Why AI's Productivity Boom Has Skipped Management
François Bossière — Founder & co-CEO, Polynom
In software engineering, AI productivity has stopped being a thesis. It has become a measurement. Coding with Copilot is roughly 55% faster than coding without it, and the tool is now in use across about 90% of the Fortune 100. The line on the chart goes up and to the right.
Walk down the hall to any management floor and the chart looks nothing like that.
Managers, directors, partners, and the C-suite have all dutifully adopted AI assistants. They draft emails with them. They summarize documents. They occasionally ask one to clean up a slide. The gains are real but small — minutes a day, not orders of magnitude. The calendar looks the same. The organization runs at the same pace.
This gap is not a story about technology. It is a story about where the time actually goes.
The work is not the document. The work is the meeting.
Harvard Business Review has tracked the number for decades, and it now sits around 23 hours a week in meetings for senior executives — more than double the figure from the 1960s. The layer below is not much better. Managers, partners, account directors — anyone whose job is to coordinate other people — spends the majority of the week in synchronous conversation.
This matters because it changes what "office productivity" means.
For a software engineer, the unit of work is the function, the file, the pull request. AI met the engineer at that unit. That is why the gains showed up.
For a manager, the unit of work is the meeting. The document is residue — written before the meeting to prepare for it, written after the meeting to record it. The decision itself, the alignment, the commitment — all of it happens during the call.
Office AI has been pointed at the residue. It has barely touched the meeting itself.
That is the entire gap, and it explains everything else.
The short-term move: stop single-threading the hour
A manager's hour is currently bounded by a single thread of attention — the meeting in front of them. Everything else queues. There are two ways to break that today, and they compound.
Put AI inside the meeting. Most current tools sit at the edges: a notetaker before, a summarizer after. The hour itself remains an analog event in which six to twelve highly paid people consume each other's time at full attentional cost, then spend the next two days producing the artifacts the meeting required. The technically feasible move now is to embed AI as an active participant during the meeting. Retrieve the contract clause when it is referenced, not three days later. Draft the customer reply as the decision is made, not the next morning. Push the action items into the project tool, with owners and dates, before anyone stands up.
The reason this is now realistic is that the assistants are connected. The current generation of agentic AI — the kind that plugs into your inbox, your drive, your CRM, your project tracker — already has access to the answer for most of what comes up in a meeting. In a typical management discussion, something like 80% of the questions raised are easily derivable from documents that are already on hand. What did the client say last week? — in the inbox. Where did we land on the pricing? — in a slide deck three meetings ago. What is the current burn? — in a spreadsheet. These questions still get answered with "I'll get back to you" not because the information is hard to find, but because no one in the room is searching for it in real time.
But this only works if the meeting has a real agenda. Most do not. Most have a title, a vague topic, a list of attendees, and a conversation that finds its own course in the room. An assistant cannot operate against that. A well-posed agenda — specific items, specific decisions to make, the documents each item depends on — is what allows the AI to anticipate the retrieval, prepare the artifacts in advance, and help pace the conversation. The agenda is the input that turns the assistant from a passive listener into an active operator. If you want AI inside your meetings to work, the first thing to fix is not the AI. It is the agenda.
Put AI outside the meeting, working on something else. While you sit through the call, your AI does not have to wait for you. Agentic compute can run in parallel on unrelated work — answering the customer emails sitting in your inbox, building the Excel report your CFO asked for, preparing the briefing you need for your next call, kicking off a research sequence that will be ready when you stand up. You are in one meeting; your assistant is producing the output of two more. The hour is no longer single-threaded.
Put together, these moves change the math of a manager's day. The same calendar produces materially more output, without anyone touching the calendar itself. That is the first real productivity step-change for management work. It does not require a new model. It requires pointing the existing ones at the right hour.
The long-term move: representation, not attendance
The deeper problem is that most participants in most meetings are not the decision-maker for most of the agenda. They are present for context, for the few moments where their input is actually needed, and for the political cost of being absent. Senior people know this. They sit through hours of discussion to contribute for minutes.
The long-term lever is representation.
An AI proxy that attends on behalf of a manager. Holds full context on their priorities, positions, and constraints. Listens through the entire conversation. Pings the human only when a decision is genuinely required.
The rest of the time, the manager is doing other work — or, more honestly, doing the work.
This is not a transcription tool with a chat interface bolted on. It is a different topology of the workday. The unit of attention becomes the decision, not the hour-long block. Synchronous time collapses into the moments that actually need a human.
The trust and design problems are real. Faithful preference modeling. Escalation policies the principal actually agrees with. Real-time disagreement detection. Protocols other participants will accept. None of this is trivial. None of it is unsolvable on a horizon of two to three years.
This is where the next order of magnitude in office productivity sits.
The reframe
The software story worked because AI met the engineer at the unit of work. The office story has under-delivered because AI has been pointed at the wrong unit — the artifact instead of the meeting.
The organizations that close the gap will not get there by adding more AI to Outlook. They will get there by treating the meeting hour as the atomic unit of management work — parallelizing it and instrumenting it in the short term, replacing attendance with representation in the long term.
The productivity curve management has been waiting for is real. It will arrive not when the models get smarter, but when we stop forcing managers to work one hour at a time.
Polynom designs and operates agentic AI services for European enterprises — the production-grade systems behind the moves described above. Operated, governed, accountable. Not pilots, processes. If you want to discuss what this looks like inside your organisation, find us at polynom.io.
