A few years ago, I remember sitting in an office with someone who said they would never write with AI.
Now virtually everyone is writing with AI.
That shift alone tells you what sort of moment we’re in. What looked fringe, awkward, and vaguely embarrassing has become normal with frightening speed. And not just normal in tech. Normal in offices. Normal in homes. Normal in schools. Normal in search. Normal in daily life.
Even this piece is part of that change. I’m not writing it in the old-fashioned sense. I’m dumping voice notes into a workflow I’ve built, then using AI to turn raw thoughts into something structured, readable, and hopefully worth your time.
That makes perfect sense to me, because voice is a faster interface. I can get ideas out while they are still warm, and AI can help shape them into something clearer.
That’s just a better workflow.
And I think that’s the first thing people still haven’t fully absorbed: the future isn’t arriving. It’s already here, and it’s starting to speed up.
From novelty to infrastructure
A few years ago, AI output was often poor. But that was always going to be the worst it would ever be.
A lot of people looked at the rough early versions, laughed, and assumed the whole thing was overhyped. They mistook an early snapshot for the final product. Meanwhile, the systems kept improving. Quietly at first, then suddenly.
Now you can create detailed infographics, images, summaries, research breakdowns, code, tools, drafts, workflows, and analysis in seconds. Things that would once have required a specialist, a team, or at the very least a block of focused time can now happen absurdly quickly.
That changes more than productivity. It forces us to rethink what work is, what value is, and what we are actually paying humans for.
Because once the machine can do the first draft, the formatting, the admin, the synthesis, the comparison, the QA, the optimisation, and increasingly the implementation too, the question is no longer, “Can AI do useful work?”
The question becomes, “What is left for the human to do?”
AI is already crossing generations
This isn’t some niche behaviour restricted to people on X posting agent screenshots.
My father is in his 70s and happily using AI Overviews. That should tell you something if you know how long I’ve lived in the search world. My daughter attends an AI-powered school setup I built for her because she’s home educated. My son is revising for his GCSEs with an AI-powered homework assistant that loads questions and helps him learn faster. His grades are improving.
So yes, adoption stats might still look low in the abstract. Fine. But the shape of what’s happening is obvious. AI is crossing generations because it is useful. Once a tool is useful, it spreads.
And yet huge parts of society still seem to be living in a weird bubble.
My son still has to sit one maths paper without a calculator. That result then gets bundled in with the calculator papers, so weak mental arithmetic can drag down his broader performance, even though that is not remotely representative of the world he is entering. It’s a bonkers mismatch. We are preparing children for environments that are already disappearing.
There is probably a generation now that risks leaving school without the right skills for an AI-shaped world, while the adults in charge carry on pretending the old model still fits.
It doesn’t.
I’m not talking about AI anymore. I’m embedded in it.
Part of why I started AI Growth Lab was to share how I’m actually using these systems in real life, not just comment on them from the sidelines.
And the truth is, I’m now more embedded with AI than I’ve ever been.
I’m using Google AI Studio to build real-world marketing apps that help me day to day.
I’m using ChatGPT as a kind of central thinking environment: for structuring ideas, learning, stress-testing, and spotting what’s missing.
I’m using Gemini because it can do things the others can’t, especially when Google’s wider ecosystem becomes part of the workflow. NotebookLM is excellent for learning, source synthesis, and turning lots of material into something usable.
Claude is increasingly useful when I want to get work done, especially when the task is substantial, and the output needs to move.
And what has become obvious is this: none of these tools are perfect on their own.
They each have strengths. They each have weaknesses. And the real gains often come not from betting on one system, but from learning how to use several of them together.
In one workflow I’ve been using, Gemini gave me the first draft. Claude took over for the stronger writing passes. ChatGPT, with extended thinking switched on, became the critic and editor, pointing out what was weak, what was missing, and what needed to be pushed further.
That surprised me. It ended up feeling like an agency structure: Gemini as the junior, Claude as the senior copywriter, ChatGPT as the editor or strategist.
That is a useful way to think about this next phase.
Not one magic model.
A stack.
A workflow.
A team of systems with different roles.
The big unlock is agentic capacity
We are also starting to see where this goes next.
The combination of better models, better tooling, and more agentic workflows means you can now set up systems that move far beyond the old prompt-in, answer-out pattern. We are entering a phase where machines do things, not just say things.
That matters because it starts to change the economics of knowledge work.
A few years ago I wrote about the possibility of the first billion-dollar one-person SEO agency almost as a joke. Now I’m not joking quite so much. One person with the right stack, the right process, and the right literacy can suddenly become absurdly productive.
Client meeting gets recorded. Notes are taken automatically. Tasks are identified. Agents begin work. Drafts get produced. Checks get run. Comparisons get built. Recommendations get surfaced. Revisions happen fast. The work moves while the human is still thinking about the next decision.
That is not theory anymore. It’s starting to become practical.
But this is where most people stop their thinking too early. They see infinite scale and assume the human disappears.
I don’t think that’s right.
The human bottleneck
The biggest insight I’ve had lately is that AI can scale output close to infinity, but it cannot scale human judgment.
That is the bottleneck.
You can produce more content, more recommendations, more research, more pages, more campaigns, more analysis, more everything. But you cannot review, understand, supervise, and care about that work at the same speed. A human still has to read it. A human still has to think about it. A human still has to decide whether it is right, good, useful, on-brand, and strategically sound.
That takes cognitive bandwidth.
