A few weeks ago I sat with a senior executive whose team had an agency working on a piece of research for them. A brand and positioning audit across several of their businesses, a look at their strategy, the gaps in how they were positioned, the places where the marketing did not line up with what the products did, and then a new plan on top of it, positioning built around real user personas and the openings that had been missed. It was substantial work, the kind a company commissions from a good agency and waits a few weeks to receive.
We did a version of it in an afternoon. About four hours, most of it spent thinking rather than producing.
We spent the first two hours on the least glamorous part, working out what the research needed to answer and what it would take to answer it well. What to look at, in what order, what good would look like, how we would know the output could be trusted. Then we gave each part of the problem to a different model, set up checks to test what they produced against the real sources rather than taking it on trust, and pointed the whole thing at the material, mostly public, three of the company’s own public facing websites, search, media, news, joined with some internal information the team had given us access to.
Once that was built, it ran. Several models working at once, each on a different part, each checking its own output, noticing where it fell short and going again. None of them was waiting to be told the next step. My part was to set it up and keep an eye on it. At the end we had something close to what the agency was going to hand over. Not identical, but close enough that the executive sat back for a moment before asking how it had been done.
The agency almost certainly had the same models we did, and almost certainly used them. Same tools, available to anyone with a browser and a subscription. So the question is not how AI beat the agency. It is that they had the same technology, were using it, and it was still going to take them weeks.
What those weeks cost is not only time. It is attention. The people on that engagement were spending their days gathering, formatting, checking one figure against another, assembling, and those are days their minds were not free to do anything else. Attention is the one thing in a business you cannot buy more of. You can add people, add money, add tools, but each person still has one mind and one day, and most of it is being spent on work that holds their focus without needing their judgment. The four hours did not just save time. They handed a set of people back their attention, which is the scarcest thing any of them had.
The difference was not the model
The answer is not that we had a cleverer AI. We did not. The difference was in how the work was arranged.
They were almost certainly using AI, and using it well, the way most good teams now do. A research assistant that reads faster than any person, a drafter that turns a rough brief into a first version in seconds, a second pair of eyes on a document. That is a real help, and it is not nothing. But look at the shape of it. One person, opening the tool, asking it something, reading what comes back, tidying it up, asking the next thing, holding it the whole time, one turn after another. The tool is quick on each turn, so each task gets a little faster, but the work itself is exactly the shape it was five years ago. A person, doing steps, in order, by hand, with a good assistant beside them. The AI is sitting where the search engine used to sit, and doing that job well.
What we did was different in kind, not degree. We did not give a person a better assistant. We took the person out of the middle of the steps altogether. Instead of one person asking and reading and asking again, we set up a set of small runs of work that could go on their own, several at once, each on a different part of the problem, each checking itself and carrying on until it was done. My job stopped being to do the steps. It became to set those runs up, watch where they drifted, and make the calls they could not make themselves. That is the line. Not a sharper tool inside the old process, but the process itself rebuilt so the tasks run in parallel and the person moves to the top of it.
There is a name for this now, and it comes from the people building the tools. They call it a loop, and how to write loops rather than prompts became one of the live questions among builders earlier this year. Andrej Karpathy, one of the founders of OpenAI, has started calling the discipline agentic engineering, and describes the job of a developer now as directing and supervising a fleet of agents rather than writing the work by hand. The person who leads Claude Code at Anthropic says he no longer prompts the model at all; he writes the loops, and the loops do the prompting. This is how a lot of people building agents now work, and most of it has not travelled beyond them yet, which is the only reason it still sounds new.
A loop is simple to describe. It is a small piece of work that finds the thing to be done, hands it to a model, checks the result, keeps what worked, and comes round again, without a person kicking off each turn. You are not in the loop any more. You wrote it. You read what it produces and you decide what happens next, and that is the part that stays yours.
That is still abstract, so let me run one. Take the most ordinary job in an office, the Monday morning competitor update. The version everyone knows goes like this. Somebody remembers it is Monday. They open the tabs, the competitor sites, the changelogs, the pricing pages. They read through looking for what has changed since last week. They copy the relevant bits into a document, write a short summary, format it the way the team likes it, paste it into the shared space, and post a message so people know it is there. Forty minutes, give or take, and it is the same forty minutes every week.
Now the same job as a loop. The trigger is the calendar: it is seven on Monday, so the loop starts itself, and nobody had to remember. The first run goes and gathers, the same sites and changelogs, and pulls back everything that has changed since the last time it ran. A second run takes that pile and writes the summary, in the format the team uses, because you gave it three old summaries once and it learned the shape. A third run checks the second one: did it cover all six competitors, did it miss anything obvious, does it read cleanly. If something is off it sends it back to be done again, on its own, until it passes. Then it posts the summary in the shared space and drops the message in the channel. The first anyone hears of it is the finished brief sitting there when they get in with their coffee.
