The Great Flattening Won’t Take Every Manager. Here Is What Keeps You Off the List.

A single lit desk standing firm while the surrounding office empties out, representing the judgement work that survives AI-driven management cuts

Judgement work is what survived when GitLab removed up to three layers of management this year. On 12 May 2026, the company told staff it would remove up to three layers of management in parts of the business and rebuild its engineering organisation into roughly sixty smaller teams, each one given full ownership of its own slice of the product. Atlassian moved before GitLab did. On 16 March 2026, it cut 1,600 roles, about ten per cent of its workforce, and co-founder Mike Cannon-Brookes said the business was reshaping its skill mix because AI had changed what kinds of roles it needed. Block went further still. On 26 February 2026, Jack Dorsey cut roughly 4,000 jobs, close to 40 per cent of staff, and later described on a Sequoia Capital podcast the thinking behind it: a system he called an “intelligence layer”, carrying the status updates, handoffs and decision routing that used to run through people.

Three companies, three different timelines, one shared bet: that a layer of management built mainly to move information between people can now be run by a system instead of a person. The part of this story nobody is covering is which managers inside these same three companies were not cut, and why.

The Work That Got Cut, and the Work That Didn’t

Watch what a middle manager actually does across a normal week. A fraction of the time goes to judgement calls that genuinely need a human: resolving a conflict between two people who both believe they are right, deciding which of three priorities loses when none of them can, coaching someone through a mistake they need to understand rather than just correct. The rest goes to relay work: taking a status update from one team and translating it for another, chasing a decision sitting unsigned with someone senior, turning six scattered updates into one summary a director will actually read. Relay work needs availability, memory and the discipline to follow up, exactly what an AI agent does without complaint, and exactly what GitLab, Block and Atlassian have already put into production.

Judgement work has not been touched. No system built by any of the three can resolve a conflict between two people who both believe they are right, or decide which of three priorities loses when none of them can, and live with the consequence of getting it wrong. Crossing that boundary is still what makes a manager’s job a manager’s job.

What Actually Keeps a Role Off the Cut List

Three things separate the managers who are safe from the ones who are not, and none of them are about looking busy.

First, owning decisions rather than carrying them. A manager who can point to a call they made this month, one with a real downside if they got it wrong, and the reasoning behind it, is doing work no relay system performs.

Second, holding context nobody else has access to. Institutional memory, the history behind why a client reacts the way it does, which two people on a team cannot be put in the same room, the unwritten reason a project failed two years ago: none of it lives in a dashboard, and none of it transfers to an intelligence layer no matter how much data that layer ingests.

Third, automating your own relay work before someone else does it for you. A manager already using an agent to chase status and summarise updates, and spending the time saved on judgement instead, is building the one profile none of these three companies cut.

The Test Worth Running on Your Own Role

I have led programme teams large enough that I needed a layer like this myself, once, and I have also sat in enough steering committees to know how often the person in the room was relaying a decision rather than making one.

The honest test has nothing to do with whether your week looks full. It is whether removing you from it tomorrow would cost the organisation a decision, or only cost it a messenger. Run that test on your own role before GitLab, Block or Atlassian’s next earnings call runs it for you.

Name the last call you made that carried real consequence if you got it wrong, and write down why you made it the way you did. If nothing comes to mind for last month, start building the habit this month.

Scroll to Top