And bandwidth is finite.
I’ve felt this very directly in complex client work. We were producing complicated buying-guide-style material about products and services that were difficult, technical, and nuanced. AI was brilliant at synthesising information faster than I could. It could compare, draft, refine, and keep iterating. But by the ninth round of revisions, the real issue was no longer output.
The issue was direction.
Was the goal clear enough at the start?
Had the desired end state been defined explicitly enough?
Was the machine solving the right problem or just doing impressive-looking work around the edges? And had enough human insight been added at the right moments to stop the process from drifting?
That’s where the human in the loop becomes everything.
Because the machine is powerful enough to go very fast in the wrong direction if you don’t manage it properly.
Unlimited productivity, limited headspace
This is why I think we are heading into a strange economy: one with near-unlimited productivity, but deeply limited human headspace.
You can absolutely imagine an agency where one person has 40 clients and an army of agents doing the production. In theory, that person looks superhuman. In practice, they still only have one brain. They still only have so much strategic capacity. They still only have so much reading time, so much concentration, so much ability to care.
That last word matters more than most people realise.
Because in the end, the premium is not going to sit in “doing the work.” AI is rapidly making a lot of doing cheaper.
The premium shifts to caring about the work. Supervising it. Challenging it. Improving it. Taking responsibility for it. Holding the client, the context, the nuance, and the desired outcome in your head long enough to make sure the output is actually any good.
That is why I think we may be moving beyond the knowledge economy into something closer to a care economy.
Not “care” in some fluffy brand-values sense. Real care. Deep attention. Judgment. Taste. Oversight. Emotional labour. Cognitive labour. The willingness to stay with a problem until it is genuinely solved.
That becomes scarce when machines can produce endlessly.
When the basics get automated, value moves upstream
SEO is the easiest example for me, but this applies almost everywhere.
Imagine a client comes in wanting better search visibility. In the old world, optimisation was a months-long process involving audits, briefs, content changes, dev queues, review cycles, and endless chasing. In the new world, agents can identify issues, update templates, improve internal linking, generate missing sections, ship code, and check each other’s work in parallel.
What took months could take hours.
Great. Then what?
Well, if everyone can get to “well optimised,” optimisation stops being a differentiator. It becomes table stakes. So value moves upstream into the next thing machines can’t flatten as easily.
That might be PR. It might be positioning. It might be events. It might be partnerships, distinctive brand assets, reputation, trust, community, or attention. But the pattern is the same. Once the basics are automated, humans are pushed up the stack into the next scarce layer.
This is why I don’t think AI kills marketing. I think it makes real marketing more important.
If everyone can produce, then being chosen matters more. If everyone can execute, then preference matters more. If every business can use machines to get the basics done, then brand, positioning, and distinctiveness become even more valuable.
The machine does more of the doing.
The human still has to create preference.
Cheap AI versus cared-for AI
I also think there’s a commercial divide coming.
On one side, you’ll have cheap, mostly automated output with little human supervision. Good enough, fast, affordable, lightly checked. On the other side, you’ll have AI-accelerated work with real humans in the loop: reading it, challenging it, improving it, managing the relationship, and taking responsibility for the result.
Those are not the same offer.
One is scale.
The other is trust.
Some buyers won’t care. Some will happily take the cheaper version and accept the risks. But wherever quality, nuance, accountability, or strategic consequences matter, human-supervised AI will command the premium.
Not because humans are doing more typing.
Because they are doing more thinking.
The super-user tip that changed my results
One of the best practical techniques I’ve picked up recently is deceptively simple.
Don’t ask the AI to do the task first.
Ask it to build the best possible prompt for completing the task.
So instead of saying, “Write this article,” or “Build this workflow,” or “Analyse this problem,” I explain the outcome I want in plain English. I explain the context, the constraints, what success looks like, what the final output should do, and what I’m trying to achieve. Then I ask the model to write a detailed prompt that would force the AI system to complete that task properly.
I first came across a version of this through Reid Hoffman talking about using AI more strategically, and once I started doing it myself, the jump in quality was obvious. Add Extended Thinking in ChatGPT on top and it unlocks something close to god mode for depth. Not perfection. Not magic. But materially better thinking, better structure, and better results.
Why does it work?
Because most people prompt too early.
They rush straight into execution before the machine has properly modelled the problem. When you ask AI to generate the prompt first, you force it to slow down and think about the task before trying to complete it. It has to infer your objective, define the scope, identify the missing context, specify the output, and build a better frame.
And with these systems, framing matters enormously.
A vague instruction gives you plausible sludge. A sharply defined brief with clear constraints, role, context, and success criteria gives the model a much better chance of doing something genuinely useful.
So the real trick is not prompting harder. It’s prompting one layer earlier.
You make the machine think about the work before it does the work.
That, in my experience, is where a lot of the leap happens.
The real scarce asset now
So yes, learn these tools. Learn all of them if you can. Learn where each one is strong and where each one is weak. Learn how to move between them. Learn how to build workflows rather than just ask one-off questions. Learn how to get them to build the prompt before they do the task. Learn how to manage them, critique them, and use them without becoming lazy yourself.
Because the work is getting cheaper.
Execution is getting cheaper.
Output is getting cheaper.
But judgment, care, taste, strategic thought, and real human bandwidth?
That’s where the value is heading.
The machine can do more and more.
But the human still has to know what good looks like.
Andrew Holland
Originally published on LinkedIn on 30 March 2026.