In short: AI is a capable assistant that cannot manage itself. Somebody has to think, assign, delegate, critique, tune, refine and test, and that somebody needs enough depth in the subject to know when the answer is wrong. Whether AI ends up making your business more profitable or just wasting everyone’s time depends almost entirely on the effort and understanding of the people directing it. And the direction of travel is not in doubt: this is the biggest change I have seen in thirty-two years of IT.
What it is genuinely good at
I want to be fair to it before I start listing limits, because I use this stuff constantly and I would not go back.
I ask for its perspective all the time. It contributes meaningfully to building things and to working out why something is broken. It will spot the thing I have stared past for an hour. It reads faster than any human alive, it never gets bored on the four hundredth repetition, and it is genuinely good company on a hard problem at eleven at night.
That is not a small thing and I am not damning it with faint praise. It is a real capability and businesses that ignore it will be at a disadvantage.
What it cannot do
It does not have the depth of knowledge, and it does not have the intimate experience, to make the judgement calls.
Those are two different gaps and both matter. Depth is knowing that the obvious answer is wrong for a reason that is not written down anywhere. Intimate experience is knowing this particular business, this particular client, what was promised in a meeting three years ago, and which of the two technically correct options will actually get used by the people who have to live with it.
It will give you a confident answer regardless. That is the part people underestimate. It does not hedge when it should, and a wrong answer arrives in exactly the same tone as a right one. We have measured this on our own systems, where models given a hostile instruction hidden in their own data produced professional-looking recommendations that would have destroyed a machine’s ability to recover. Nothing about the output looked wrong. That is the whole problem.
So it needs a manager
Think, assign, delegate, critique, tune, refine, test. That is not a metaphor, it is a job description, and it is the same one that has always applied to running capable staff.
Think about what actually needs doing before you ask for anything. Assign the parts that suit it and keep the parts that do not. Delegate properly, which means scope and context and an example rather than a vague question. Critique the output against a standard you hold, not against whether it sounds good. Tune the instructions when it gets something wrong, because the fault is usually in the brief. Refine as you learn what it is reliably good at. Test it against how it fails, not just how you hope it works.
Skip those and you get a plausible-looking mess produced very quickly. That is the actual risk here, and it is not a technology risk. It is a management one.
The businesses seeing real returns from AI are not the ones with better tools. Everybody has the same models. They are the ones where somebody took the manager’s job seriously.
Where this is actually heading
Now the part that should concern anyone running a business, because the direction is much clearer than the timeline.
Look at how AI is being used rather than at what is being said about it. Anthropic publishes an economic index built from samples of its own traffic, and it separates two things worth separating: augmentation, where a person does their work with help, and automation, where the task is handed over and the output flows straight on to something else.
The split between people and businesses is stark.
Look at it task by task and the gap widens further: on that same sample, 97% of economic tasks done through the API showed automation-dominant patterns, against 47% on the consumer side. Only 12% of API transcripts looked like augmentation at all.
The pattern has held since. The following report, covering November 2025, found consumer use drifting back toward augmentation at 52% against 45% automated, while automated use stayed dominant through the API. Computer and mathematical work made up 34% of consumer conversations and 46% of API traffic.
That is the real signal. The visible half of AI is people typing questions into a box. The half that will actually change the economy is programmatic, running without a human in the loop, and it is the half businesses are buying.
Third-party figures circulating this year suggest automated agent usage overtook direct human usage during 2026 and grew many times over in the months after. I have not verified those independently and I would not build a business case on them, but the shape matches what I see: the human-typing-questions line is rising steadily, and the machines-calling-machines line is going up at a much steeper angle.
Big business can see this. The investment is heavy and much of it is quiet, because nobody announces this sort of programme until it works.
The bit that surprised me
The conventional line is that the low-paid and unskilled work goes first, then the repeatable white-collar work.
The data does not currently show that. In the same index, computer and mathematical work is the single largest category of AI use: 34% of consumer conversations and 46% of API traffic in November 2025. On the consumer side it has actually come down from a peak of 40% in March 2025, which suggests everyone else is catching up rather than that programmers are easing off.
In other words, the first occupation to hand a large share of its daily work to machines was one of the best-paid and most skilled professions in the economy. Software engineering. It happened first there because the work is text, the feedback is immediate, and the people involved were closest to the technology.
I think that reorders the question. It is less about which jobs are lowest paid, and more about which work is text-shaped, high-volume, and quick to verify. That description covers a great deal of what happens in an accounting practice, a law firm, a real estate office and a medical clinic, and comparatively little of what happens on a building site or behind a coffee machine.
The new industrial revolution, and what it means at your size
This is the biggest change I have seen, and I have watched computerisation, the internet, the cloud and smartphones come through. Some of what it brings will be good and some of it will not be, and it is going to keep increasing either way. Every previous wave worked the same: the businesses that refused to adapt did not lose a bit of market share, they disappeared.
But adapting has never meant buying the maximum amount of the new thing. For a business of ten or fifty people, this is not a story about replacing your staff. Most small and medium businesses are not overstaffed, they are underwater. The realistic outcome is not fewer people, it is the same people doing less of the work that grinds them down, and more of the work only they can do.
Which brings it back around. If AI is the assistant, the roles that survive and grow are the ones that manage it: the people who can think, assign, critique and test. Those are judgement jobs, and judgement is exactly what it does not have.
The businesses that come out of this well will be the ones who worked out early which half was which.
If you want a straight answer about which half is which in your business, that is what an AI opportunity audit is for. If you would rather start with the skill, I have written about what good delegation to AI actually looks like and how to work out whether it is paying for itself.
Usage figures and the chart above are drawn from Anthropic’s published economic index, which states its own sampling method and data periods; the chart is our own, built from their numbers rather than reproduced from their report. The agent-versus-human growth figures are third-party and are flagged in the text as unverified, because that is the standard we hold ourselves to at /experiment. Drafting and structure: AI, working from my notes, with the judgement calls mine.