AI Flips the Org Chart: From Adding Management Layers to Absorbing Them

How does AI change the organizational structure of a company? Layers exist to relay, check and coach. What AI can absorb, and what stays with a person.

Short answer: When a company grows, the usual move is to add a management layer: a supervisor above the team, then a manager above the supervisors. Layers exist because one person can only collect information from, check and coach so many people. AI can now take on parts of that work. In one large study it took over some coaching, and it is likely to handle much of the information-passing and routine checking. That can flip the default question. Instead of asking "who do we add above this team?", ask "which parts of this layer's work can AI absorb, and who still owns the result?" Research on earlier technology adds a catch: tools that spread knowledge pushed decisions down to the front line, while tools that made instructions cheaper pulled them up. AI can likely do either. A flatter company is a design choice, not something AI does for you.

A familiar pattern in growing companies: the chart gets taller, and a few years later decisions are slow and the people closest to the customer wait for answers. This article is about how work is designed, not a staffing plan.

Why companies add layers in the first place

I started as a technician in appliance repair. Then I hired a technician to do the jobs, and then a manager. That was my first layer. The usual reason for adding one is simple: one person cannot be on every job, answer every question and check every result.

One line of research treats management hierarchies as devices to acquire and pass on knowledge and information (Bloom, Garicano, Sadun and Van Reenen, Management Science, 2014). The front line solves the problems it knows how to solve. Harder ones go up to someone who knows more. Each layer is a filter for problems and a relay for information.

Flattening is not new. Raghuram Rajan and Julie Wulf studied reporting lines in more than 300 large U.S. companies from 1986 to 1998 (NBER working paper; published in Review of Economics and Statistics, 2006). The number of positions reporting directly to the CEO rose from about 4.5 to about 6.8 on average, and from 4.4 to 8.2 in the 51 firms observed for the whole period. Levels between division heads and the CEO fell. Firm growth, new positions and mergers explained only part of it. This is about large public companies in the 1990s, at senior levels, not a regional HVAC company today. The lesson: when someone credits AI for a flatter company, ask what else changed.

The four jobs inside a management layer

A practical way to see what a layer does is to break it into four jobs. This is a working model, not an established management theory.

Relay. Collecting status, summarizing it and passing decisions back down: the daily report, the meeting that exists to share updates, the question forwarded one level up.

Check. Comparing work against a standard. Was the job closed out correctly? Is the estimate complete? Did the follow-up call happen?

Coach. Helping people get better: listening to calls, reviewing jobs, showing a new person how the best person handles a hard customer.

Own. Making the call when the rule doesn't cover the situation, setting priorities, developing people, and answering for the result when it goes wrong.

Most roles mix all four. The working hypothesis here is that Relay is the easiest to absorb, Check and Coach can be partly absorbed when the standard is written down, and Own stays with a person.

What building an AI executive team taught me

I have written about how I built a second executive team made of AI seats: role-specialized AI agents. They are not officers of the company. Their job is to evaluate, challenge, recommend and coordinate. The principle is that the councils provide intelligence, not authority. The founder still decides and remains accountable.

When I was explaining agentic AI to other founders recently, I used the org chart. A project manager can hire frontline people but cannot hire directors. Directors can hire the managers who hire the workers. VPs hire the directors. My view is that agentic AI only really pays off for people who can manage many layers of an organization. It depends on your thought process: how you start the process and how you end it.

Working with AI agents is management work. Someone still decides what each agent is for, what good output looks like, and when it should stop and ask. In this article's terms, the AI seats help with Relay and Check, while Own stays with a person.

What the evidence says AI can absorb

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a generative AI assistant rolled out to 5,172 customer support agents at a Fortune 500 software company, most of them outside the U.S., mainly in the Philippines (Quarterly Journal of Economics, 2025). Issues resolved per hour rose 15% on average and about 30% for less skilled and less experienced agents. The most experienced and skilled saw little gain in speed and a small drop in quality. The authors offer suggestive evidence that the tool spread the practices of the best agents to less skilled ones.

Two details are about management work. Before the tool, managers coached agents in brief weekly sessions and could give feedback on only a small fraction of conversations. The tool gave suggestions in real time, during the conversation. And customer requests to speak to a manager fell by almost 25%, from a baseline of about 6% of chats, with the point estimates largest for less experienced agents. Read through the four jobs: part of the Coach job and part of the escalation path, work that normally sits in the layer above, moved into the tool at the front line.

