I suspect AI means more managers and meetings, not less

There have been plenty of headlines over the past few years about tech companies eliminating middle managers with "AI" being given as the justification:

• Amazon cuts ~14,000 corporate roles (Oct 2025) to be "organized more leanly, with fewer layers"
• Block lays off ~40% of its staff (Feb 2026) because "a significantly smaller team, using the tools we're building, can do more"
• Coinbase cuts 14% and replaces "pure managers" with "player-coaches" (May 2026), each with 15 or more reports

Sometimes by implication and sometimes explicitly, the idea is that AI should enable the existing manager to handle more reports and/or enable them to perform as individual contributors (ICs) alongside their managerial responsibilities.

But to me, this seems backwards. It's clear at this point that AI has made me as an IC more productive, and some of my coworkers significantly more productive. But it hasn't moved the needle nearly as much on the core of what good managers do: communication. As of today, LLMs are great at writing PRs, mediocre at writing design docs, and terrible at keeping people abreast of all the decisions that are relevant to them. All of that suggests that we need more management, not less.

To formalize that, here's a toy model. Let's say a company is a balanced tree, with a CEO at top, managers in the middle, and ICs at the bottom.

• $\boldsymbol{G}$: the company's goal direction. The farther in this direction we move, the better
• $\boldsymbol{C}$: the efficacy of communication. Higher => alignment is easier
• $\boldsymbol{P}$: productivity. Higher => the more stuff each IC does in their direction
• $\boldsymbol{D_i}$: the "direction" each employee is working in, randomly initialized
• $\boldsymbol{W_i} = P \cdot D^*_i$: the work done by each employee
• $\boldsymbol{W_c}$: the work done by the company as a whole
• $\boldsymbol{S}$: success, how far the company moved in the desired direction
• $\boldsymbol{R}$: the number of reports each manager (including the CEO) has
• $\textbf{\textit{IC}}_{\boldsymbol{m}}$: the percent of time the manager spends on IC work
• $\boldsymbol{A}$: time managers spend "aligning" their reports
  • $A = \frac{1 - \mathit{IC}_m}{R + 1}$
  • "The manager has to split their time between IC work and aligning each of their reports"

The trickier question is, how do we define "alignment". I went with:

• $\boldsymbol{D^*_i}$: direction after alignment, $D^*_i = \operatorname{slerp}(D_i, D^*_{\mathit{mgr}}, \lambda)$
• $\boldsymbol{\lambda}$: amount of alignment, $\lambda = 1 - e^{-C A / W^*_i}$
• $\boldsymbol{W^*_i}$: the weighted amount of work $i$'s sub-org is doing:$$\begin{aligned} & W^*_i = P \cdot \mathit{IC}_i \\ & \quad + \ln\Bigl(1 + P \sum_{j \in \operatorname{subtree}(i)} \mathit{IC}_j\Bigr) \end{aligned}$$

Which says basically "Aligning a little is easy, aligning a lot is hard, and the amount of work it takes is proportional to the weighted amount of work your sub-org is doing"

Obviously very simplistic (doesn't account for efficiencies of scale, cross-team communication, parallelizability vs. sequentiality of work, etc.), but as a simplification, I feel you can do worse.

When you plug all those things together, you find it's true: The greater the productivity, and the harder communication is relative to it, the more managers, with fewer reports you should have:

Some fun details this calls out:

1. When communication is hard enough, and productivity is high enough, there's a flipping point where the ideal share of a manager's time spent on IC work is 100% instead of 0% (the "Screw it, I'll do it myself" point)
2. The only other time it makes sense for a manager to do things themself is when communication is really easy relative to productivity (the "I have nothing left to teach you" point)
3. As productivity increases relative to communication ease, the advantage of a smaller firm in per-employee productivity over a large firm grows
   1. In this model small firms are always more productive per-employee than large firms, because we have no term accommodating for efficiencies of scale. The important aspect is the relative ratio

Of course, it's very possible the regime we're in now, where AI improves productivity, while doing relatively little for communication is a temporary one. AI bots with superhuman ability to understand what context you need to do your job, and how to route it to you are eminently plausible. But assuming that's still a little ways out, I expect we'll see companies quietly walking back their push towards high report counts and "player-coaches".