Earlier this year I wrote an article called It's Time to Watch the Canary. Software engineers were the first profession to have their core work taken over by AI, and I said the rest of white-collar work had twelve to twenty-four months before the same thing arrived.
I stand by that timeline. But there's a milestone inside it worth marking, because from here the question changes. The capability has arrived. What remains is adoption.
I now believe we've crossed the point where most of the white-collar doing jobs I've seen across my career can be done by well-designed, well-deployed AI agents.
I don't say that lightly. I say it as someone who has built operating models for more than 25 years, who has hired for these roles, built and run the teams that did them, and led global transformations that reorganised them. And I say it as someone who runs an AI-native business today, with eleven specialist agents doing real work every day. I know what these jobs look like from the inside. I know a good one from a bad one. And I know that a well-designed agent can now do them.
That's not a prediction anymore. It's an observation.
What do I mean by a doing job? The roles where the work is production. Someone takes a brief and produces an output. The analyst who pulls the numbers and builds the report. The planner who turns a strategy into a schedule. The coordinator who chases the approvals and compiles the status update. The specialist who processes, reconciles, drafts, checks, and files. These roles fill most org charts, and I've spent a large part of my career hiring, training, and managing the people in them.
They're distinct from the roles that decide what should be done in the first place. And they're distinct from the roles that exist to deal with people. I'll come back to that second group, because that's where the line still sits.
This isn't the first shift and it won't be the last. I've been tracking these step changes since the start, and each one has had the same shape. A capability arrives, it feels like a curiosity for a few weeks, and then it quietly becomes the floor.
ChatGPT arrived in November 2022 and signalled that something new and meaningful had shown up. Late 2024 brought models that reason before they answer, rather than blurting out the first plausible thing. Late 2025, around the release of Opus 4.5, was another step change. I wasn't the only one who struggled to pin down exactly what changed. The models didn't do anything obviously new. They just started really working, reliably, on tasks that used to need babysitting.
Now I'm making the call that we've crossed another boundary. Fable 5 was the starting gun. Fable 5.1 and now GPT-6 Astra have followed, and this generation is in another league altogether from the one before it.
The benchmarks tell part of the story. The new models are saturating tests built to stretch the last generation, and new tests keep having to be written to measure the gap. But I work with these models every day on real work with real deadlines, and the difference I notice most isn't in a benchmark. It's in how I work with them. Two things have changed.
The first is stamina. I used to have to drive the agents to keep working. I relied on a command that forces an agent to keep going until it meets defined criteria, because left alone it would stop early, declare success, or drift. I still use it where it makes sense. But since Fable 5 I mostly don't need to. The agent understands what needs to happen and keeps working until it's actually done.
The second is how I brief them. I've noticed that I now start far more prompts with a broad objective than with a specific task. I noticed it enough to take a photo of the screen. After a long back and forth on a hard problem, my prompt was: "Use your expertise and judgment to bring this to a resolution and into our build in an appropriate manner." Another, a few days later, was a version of: "I'm happy to take your recommendation. You've been closest to the work and you understand how the instructions get absorbed as we go, so go ahead with whatever you think is right."
Those are delegations, not instructions. A year ago I would have specified the path, the steps, and the checks. Now I describe the destination and trust the agent to find the route and walk it. That's a fundamentally different way of working with AI, and it's only possible because the models, the harness, and the context are all good enough at the same time. Take any one of them away and it collapses back into micromanagement. With all three in place, you're working on the other side of a threshold.
There's a second dimension to this shift, running alongside the first. Everything I've described so far is about how capable the agents have become. The other dimension is how far they can reach into the human world.
AI mastered text first. That was a bigger deal than it now sounds, because language is how most knowledge work moves. But text on its own is limited. An agent could only touch the parts of the world that could be typed to it and typed back. Then it was voice. Real conversations, in real time, with a person on the other end. And now it's the screen. Agents that can look at a user interface, understand it the way a new hire would, and operate it. Click the menu. Fill in the form. Read the table. Move to the next system.
Text, voice, screen. Each mode is a door into the human world, and each one opens onto work the previous one couldn't reach. Each one also removes an excuse.
The screen is the one that matters most for traditional organisations, because it dissolves the biggest excuse of all. "Our data isn't ready." I hear it in almost every conversation about AI deployment, and it's often true. Legacy systems. Information scattered across a dozen platforms. Nothing structured, nothing connected. For the last two years that was a real blocker. You had to get your information into an AI-native form before an agent could do anything useful with it.
That's no longer true. An agent that can operate a screen doesn't need your systems to be AI-native. It uses your systems the way your staff do. Through the same interface, filling the same forms, reading the same reports. The environment doesn't have to change first. The agent comes to the environment.