No one typed a prompt or opened a tab. The person who used to spend forty minutes on it now spends five, reading the brief and deciding what actually matters in it, which is the judgment that needed them in the first place. The gathering, the formatting, the posting, the remembering, all of that was the loop, and the loop now runs on something other than a person. What is left for the human is the deciding at the end, the part that was worth their time all along.
You are already the loop
The thing that changed here is not about AI at all.
Look again at who was running that loop before you wrote it. A person. And it is not only the competitor update. The same shape is everywhere once you start looking: the month end pack, the board report, onboarding a new supplier, the same tabs opened, the same fields copied, the same approvals chased, every time. Somebody is the trigger and the gatherer and the maker and the checker and the messenger, all of it, by hand. They have just never been called the loop, because until now there was nothing else for a loop to run on except a person.
Almost all of what we call knowledge work is this. Not the judgment. The running of the loop around the judgment. The remembering, the gathering, the formatting, the chasing, the passing along. We built jobs out of it because for a hundred years the only thing that could hold a loop together was a human being paying attention.
That is what has changed. The loop can run on something else now. Which means the question stops being how fast your people can run it, and becomes why a person is still running it at all.
What this asks of a leader
Once you see the work this way, the same thing happens in every part of a company. The recurring, by hand part of a job goes to the loop. The judgment stays with the person. Finance is the clearest case. The close, the reconciliation, the variance analysis that eats the first week, the pack rebuilt from the same sources every month. The loop can gather it and do the first pass and flag what does not add up, and the finance team moves up into what the job was for underneath, which is deciding what the numbers mean and what to do about them. I watched a financial controller work this out halfway through a session, that the three days she lost every month to pulling the pack together were about to become an afternoon. Her first reaction was not relief. It was a kind of annoyance that she had spent eleven years being the loop for the part that mattered least. Then she started listing everything she would finally have time to look at properly.
The headcount does not vanish. What people do with their week changes, from running the loop to deciding what the loop is for and whether to trust what it produced.
So here is where it leaves the executive across the table from me, once the demonstration is over. The question was never which AI tool to buy. That question is the easy one, and it gets you a small result. The harder question is not about technology at all. If the loop can run on something other than a person, what is the person for?
Not gone, which I want to be plain about, because the replacement framing is lazy and it is mostly wrong. This is the opposite of losing people. It is getting them out of the mundane work that filled their week and into the work that was the reason you hired them, the thinking, the judgment, the relationships, the decisions that move the business. The loop takes the part of the job that was never the point. The person keeps the part that was, and finally has the room to do more of it.
And directing that work is a real skill in itself, a new one. Knowing how to break a job down, hand the right pieces to the right models, set up the checks, and step in where judgment is needed is not something people already have. It is something teams learn. The ones who learn it do not just save time. They take on work they could not have attempted before, because a small team directing loops can cover ground that used to need a department.
Every role in the business now divides along a line, between the part that was a person running a loop by hand and the part that was the judgment. The leader’s job is to find that line, role by role, hand the loop to the machine, and move the person up into the work that was always worth more of their time. That is not a purchase. It is a redesign of how the whole place works, what a role is, what a task is, how the work moves through the building.
That is the conversation leaders should be having. Most are still having the other one, about which tool to buy, because that one is comfortable and this one is not. This one is about people they know by name. It is slow, and it should be slow. But slow is not the same as waiting, and most companies are waiting.
One last thing
Back to that agency, taking their several weeks, almost certainly with the same models open on their screens that we had on ours.
There is nothing wrong with the agency. They are good at what they do. They are doing it the way the work has always been done, with the people still inside every loop, because that is what nearly everyone does, and because the alternative means rebuilding the thing from the floor up while the work is still going out of the door, and almost nobody can stop to do that. Which is the whole problem, and it is not a technology problem.
So be honest about your own company. Somewhere in your building this morning, someone is being the loop. They are opening the tabs, copying the figures, formatting the pack, sending it round, the same as last week and the week before. They are good at it, and that is the hard part, because being good at it was what the job asked of them, and the job is changing underneath them while they do it. The reason it takes them all morning is not that they are slow or the tools are old. It is that nobody has been given the time to write the loop so they do not have to be it.
For as long as there has been work, the person has been the most intelligent thing in it. That is the part that is ending. The intelligence in the work is becoming something you point at a problem rather than something a person carries, and the job left to us is the older one underneath it, deciding what is worth doing, what good looks like, what matters and what only feels urgent. We are handing off the doing and the working out. What stays ours is the judgment about where any of it should go. That was always the valuable part. We just could never get to much of it, because the tasks took all the time we had. That is the part now coming off our hands, and with it the chance to spend people where they were always worth most, on the relationships, the ideas, the thinking, instead of the loop.
If this was useful, subscribe. The next piece looks at how these loops actually get built, and why the company that connects them well ends up with something a competitor cannot buy.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.