Jason Vs The Noise Field notes, insights and projects

Insight / signal

If one person can support 50 product managers, the job has changed

Published12/07/2026
FormatInsight

The most interesting line in OpenAI's ChatGPT Work launch was not "GPT-5.6".

AIagentsoperationsJasonVsTheNoise

The most interesting line in OpenAI’s ChatGPT Work launch was not “GPT-5.6”.

It was not the desktop app. Not the connected tools. Not the part where it builds docs, decks, sheets and little apps out of your work materials while you get a coffee.

It was a customer example buried in the middle.

OpenAI says a manager at RingCentral used ChatGPT Work to turn manual monthly launch checks into a repeatable workflow. It reviewed release plans, Jira tasks and go-to-market schedules. It flagged missing steps, blockers and unclear ownership. It produced source-backed reports naming owners and next steps. According to the case study, that took him from supporting one product manager to supporting roughly 50.

Vendor case studies need salt. A proper handful. This one was selected, polished and written to sell a product.

Even with the salt, that is the line worth paying attention to.

Not because AI quietly replaced 49 people. That is the lazy reading, and it is the one every LinkedIn post will go with.

The useful reading is that AI changes span of control.

One good operator can now supervise far more parallel work, as long as the work is structured, visible and reviewable. That is a different claim from “AI saves time”. It is also a much more awkward one for business owners, which is usually a sign it is worth sitting with.

Task speed is the small story

Most AI conversations are still stuck on the stopwatch.

Can it write the email faster. Can it build the spreadsheet faster. Can it summarise the meeting faster. Sometimes yes. Sometimes it produces an impressive mess at speed and everyone nods along because the demo had nice music.

The bigger shift happens when tasks stop arriving one at a time.

ChatGPT Work sits across apps and files, runs scheduled tasks, and keeps working while you are away. Google is describing agent platforms with memory, sandboxes, tool retrieval, agent-to-agent communication, identity and threat detection. The Grok and Hermes operator demos doing the rounds this week show the same thing from the other side: persistent cloud agents, parallel sessions, connected tools, fast loops.

That is not a chatbot writing you a paragraph. That is a new management surface.

The human role moves from doing the assembly to deciding what should happen, what counts as done, what evidence is required, which exceptions matter, and when the machine is allowed to act on its own.

Expertise is the multiplier, not the model

Anthropic’s research on agentic coding lands in the same place from a different angle. Across Claude Code sessions, humans made most of the planning decisions and Claude made most of the execution decisions. The more expert the user, the more useful work came out of the agent, because they knew how to frame the problem, inspect the output and recover when it went sideways.

That feels right, and it is the part the hype skips.

The agent gives leverage to the person who understands the work. It does not create judgement in someone who cannot tell good from bad.

Which is why both loud camps are wrong. AI will not replace management. It is not autocomplete either. It creates a new kind of management.

Less “have you done the spreadsheet”.

More “show me the source trail, the exception list, the owner, the approval state and what changed since last week”.

That sounds dry. Good. Dry is where the money is.

What this does to a marketing business

A strategist used to wait. Wait for research, notes, performance data, client context, campaign assets, reporting inputs. Different people, different tools, usually assembled in a rush the night before the call.

Now an agent loop can gather the raw material, draft the brief, flag missing proof, surface anomalies, prepare the client report and suggest next actions.

That does not make the strategist useless. It makes a weak strategist more exposed and a strong strategist more dangerous.

Same in client service. Same in sales ops. Same in finance. Same in delivery.

The work does not disappear. The centre of gravity moves.

Span of control only works if the control part is real

Here is where most companies will trip over their own enthusiasm.

They will read “one person can support 50 product managers” and hear “we can make everyone 50 times more productive”. Then they will connect every tool, relax every permission and wonder why nobody trusts the output six weeks later.

If one operator is supervising 50 workflows, those workflows have to be boringly legible.

A hidden agent doing mysterious helpful things in the background is not leverage. It is a liability with a friendly interface.

