Reorganization memos have started to carry a particular pair of decisions. Management layers come out and delivery is regrouped into small pods with agent fleets attached — and then, usually in the last paragraph where nobody is looking, a line confirming that every stage gate still requires a named human approver.
Read those two decisions together and something uncomfortable falls out. The organization just removed most of the people whose job was to know what was happening, and kept every obligation to prove what happened.
Status reporting was always testimony. You cannot take testimony from an agent, so the only thing left is to count what it produced and record who accepted the result.
That memo is not one company’s bad quarter
It is what the firms advising the board are all describing at once. Gartner expects 60% of organizations to be running smaller software engineering teams at scale by 2029, up from 15% in 2026.60 McKinsey puts the professional’s job as moving from producing artifacts to supervising the systems that produce them.61 BCG Platinion writes about software factories where as few as three engineers run delivery and humans no longer write code — and, in the same piece, that every stage gate has a human accountable for approval.62 Forrester states it in a line: humans stay accountable, but AI does more of the execution.64
Read them together and the pattern none of them makes the headline is the one that matters here. Every single forecast pairs a smaller producing layer with a named human who signs. The middle of the pyramid goes; the apex gets heavier. That is precisely the memo, and it is why the memo is not a contradiction.
PwC’s agent-governance guidance then writes the auditor’s checklist for them: a verified identity, a defined role, task-specific permissions, auditable activity records, and clear limits on autonomous action.63 In most estates today the agent has a name, a token and none of the rest.
Why the old instrument stops returning signal
A weekly status report is an interview. Someone who was present is asked how it is going, and their answer is compressed into a color and a paragraph. The whole apparatus — the RAG picker, the confidence field, the commentary box — is built around a witness.
The fallback everyone reaches for — just ask the engineers — has been measured, and it does not hold either. In METR’s randomized trial, experienced developers working in their own repositories were 19% slower with AI tools while believing they had been 20% faster.65 Self-report is not the backup plan. It is the thing that broke.
Take the witness away and the instrument does not fail loudly. It fails quietly, which is worse. Someone still picks amber. Jira Align’s own documentation concedes that its five health dimensions keep no history at all, 5 so by the time an auditor asks when the project turned, the only record of the judgment is the judgment standing today. ServiceNow’s answer is to have AI predict the RAG instead 4 — which replaces a witness with a model and leaves the auditor in exactly the same position, holding a verdict nobody signed.
The three questions that survive
Ask a compliance reviewer what they actually need from a delivery record and it reduces to three things, none of which require anyone to have been present.
What was produced, and against which version of the system? Not “the team completed the integration” but: this requirement, this citation, resolved at this commit. A pinned commit is checkable a year later by someone who was never in the room.
Who accepted it? Not who ran the build — a person, with an account, who reviewed evidence and said yes. This is the load-bearing one. An agent can propose all day; the acceptance is the governance event, and it either has a name attached or it does not exist.
Why was the decision made? Recorded at the time, by the person who made it, not reconstructed in October to explain a July reversal.
Notice that all three are answerable by counting and by custody. Notice also that none of them are answerable by a health score.
The objection: isn’t this just more paperwork?
The fair challenge is that a flattened organization removed those layers precisely to reduce overhead, and a governance model demanding rationale on every decision looks like putting the layer back.
The difference is who does the work. The old layer produced the record — a person spent Thursday afternoon assembling a status pack from other people’s recollections. In the new shape the record is a byproduct of the work happening: the agent’s proposal is already a row, the citation is already a path at a commit, the acceptance is already a click by a named person. What a PMO adds is not transcription. It is the standard that says which artifacts a project of this shape owes, and the judgment about whether the evidence is good enough — which is the part that was always the actual job, buried under the reporting.
Wellingtone’s practitioner survey has spent years reporting that PMOs cannot trust their own project data. 57 The teams that trusted it least were always the ones whose tooling asked people how things felt. That problem does not get better when the team is half agents. It gets structural.
What changes on Monday
A PMO director operating this way opens a register, not a dashboard. They read how many requirements carry evidence and how many are still waiting on a person. They see which acceptances have no name against them. They staff the accepting role rather than chasing the status update, because the status update no longer exists and the acceptance is the bottleneck.
It is a smaller job than running a reporting cycle and a more consequential one. That is roughly what everybody said the flattening was for.