A CEO fires off a request that hasn’t been thought through. A few days later, first review. The work comes back and the CEO is immediately frustrated — “this isn’t what I wanted.” Nobody in that room stops to examine the ask. The team gets the pressure. The brief stays broken.
The great PM or CoS doesn’t let it get there. Before the work starts, they ask questions. What are we actually trying to decide? Who’s the audience? What does done look like? The CEO, pushed to articulate what they actually want, usually realizes mid-answer that the ask was half-formed. The brief tightens. The work that comes back is sharp — and the CEO can’t be angry at it, because they co-created it. The questions weren’t friction. They were the job.
Some AI models do this too. Claude will stop and ask before executing when the brief is thin. It surfaces the ambiguity instead of burying it under a confident wrong answer. Research on structured prompting consistently finds that engagement with the brief before execution — whether through clarifying questions or tighter framing — produces measurably better output than guessing. Most people have never noticed this, or treat the questions as an obstacle and skip past them. That impatience is the same mistake the CEO makes. You are choosing the wrong variable to optimize.
The same dynamic plays out in any AI session, with one important difference. I gave the model a brief. Three paragraphs. Context, goal, what I wanted out the other side. It came back with something polished, well-organized, and completely wrong. My first reaction: you were not listening. But the model cannot listen or fail to listen. It processed what I gave it and returned its best output against an incomplete brief. That is exactly what happens when a team fills in the gaps from unclear direction — except the model does not fill gaps carefully. It hallucinates. It returns invented certainty where there should have been a question.
If the output was wrong, one of two things happened: the brief was bad, or the tool was wrong. Both are fixable. Neither is where the frustration usually points.
At small agencies and lean startups, the chief of staff or senior PM fills a specific operational gap. They hold the founder’s context. When the CEO is back-to-back and a decision needs to happen, the CoS makes it, because they know what the CEO would say. When three teams are pulling in different directions, the CoS translates strategy into direction.
Matt Mochary, in “The Great CEO Within”, defines this as the core function of the role: context holding. The CoS exists to store and apply the CEO’s thinking so the CEO does not have to be everywhere at once. Not cheerleading, not organizational politics — those are real skills, but they are in service of the underlying function. The function is information synthesis and context distribution.
That is also what a well-run assistant session does. Jensen Huang, at Davos in January 2026, described this shift directly: the IT department, he said, is becoming “the HR department of agentic AI” — responsible for onboarding, managing, and directing digital employees who work alongside biological ones. Elad Gil, writing in April 2026, noted that revenue can grow 30 to 100 percent while headcount stays flat. The operations and PM functions are the first to absorb the change.
The context-holding function did not disappear. It moved.
But then the model misses. And the question is not just “what went wrong.” It is which variable failed.
There is a cost specific to the model that human teams do not pay. Ambiguity does not produce imprecise output. It produces hallucinations. A team might execute carefully in the wrong direction. The model fills the gap in your brief with the most plausible-sounding answer it can construct — it invents with confidence because that is the only option available when direction is missing. Ethan Mollick, whose research informs his 2024 book “Co-Intelligence,” has documented that when people assign ongoing roles to AI systems, they apply the same accountability frameworks they use for human colleagues. When the model misses, the frustration is calibrated to the trust extended, not the technical severity of the failure. We experience the consequence of a bad brief as if it were a personal betrayal.
But that is only one side of the equation. Output quality depends on two variables: the brief and the tool. And not all tools are equal.
A great chief of staff interprets a rough ask and returns something sharp — but more than that, they stop you before the ask is rough. They ask the clarifying questions that surface what you actually need before the work begins. A mediocre one receives an ambiguous ask and executes it literally, precisely, and in entirely the wrong direction. They never push back. They just deliver.
The same difference exists between models. Claude asks. When the brief is thin, it surfaces the gap rather than papering over it with a confident wrong answer. Treating AI models as interchangeable is the same mistake as treating any warm body as a CoS. Claude is not GPT is not Grok is not Copilot. Each operates differently, fails differently, and has different patterns that bring out its best work. The 2025 Stack Overflow Developer Survey found Claude Sonnet the most admired LLM among developers — 67.5 percent admired rating — with Claude Code already at 10 percent IDE adoption in its first year. Professional developers, whose output is immediately testable and who have no reason to prefer one tool over another for political reasons, chose Claude. That preference was strong enough to show up even inside Microsoft.
In December 2025, Microsoft gave thousands of its own engineers access to Claude Code. By May 2026, they were canceling the licenses. The reason given: cost. Per-engineer API spend was running $500 to $2,000 a month, with usage rates between 84 and 95 percent. The fix: replace Claude Code with Copilot CLI, the tool Microsoft builds and sells. As reporting at the time noted, Microsoft had “invited a rival tool into its own engineering culture, watched developers adopt it enthusiastically, and then chose platform discipline over tool pluralism.”
But cost is not what the numbers describe. Usage at 84 to 95 percent means engineers were actively choosing the tool every day. At that adoption rate, a $500-to-$2,000 monthly spend is not a cost problem. It is a communication problem. Every unnecessary token is a reprompt, an iteration on a vague brief, a loop that a tighter ask would have avoided. Engineers generating the most tokens were prompting least efficiently — burning compute iterating on ambiguous output, reprompting instead of thinking first. Leadership read the bill as a budget line item and pulled the better tool. They optimized for the wrong variable. A company that sells AI infrastructure to the world apparently could not build AI-fluent employees. That is the tell.
The CoS role traces to Sherman Adams in Eisenhower’s White House in the early 1950s. The president had too many inputs, too many decisions, too many people who needed answers. Adams became the filter: information in, decisions out. He was the first person to hold the explicit title of “Chief of Staff” — Eisenhower copied the structure directly from military practice. The role spread into corporate America not because executives needed someone to talk to, but because the information architecture of large organizations created the same bottleneck Eisenhower faced. Mochary’s “context holding” definition is not a soft or cultural concept. It is structural. We put a person in that function because there was no better option. Now there is. When we feel betrayed when the model misses, we are revealing how thoroughly we absorbed the fiction that the role was ever about the person rather than the function.
When you are angry at Claude, two honest questions: Was the brief good? Is this the right tool? The organizations winning with these systems answer yes to both. They invest in communication — how to give direction, how to define done, how to scope a problem before handing it off. And they use the tools their best people actually want to use, not the ones that make the budget spreadsheet easier to read.
The great CoS and the mediocre one are not interchangeable. Neither is the brief.