So you need a work queue, not a chat history. You need real task states: requested, running, blocked, needs human, approved, shipped, failed. You need source trails, because “the AI said so” will not survive contact with a client, a finance director or a regulator. You need exception reports, because nobody inspects every line of every generated artefact forever. You need permissions matched to the risk, because reading a dashboard is not the same as emailing a customer, drafting a report is not the same as sending it, and finding a payment problem is not the same as moving money. You need cost caps, because faster agents burn budget faster.

And you need a named human owner.

That bit is unfashionable, which usually means it matters.

Every serious AI workflow needs someone who can say: I own this loop. I know what it is allowed to do. I know what good looks like. I check the exceptions. I approve the risky steps. I change the procedure when it fails.

Without that, you get automation drift. Small mistakes get normalised. Weird shortcuts creep in. Reports look plausible but lose contact with the business. Everyone assumes someone else checked the source. Nobody did. Lovely.

The post-agency opportunity sits in exactly this gap

Old agency work was sold as output. Campaigns, posts, reports, landing pages, decks, dashboards. AI is pushing the cost of first drafts through the floor. That is not a moral crisis. It is just true.

The value now is the managed operating loop around the output.

A campaign loop that watches the market, spots real changes, drafts assets, checks claims, routes approval and learns from performance. A client-reporting loop that pulls the data, names the exceptions, attaches the evidence, drafts the narrative and asks a human before anything goes out. A sales follow-up loop that catches missed pipeline moments without spraying emails because someone got excited on a podcast. A content loop that turns source-backed research into an actual opinion instead of beige LinkedIn mush at scale.

That is not AI content. That is an operating system for commercial work.

And the person running it is not a prompt monkey. It is someone who understands the business well enough to decide what the machine should do, what it must prove, and where it must stop.

The question to take into Monday

Do not ask how to give everyone an AI assistant. Too broad, and it ends in forty logins and no accountability.

Ask this instead: which five recurring workflows could one strong operator supervise, if the intake, evidence, exception reporting and approval path were designed properly?

That question is uncomfortable, because it forces you to look at the business as a system. Where does work enter. Where does it get stuck. Where do people copy and paste between tools. Where do managers burn hours interpreting dashboards that should have produced a decision already. Where does client trust depend on somebody spotting the one weird number before the meeting.

Start there. Pick one workflow. Build the queue. Define the artefact. Attach the sources. Decide the approval rule. Keep the agent read-only at first. Watch the exceptions. Measure time saved, yes, but also measure correction rate, missing context, false confidence, cost and trust.

Then widen the span carefully.

Less glamorous than announcing you have an AI workforce. Much closer to how real businesses will absorb this without making a mess.

The next useful AI skill is not prompting. It is supervision.

The companies that learn it will not just move faster. They will know which machine work to trust, which to ignore, and which should never have been delegated in the first place.

That is the difference between leverage and chaos with a subscription plan.

Pull quotes

  • AI changes span of control. That is more useful, and more dangerous, than saying it saves time.
  • The agent gives leverage to the person who understands the work. It does not create judgement in someone who cannot tell good from bad.
  • A hidden agent doing mysterious helpful things in the background is not leverage. It is a liability with a friendly interface.
  • If one operator is supervising 50 workflows, those workflows have to be boringly legible.
  • The next useful AI skill is not prompting. It is supervision.

Sources

Editorial notes

  • The RingCentral “one product manager to roughly 50” claim is from OpenAI’s own launch page. Vendor-selected case study, not independent proof of a general productivity multiple. The blog says this out loud on purpose.
  • Codex task-duration figures are model-estimated from usage data, not stopwatch comparisons. Directional only. Not quoted as proof.
  • Anthropic’s expertise findings come from agentic coding sessions. Highly relevant to knowledge work, but the same ratios should not be assumed for marketing, sales or finance without testing.
  • Google Cloud’s 20-question post is product-led guidance. The operating questions are useful; the implied answer points at Google’s platform.
  • The piece deliberately avoids any headcount-replacement claim. The stronger and safer argument is span-of-control expansion under good workflow design.