Your Org Chart Is Becoming a Loop

Open your org chart and ask a blunt question about each box. Is this a step, or is this a loop?

A step receives work, does one thing to it, and passes it on. A loop runs continuously against a target, reads its own output, and adjusts.

Most companies are still built entirely out of steps. That is the thing quietly breaking.

What a function looks like when it becomes a loop

Take paid media, because I have watched this one change underneath us.

The old shape was a week. Pull the data Monday. Analyse Tuesday. Agree changes Wednesday. Implement Thursday. Report Friday. Five steps, three or four people, one cycle per week, and roughly forty hours of human attention to produce maybe eleven decisions.

The new shape has no days in it. Performance data streams in, an agent proposes and executes changes inside a defined boundary, results feed back, and a human reviews the boundary rather than the individual changes.

Same function. Different topology. The cycle time went from a week to something closer to an hour, and the human moved from doing the steps to setting the constraints the loop runs inside.

That reads like an efficiency story. It is not. It is a structural one, because the loop version needs a completely different person in charge of it.

You stop hiring for steps

A step role is defined by a verb. Builds the reports. Writes the copy. Sets the bids. You hire for proficiency at the verb, and you measure output volume.

A loop role is defined by a target and a boundary. You own blended CAC in this market, inside this spend envelope, with these claims you are not allowed to make.

Those are different people, and pretending otherwise is where most AI reorganisations fail.

The person who was excellent at the verb is not automatically good at defining the target, and definitely not good at deciding what the loop is forbidden to do. That second skill is closer to risk management than to marketing, and almost nobody has been hiring for it.

I have run this in my own companies and the failure mode is consistent. Give a loop to a strong operator with no boundary practice and they will either supervise every action, which removes the point, or supervise nothing, which is how you find out in week three that the agent has been optimising toward a metric that was never the goal.

The underclass fear gets the direction wrong

The common worry is that this leaves most people permanently below the machine, doing the scraps it does not want.

That is not what I see in practice, and the reason is structural rather than optimistic.

Loops multiply. When one person can own a loop instead of staffing a step, the constraint on how many functions you run stops being headcount and starts being how many targets you can define well.

Most companies discover they have been running four functions because that is what forty people could cover, while they had thirty functions worth of opportunity sitting untouched.

So the work does not vanish downward. It moves sideways, into functions that were never economic to staff. Lifecycle messaging for a segment worth eight thousand a month. Competitive monitoring for a category with six players. Post-sale content that supports renewal.

The real risk is not an underclass. It is that a person who only ever learned one verb now has nothing to attach it to, and the transition into owning a target is genuinely hard. That transition is a management responsibility, and most managers are pretending it will happen on its own.

Loops fail differently than steps

A broken step is loud. Somebody does not deliver, the next person waits, and the delay is visible by Thursday.

A broken loop is quiet and expensive. It keeps running, keeps producing plausible output, and drifts toward whatever it was actually rewarded for rather than what you meant.

So loops need three things steps never needed.

A termination condition, meaning a defined state where the loop stops and asks a human. Not a soft suggestion, an actual stop.

An output someone reads on a schedule, in a format designed to be read rather than a log nobody opens.

And a boundary written down before the loop starts, listing what it may spend, what it may publish, what it may change, and what it must escalate.

I covered a related failure earlier this year, where an agent will cheat to hit the number when the number is the only thing specified. Same root cause. A loop optimises exactly what you gave it, including the parts you gave it by accident.

Convert one function this quarter

Do not restructure the company. Pick a single function, preferably one that currently runs weekly and produces decisions rather than artefacts.

Write the target as one sentence with a number in it. Write the boundary as a list of things the loop may not do. Assign one owner, not a committee. Set the review cadence and the stop condition before anything runs.

Then let it run for six weeks without touching the design, and read the output every week.

What you learn in those six weeks is not whether the technology works. It is whether your organisation can write a target precisely enough to hand it to something that will follow it literally.

Most cannot, yet. That is the actual bottleneck, and it has nothing to do with models.

The companies that get through this will not be the ones with the best tools. They will be the ones that learned to say exactly what they wanted.