The limits: one company, chat support, not field service, and requests for a manager rather than actual escalations. The authors say the study is not designed to show effects on overall employment or wages. It shows that a layer's work can move. It does not show that the layer is no longer needed.

There is also firm-level evidence. Using data from 2010 to 2018, before today's generative AI tools, Tania Babina and co-authors found that U.S. firms investing more in AI tended to flatten, with higher shares of junior workers and lower shares in middle-management and senior roles (NBER working paper, 2023). That is an association, not proof that AI caused it.

Another study shows where absorption fails. In a preregistered experiment with 758 knowledge workers, run with Boston Consulting Group (Dell'Acqua and co-authors, Harvard Business School working paper), people using GPT-4 on 18 realistic tasks inside what the researchers call the AI "frontier" completed 12.2% more tasks, 25.1% faster, with significantly higher quality. But on a complex task chosen because it sat outside that frontier, people using AI were 19% less likely to reach a correct solution than people without it, in the authors' words. The frontier is jagged: two tasks that look equally hard can sit on opposite sides of it.

One reading for the org chart: being good at many tasks does not make AI good at every task that looks similar. The judgment calls that make up Own are where an unnoticed error costs most, so that is where AI deserves the least trust.

None of these studies measures relay work directly. That Relay is easiest is a judgment from how the work is structured, not a finding.

The mechanism: the default flips, but in which direction?

If Relay and routine Check get cheap, one person may be able to see more of the business without losing track of it. In this model, a layer that existed mainly to pass on information loses its main reason to exist, and the next one can wait.

Bloom and his co-authors studied firms in the U.S. and seven European countries and separated two kinds of technology. Information technology that gave people more knowledge (planning software for plant managers, design and manufacturing software for production workers) was associated with more autonomy and a wider span of control. Communication technology, such as data networks that made it cheaper to send instructions, was associated with less autonomy for both workers and plant managers.

A reasonable inference is that AI can do either. An assistant that helps a technician diagnose a unit gives knowledge to the front line. A dashboard that lets the owner watch every job and approve every exception makes instructions cheap, so decisions flow to the top. Both can remove a layer. One builds capable people with good tools; the other, a company where the top approves everything, faster.

I prefer the first. As I wrote in that earlier essay, our vision isn't a company without people. It's a company where every person has access to more intelligence. That is what I want, not evidence that it works. Which way a company goes depends on where it points the technology.

The strongest case against flipping the org chart

The best argument against this idea is that layers do things a task list doesn't capture.

Good managers create lasting value. Using personnel records from a large multinational (200,000 white-collar workers and 30,000 managers over 11 years in 100 countries), Virginia Minni found that workers who gained a fast-promoted ("high-flyer") manager when managers rotated across teams made 40% more lateral moves and earned 13% more seven years later (Quarterly Journal of Economics, 2026). The gains lasted after the manager moved on. In this article's terms, matching people to the right work is Own and Coach, not Relay.

Layers are where people learn to lead. The supervisor job is where the next managers come from. Remove the layer and you may remove the training ground. Even Gartner, which predicted in October 2024 that through 2026, 20% of organizations would use AI to flatten their structure, eliminating more than half of current middle management positions, listed the risks: managers overwhelmed by more direct reports, broken mentoring pathways, and fewer chances for junior workers to develop. That is a forecast, not evidence, and in our search we found no measurement of whether it came true. The risks are the ones to take seriously.

Accountability and trust are human. A customer with a failed installation or an inspector with a permit issue wants a person who can decide and answer for it. And the jagged-frontier result shows AI can make work worse outside what it does well, where it is not always obvious which tasks those are.

Smaller companies don't have many layers to absorb. A $5–20M home service company may have one or two layers between owner and technician. Its question is when to add the next one, and whether it should look like the last.

AI work isn't free either: someone has to set up the agents, check their output and correct them, and that oversight is itself management work.

What does this reveal? The thesis survives in a narrower form. AI can plausibly absorb much of Relay, and in one setting the evidence is consistent with it taking over part of Coach and escalation. Nothing we reviewed shows it absorbing Own, or the development path layers provided. Take the work out of a layer, and you have to rebuild the learning on purpose.

When this fails

  • There is no written standard. AI can't check work against a standard nobody wrote down. It will check against a guess.

  • The data is messy. If job records and statuses are incomplete, AI relays the wrong picture faster.

  • The work is physical and high-stakes. Coaching an installer on a roof happens in person, and licensing and code rules put responsibilities on specific people.

  • The decision is new. Novel, high-stakes or ambiguous decisions are more likely to sit outside the frontier.

