Work in Superposition | Work Is Code

AI allows work to exist in a superposition state, where it can be rendered in real time into whatever shape it needs to take to be consumable.

That sentence came out of me mid-conversation months ago, and I have been testing it since. On October 27 I gave a talk to a room of project managers about the most ordinary version of it: the status update you rewrite five times a week. This is the longer argument behind that talk.

The Status You Already Rewrite

One project. Five audiences. The executive wants the headline. Finance wants the number. The client wants the milestones. The team wants the detail. The PMO wants its template.

So you write it five times. Asana found that 53% of knowledge workers’ time goes to work about work, and chasing the status of tasks is named in the definition.¹ For a PM, that is most of the week.

The time is the visible cost. The invisible one is drift. Go-live moves from November 20 to December 4. The exec email says “minor schedule adjustment.” The client hears “early December.” Finance hears “no budget impact.” The standup hears “we slipped two weeks.” The PMO template says Amber. Nobody lied. Five honest versions of one fact tell five different stories.

Knowledge Lock, One Move Further

In The Knowledge Lock I argued that every organization runs an invisible permissions system enforced by format. Engineers read the repo. Finance reads the model. Executives read the deck. The PM ferries facts across each border by hand. AI breaks the lock because it can translate between formats.

That framing still assumes the formats are fixed. Repo on one side, board summary on the other, translation in between.

Superposition is the next move. The work doesn’t have a shape at all until someone needs it. The kanban board, the exec summary, the steering deck, the client email, the daily note: none of those is the work. Each is a rendering of the same underlying source for one reader at one moment.

When rendering was expensive, because a human wrote every version by hand, organizations rationed renderings. They picked one canonical shape per audience and called it the truth. The repo was the engineer’s truth. The deck was the executive’s truth. The status report was the PMO’s truth. Drift between them was the tax everyone paid.

When rendering costs almost nothing, the canonical shape stops carrying weight. The truth is the source. Every shape is a view onto it.

Free Rendering Is Not the Point

Here is where most teams went wrong with AI. Rendering got cheap, so they rendered more. Same workflow, more output. Every status pack got longer. Every summary forwarded everything.

That clogs the filter. An organization is a stack of filters: each layer absorbs detail and sends up what the next layer needs to decide. AI made it effortless to forward all forty items instead of the one decision. It feels thorough. It pushes the choice up a level, to the people with the least context and the least time. Forty equal items is cognitive load for the reader, who now has to do the filtering you skipped.

It also builds cognitive debt for the sender. Polished packages go up the org from people who don’t understand them well enough to say what matters. An MIT Media Lab study found that people who wrote essays with ChatGPT struggled to accurately quote what they had just written.² It is a small, early study about essays, not status reports, but it names the gap: the package is cheap, and the understanding isn’t.

The teams getting real returns did something else. In Wharton and GBK Collective’s fourth annual enterprise AI study, 84% of organizations reporting strong returns from AI had significantly changed their workflows, against 69% of everyone else.³ They redesigned the work. They didn’t just speed up the old version of it.

Superposition is that redesign. Hold everything. Serve what’s needed.

What Changes When You Hold the Source

There is no canonical shape. Treating any one rendering as the truth is a category error. The PMO template is not the project. It is one more audience, and you can keep filling it without thinking inside it.

Renderings are disposable. The source is durable. The view is a cache. When it goes stale, throw it away and render a new one.

The hard part stops being translation. Any shape is cheap now. Deciding which shape, for whom, at what level of detail is the work that stays. Ben Shneiderman put the design rule in one line thirty years ago: overview first, zoom and filter, then details on demand.⁴ The executive gets the decision. The rest is one ask away, and they should feel it is there.

A single status field is a lie about a superposed system. This one took me longest to see. The truth is often complicated and keeping multiple complicated projects entirely in your head is not something we are wired for.

Status Is Not One Thing

Every status report I’ve written asks for one color. Red, amber, green. Pick one.

Work doesn’t sit in one place, though. Any piece of it is in several at once. Has the question been answered? Has someone built it? Is there evidence it works? Has the person who owns it signed off? Four questions, four different answers, and the color pretends they’re one.

I saw this on a requirements tracker for a system we were building. It read “12 verified, 455 open.” Anyone looking at it would say the project was in trouble. It wasn’t. When we split “open” apart, 12 more items were built and checked against their acceptance criteria, waiting on sign-off. Another 253 were in build. Only 190 were actually open. The tracker had one bucket for “not done,” so work underway looked exactly like work nobody had started. Every number was accurate and the picture was wrong.

More data didn’t fix it. Keeping the states separate in the source did, and then giving each reader the part they needed. Leadership got “Built. Waiting on your approval,” a decision they could make in a minute. Engineering got the acceptance criteria and the test evidence. An auditor could follow the whole chain.

The talk covered half of this idea: work doesn’t have one shape. This is the other half. It doesn’t have one status either. The source holds all of it, and each update hands one reader the slice they need to act.

Ask, Answer, Prove

In the talk I framed the states as a flow PMs already know. Ask the right questions of the right people. Answer them well enough to decide. Prove it with acceptance criteria.

That flow is judgment, and judgment is the answer to cognitive debt. AI can carry the volume. Understanding still means leaning in enough to render your true professional judgment on the problem. That is where the human value in an organization lies. People are afraid of being replaced with AI, but then double down on copying AI instead of recognizing why they were hired in the first place.

Rules for Tools That Follow

If you accept the framing, some design rules fall out of it.

Don’t be the source of truth. Render from the durable source and point back to it. Tools that demand to be the source accumulate drift between themselves and reality.
Make the rendering disposable. State lives in the source, not the view.
Give every rendering a path back. A comment, a correction, a decision made in the meeting should flow back into the source, or the next rendering repeats the mistake.
Optimize for the moment of consumption. Different moment, different rendering, same source.
Give AI a process, not a pile. A source, a template, and rules it has to follow: use only what’s in the source, mark what’s missing instead of guessing, and a human reads and signs before it goes. A pile is how you get slop.

You don’t need new software for any of this. One source in a document you control, an AI tool your organization approves, and the discipline to sign what goes out. The prompt kit from the talk is a working starting point.

The Lock That Remains

Same ending as The Knowledge Lock, because it is still true. AI can render anything into any shape. Knowing which shape matters, for which reader, at which moment, is editorial work, and it stays human.

The point isn’t that the PM goes away. The PM’s job moves from translator to editor: what gets seen, by whom, and when.

Asana. State of Work Innovation. 2024. ↑
Kosmyna, N., et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, 2025. ↑
Wharton Human-AI Research and GBK Collective, fourth annual enterprise AI study, as reported in WSJ CIO Journal, The Morning Download, October 13, 2026. ↑
Shneiderman, B. The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations. 1996. ↑