Authorship Debt: When AI Finishes Your Work

For two years, AI at work meant a faster first draft. The model gave you a rough version, you did the thinking, and the result was plainly yours. Agentic tools ended that arrangement. Microsoft’s Copilot Cowork, generally available since June 2026, is built to run a task end to end and, in the company’s own words, “return a completed result, not just a draft or a recommendation.” OpenAI and Google ship the same promise. The output no longer reads like a starting point. It reads as done.

That sounds like pure upside, and most of the time it is. But it introduces a liability few professionals have priced in. When a tool hands you finished work, your bottleneck stops being production and becomes accountability. The deliverable goes out under your name, not the model’s, and finished-looking work is exactly the kind people stop checking.

Call it authorship debt: the distance between the work you put your name on and the work you can actually defend. Every deliverable an agent hands you arrives carrying some. Like technical debt, it stays invisible until it comes due, usually in the meeting where someone asks where a number came from and you find you can’t say.

Why the debt stays hidden

Polish suppresses scrutiny. Researchers at BetterUp Labs and Stanford’s Social Media Lab, writing in Harvard Business Review, named the neighboring problem “workslop”: AI-generated content that “masquerades as good work but lacks the substance to meaningfully advance a given task.” In their survey, 42% of people who received workslop trusted the sender less afterward, and they put the drag at roughly $186 per employee per month. It is a self-reported study and worth reading with that caveat, but the mechanism it describes is the one that matters here: a rough draft announces that it needs work, while a polished artifact asks to be forwarded. Surface credibility is what carries the weakness downstream.

So the danger was never low quality. You would catch low quality. It is high polish sitting on low provenance: output you cannot trace to a source or a decision, wearing the costume of output you can. An agent manufactures exactly that. It reads across files, apps, and systems you never watched, then returns a confident result, and where it lacks a fact it does not leave a blank. It fills the gap in the same steady voice it uses for everything else. That is why the exposure is identical up and down the ladder. An analyst forwards a model to their manager; a manager forwards a summary to a VP; an executive signs a board memo stitched from a dozen systems. The more finished it looks, the less anyone inspects it, and the more debt rides along to the room where it gets challenged.

The Provenance Pass

You pay authorship debt down the same way every time, with a short pass before you forward. Three moves, about ten minutes, and they scale from a junior’s spreadsheet to a director’s strategy deck.

Verify the load-bearing claims. Pick the two or three figures a decision actually rests on, recompute them from the source, and confirm each traces to a real file. This is ordinary deck hygiene, and the mechanics are their own skill, covered in catching the hallucinated number before it hits the deck. Agents raise the stakes only in volume: one file to spot-check has become forty.

Map what it couldn’t see. This is the move an agent makes newly necessary, and it has no substitute. Name one thing that shaped this period but isn’t in the files the tool read: the deal that slipped, the budget you moved mid-quarter, the client on pause, the context that lived in a hallway conversation. The agent reports what it can see and treats everything else as silence. It cannot tell the difference between “this didn’t happen” and “I had no access to this.” You can, and filling that silence is the part of the job that is unambiguously yours.

Own the judgments, then run the forward test. An agent can say what moved; it cannot know why, and it will guess a cause in the same calm tone it uses for facts, which is the trap examined in writing the story around your numbers. Every sentence that assigns a cause or recommends an action is yours to write or cut. Then, for each claim left standing, ask the only question that counts: could you defend it if someone senior pushed back in the room? If not, it is not ready to carry your name.

There is a fair objection. Agentic tools bill by the run, and Copilot Cowork meters model use, context, and runtime against usage credits, so re-running it to fix one section feels like paying twice. That is an argument for the pass, not against it. Ten minutes of your own reading is cheaper than a second run, and far cheaper than retracting a number in front of your boss next week.

The useful instinct is to treat the agent as a sharp junior colleague, with one correction. You would never forward a junior’s analysis to the board unread, not because you distrust them but because your name turns their work into your claim. The agent is the same, minus the part where the junior learns and needs less checking next quarter. It does not accumulate judgment. You do. Let it keep the production, keep the authorship for yourself, and clear the debt before you hit send. That last step is the assign-the-work, keep-the-decision discipline behind the four ways to work with AI, and with an agent it is the whole job.