  • AI is used to watch rather than to help. If it is used mainly to monitor and send instructions from the top, decisions may move up, not down, the centralizing pattern earlier research found for communication technology.

  • Nobody measures the AI's errors. Then mistakes become invisible.

What this changes for operators, investors and founders

For operators. When a team feels overloaded, don't start by drawing a new box. Sort the manager's week into Relay, Check, Coach and Own first. If most of it is Relay, a tool may give the manager that time back. If most of it is Own, you probably need another person.

For investors and acquirers. Ask which reporting and checking work is automated, how its errors are caught, and where accountability sits when the AI is wrong. How will it develop its next managers? A company run by a few people plus AI agents depends more on those people and its systems; price that risk. Using AI agents also does not make a home service business a software company, and it should not be valued like one.

For founders. Your job moves from managing people through layers to designing the system the layers run on. I wrote earlier, more bluntly, that the old playbook is dead: the one where you grow by adding people, managers and complexity. This is the narrower, more testable version of that idea. The goal is fewer relays between the customer and the person who can decide, with AI carrying the information and a person owning the result. How the AI seats themselves are scoped is in What Is an AI C-Suite?

How to test this in 30 days

Don't reorganize anything during the test.

Week 1: Map one manager's week. Pick one team, list its manager's tasks for a typical week, and tag each Relay, Check, Coach or Own.

Week 2: Hand over one Relay task. For example, the daily status summary. AI produces it, and the manager checks it against reality every day and notes each error.

Week 3: Add one Check task with a written standard. For example, checking that closed jobs have the required photos and notes. AI flags the exceptions. The manager reviews every flag plus a random sample of unflagged jobs, to see what it misses.

Week 4: Measure and decide. How much time came back, how many errors did the AI make, and where did the freed time go? If it went into Coach and Own, the flip is working for that team. If it went into more Relay, or errors are too high, fix the inputs first.

Frequently asked questions

How does AI change the organizational structure of a company?

It changes the cost of the work management layers do. AI can likely take over parts of the information relay (reports, summaries, status updates) and some routine checking, and in one study it took over part of the coaching, so one person may be able to oversee more and a company may be able to delay adding layers. Research on earlier technology linked it to pushing decisions down to the front line or up to the top, depending on the kind, so the structure is a design choice.

Will AI replace middle managers?

The evidence so far shows AI taking over specific tasks inside roles; in customer support, real-time suggestions did some of the coaching and requests to speak to a manager fell. Forecasts of large reductions exist, such as Gartner's 2024 prediction, but a forecast is not evidence. The studies we reviewed do not show judgment, accountability or developing people transferring to AI.

Should a small home service company flatten its org chart because of AI?

Not as a goal in itself. A $5–20M company usually has few layers, so the better question is what the next layer would do. If it is mainly relay and routine checking, test AI first. If it is mainly judgment and developing people, you likely need a person.

What is supported, what is uncertain, what to test

Supported by evidence: Large U.S. companies flattened their senior hierarchies in 1986–1998, long before generative AI. Technology that spreads knowledge was associated with more autonomy and wider spans of control; technology that cuts communication costs, with more centralization. In one large customer support operation, an AI assistant raised productivity most for newer agents and cut customer requests to speak to a manager. In an experiment with knowledge workers, AI improved results on tasks inside its frontier and made them worse on a complex task outside it. In one multinational, good managers produced lasting gains in workers' careers.

Hypothesis this article argues for: that AI can flip the default question from "who do we add above this team?" to "which parts of this layer's work can AI absorb?", and that the Own job stays with a person.

My opinion: that agentic AI rewards people who can think across many layers of an organization, and that the goal should be a company where every person has access to more intelligence, not one where only the top sees more.

Uncertain: whether the customer support results carry over to field trades; whether managers can handle wider spans without burning out; how companies will develop future leaders if fewer people move up through supervisor roles; and what it really costs to run, check and correct AI agents. In our search we found no rigorous public study of AI and management layers in home services.

What to test next: After the first 30 days, keep the test going on that team for a full season, because volume swings with the weather, and keep a comparison team working as before. Track the manager's time by the four jobs, the AI's error rate, and one customer outcome, such as callbacks or response time. If the freed time doesn't move into coaching and ownership, or customer outcomes slip, the thesis needs to change for that kind of team.

The org chart grew upward because information had to be carried by people. Much of it may no longer need to be. What still has to be carried by a person is the decision, and the responsibility for it.