I wrote earlier this year about the two worlds of AI productivity. The human world of PDFs and decks and email threads, and the AI world where information moves at machine speed. My advice was to move your work across the boundary and keep it there. That advice still stands. But what I didn't expect was how quickly the agents would cross the boundary themselves, from the other side. The map I drew there is already being redrawn, which is exactly what that article said would keep happening.
It's worth being precise about what has crossed and what hasn't. For engineers, capability and adoption moved almost together. The models got good enough and within months the best engineers had stopped writing code. That was the canary. For the rest of white-collar work, we've just crossed the capability line. Adoption will run on its own curve, and it will be slower, for reasons that have nothing to do with what the agents can do. That curve is the twelve to twenty-four months I was pointing at.
Two things will shape that curve.
The first is design and deployment. I've been careful to say well-designed and well-deployed, and I want to be honest about how much weight those words carry. They carry all of it. The capability is here. The constraint has moved. It used to sit in the lab, in what the models could and couldn't do. It now sits in the organisation. How the agents are designed, what context they're given, how they hand work to each other, and who owns what. Almost every failure I've had running my own agent teams has been mine, not the model's. A handoff that dropped context. A role nobody clearly owned. An agent that made a decision and then disappeared, taking its reasoning with it. Those are operating model problems, and I've spent 25 years fixing exactly those problems in human teams. The difference now is that the fix takes days rather than quarters.
The second is momentum. There's enormous momentum in the economy and inside large businesses. Contracts, headcount plans, systems, habits, incentives. All of that slows the transition. It softens the timing. It doesn't change the direction.
So when I say the doing jobs can be done by agents, I'm not saying it will happen smoothly. The next few years will be littered with success stories and disaster stories, and the difference between them will almost never be which model was used. It will be how well the deployment was designed.
There are roles where I don't think agents are quite there yet. Most of the ones I can think of come from my own background in media and advertising, and they share something in common. They're built around reading and moving humans.
Account management is the obvious one. The good account director knows the client is nervous before the client has said anything. They manage a relationship, not a task list. Creative strategy is another. The strategist who can feel what will land with an audience is doing something the models still don't do reliably. These roles are about understanding a person and shifting how that person thinks or feels. That's the hardest thing we do, and it's still ours for now.
But that exception comes with a catch, and it's the part of this I most want you to sit with.
It only holds while your counterparty is human.
I've watched a version of this before. A client organisation restructured its roles and responsibilities. The agency serving it kept its old structure. Suddenly nothing lined up. The people on one side no longer mapped to the people on the other, and it didn't show up as a gradual decline. It fell apart fast, because the mismatch was felt on every single interaction.
Now run that forward. Your client deploys a procurement agent. It sends briefs at two in the morning. It expects structured responses in hours, not a deck in a fortnight. It has no nervous tell for your account director to read. It doesn't need managing. It needs serving, at its speed, in its formats, around the clock.
If a human is the only thing on your side of that relationship, it doesn't lag. It breaks.
When your client is an agent, you need to be one too. Your moat is your client's humanity. And that isn't yours to keep.
So what do you do with this?
If you're an individual, start with an honest diagnostic. Which kind of role are you in? If it's a doing role, the work you've been paid for is now work an agent can do, and the value is moving upstream to judgment, design, and orchestration. I've written about that shift throughout this series, and the advice hasn't changed. Build fluency now, on your own terms, while you still have the choice about when.
If you're in a human-facing role, the question is harder. How long is the person on the other side of your relationship going to be a person? In some industries that's years. In others, the first agent client will arrive as a procurement brief, not a press release, and most service organisations won't recognise it when it does.
If you're a leader, three things. First, it's probably time to stop using data readiness as the reason to wait. Some of you have real data problems and they still need fixing. But they stopped being a reason to do nothing when agents learned to use a screen. Second, audit your organisation for roles that exist only because the counterparty is human, and ask what happens to each one when the counterparty isn't. Third, put your effort where the constraint actually is. The design and the deployment. That's where the wins and the wrecks will be decided, and it's work you already know how to do.
One more thing, because the ladder doesn't stop.
The next mode is physical. Robotics brings the doing jobs of the physical world into range. The trades, the warehouses, the work that needs hands. That's later, and it will take longer, because the real world is a much harder problem than a screen. But it's the same ladder. Text, voice, screen, world. Each rung takes a class of work with it.
For a while yet, it's still your move, human.
Damien Healy is the founder of Qanara, an Australian AI consultancy helping businesses accelerate from strategy to impact. He writes about AI-native workflows, frontier AI capabilities, and practical transformation.
My LinkedIn articles are available via my post history and here: LinkedIn Articles | Damien Healy