perhaps the easier part.

Scroll down to the actual role, though, and things get more interesting.

There's quite a range.

In some, AI is another skill to have.

In others, it's becoming part of how the work gets done.

And in a few, people are being asked to rethink the work itself because of AI.

So I looked at the hiring side too.

Companies are clearly thinking about it. Work samples, simulations, AI assessments, new talent programmes. The hiring process isn't standing still either.

But I came away thinking that today's AI proficiency may not be the most interesting signal.

It has a pretty short shelf life.

What may matter more is what someone does when the way they've learned to do their job no longer makes quite as much sense.

Not just whether they can learn the next tool.

Whether they can rethink the work.

Which leaves me with a much harder question.

How do you spot that across an interview table?

[The article that started all this]

https://lnkd.in/dqZM2NjA

I've been thinking about this quite a lot lately.

Maybe because I've spent most of my career in global agency networks. Which also means I've spent most of it working with systems I didn't design.

Some worked better than others. But they were built for the technology, capacity and coordination realities of their time.

And that's the bit I keep getting stuck on.

I came across an interesting piece this week arguing that AI's next victim could be the traditional agency retainer.

The retainer question is interesting. But if we're questioning who does the work, shouldn't we also be questioning what work needs to be done, and how?

That's where my mind went. The production workflow itself.

A lot of the conversation around AI in production still starts with what we already have. Which tasks can we automate? Which handoffs can we remove? How much faster or cheaper can we make it?

Of course those questions matter.

But if AI changes some of the constraints the workflow was built around in the first place, are we always starting in the right place?

So, two questions:

How much of the current production workflow can AI automate?

And if we were designing the production system from scratch today, how much of the current workflow would still be there?

I'm increasingly interested in the distance between those two questions.

https://lnkd.in/disKiJmd