Workslop has a price, and the reader pays it

Six pages arrive in your inbox at 4pm. Headings, a summary table, three recommendations in bold. It looks like a week of work. By page two you notice it never names a customer, a number from your own books, or a decision anyone has to make. It could have been written about any company in your sector. Now you have a choice: send it back and look difficult, or rewrite it yourself and lose your evening.

There is a name for this now. Researchers at BetterUp Labs and the Stanford Social Media Lab called it "workslop" in a piece for Harvard Business Review in September 2025: AI-generated work that looks like good work and does not move the task forward.

Their survey of 1,150 US desk workers put some edges on it. About 40% had received workslop in the previous month. It made up roughly 15% of the work they received. Each instance took close to two hours to deal with. And about half of the people who received it rated the sender as less capable and less reliable afterwards.

Read that last figure from the sender's side. The person who saved an hour drafting has spent some of their reputation, and the person who received it has paid the hours.

Where the cost hides

Workslop does not show up as a line item. It shows up as the senior person who rewrites junior drafts at night, the client email that needs three rounds, the report that gets politely ignored. A business owner looking at AI adoption sees more documents produced per week. The same owner rarely sees the hours spent repairing them, because those hours are spread across the people who receive the work.

That is why so many teams report that AI makes them faster while their managers feel busier.

Why AI output goes generic

A language model asked to write about your business with nothing but the request in front of it fills the gaps with whatever is most likely. Most likely means average. It has never seen your price list, your last three client complaints, the phrase your founder hates, or the reason you stopped offering the cheaper service tier. So it writes a plausible average company, in a confident format, and the format hides the emptiness.

Better prompting helps a little. It does not survive staff turnover, and it does not survive a busy Friday when someone skips the long prompt.

What changes when the AI reads a company memory first

We run our own AI work from a written company memory. Before a session drafts anything, it reads the files that hold our rules, our verified facts and the corrections we have made in the past. A few things follow from that.

• Drafts carry real specifics. If a fact is in the memory, the draft can use it. If it is not, the draft has nothing to invent from, and the reviewer can see the gap.
• Communication rules apply to everyone. How we adapt tone to the reader, what never appears in public material, how our names are spelled. One person's correction becomes the next person's default.

If you build your own, add approved examples next to the rules. A model shown two real proposals you were happy with writes closer to them than one shown a description of "our style."

Then there is a gate. A person reads proposals and client documents before they leave the company, and the reviewer's job is to ask one question: does this say something only we could say?

What it does not fix

Memory supplies facts and voice. It does not supply a point of view. Most workslop is a substance problem dressed as a writing problem. If the sender had nothing to say, a well-informed AI will produce a well-informed nothing. That part still needs a person who has thought about the problem.

Memory also goes stale. A price list in the memory that nobody updated produces confident, specific and wrong drafts, which are worse than generic ones.

What the sender can do

Most of this post is about the receiver's cost, but the fix starts with the sender. Three habits help, with or without a company memory.

First, write the one-sentence answer before asking the AI for anything. If you cannot say what the document should conclude, the AI cannot either, and it will produce length instead.

Second, cut the draft by a third before sending. AI drafts run long because length is cheap for the model. The reader pays for every extra paragraph in minutes.

Third, add one thing only you know. A figure from last month, a comment the client made on the phone, a judgment about what to do next. If you cannot add anything, ask yourself whether the document needed to exist.

None of these are clever. They are what a good manager would say to a junior who handed in a padded report, which is roughly what an unreviewed AI draft is.

Try this for one week

Keep a tally of every document you receive that you had to redo or send back. Note roughly how long each one took. At the end of the week, sort them by what was missing: facts the sender did not have, or thinking the sender did not do. The first pile is the one a company memory shortens. The second pile is a conversation with your team.

Book 20 minutes at pgco.world/book.