Workslop: The Time AI Saves You Gets Spent by Someone Else

The failure case of lazy work is not lack of output. It’s actually over-output.

Tobi Lütke said that on The Knowledge Project on September 15, 2026. Inside Shopify, they call the things people lob at each other “slop grenades”. It’s a useful name for the document that takes seconds to send and an afternoon to untangle.

I make money helping businesses adopt AI.

I also think much of the productivity people report hasn’t disappeared. It’s landed on someone else’s desk as work. I’ve written about what AI does to jobs. This is about what happens between the person who produces something and the person who has to use it.

A Slop Grenade Leaves You to Find the Point

A slop grenade is unchecked AI output sent to a teammate or client. It uses a lot of words to say something simple, looks deep without offering much substance, buries the useful information and leans on weasel words to sound important. The receiver has to extract an answer before they can do their own work.

Stefan Neubig’s noslopgrenade.com, inspired by nohello.net, gives a good example. Someone asks, “Should we use Redis or Memcached?” A minute later, they get about 270 words, including:

Great question! The choice between Redis and Memcached is a nuanced decision that requires careful consideration of multiple factors. … I’d recommend conducting a proof of concept with your actual workload patterns to make an informed decision.

His alternative:

Redis. We need pub/sub for the notifications feature.

The short reply gives the decision and the reason. The long reply hands the decision back, with homework attached.

This can arrive as AI mockups of pages and visuals, charts nobody can read without the sender explaining them, or off-brand images and video that need an expert to review. Slide decks can need more than 30 minutes of formatting fixes before anyone gets to the argument. Any deliverable can become a slop grenade when it transfers unfinished thinking to the next person.

BetterUp Labs and Stanford’s Social Media Lab called this workslop in September 2025: AI-generated work that looks good but lacks substance. In BetterUp’s survey of 1,150 full-time US desk workers, 40 per cent had received it in the previous month, and each instance took an average of one hour and 56 minutes to sort out.

The receiver pays twice. Alongside the cleanup, there’s the damage to the relationship. Fifty-three per cent of recipients were annoyed, 42 per cent trusted the sender less, and about a third would rather not work with them again. Those are reported reactions from a survey, but they’re a fairly clear warning about what happens when colleagues become your unpaid editors.

Time Saved Can Become Time Displaced

Workday surveyed 3,200 full-time AI users at large companies and found that 85 per cent saved between one and seven hours a week. I believe it. Getting a usable draft in twenty minutes instead of an afternoon is useful.

But nearly 40 per cent of those saved hours went back into rework, and only 14 per cent said they consistently came out ahead. The same people who reported the savings also reported fixing mistakes and checking what the tool produced.

The survey follows the person using AI. It can’t account for all the work that person sends downstream. Every hour here is self-reported, and the rework estimate covers the respondent’s own cleanup.

Picture a slop grenade that saves the sender 30 minutes and creates 30 minutes of work for someone else. This is an illustration, not a survey result. Even a straight swap like that shows up as a gain because only the sender’s half gets measured.

The saving goes on one timesheet. The cleanup goes on another, perhaps in a different department under a different manager. If the first team reports faster delivery and the second reports being busy, the organisation can miss that they’re describing the same document.

Polish Hides Where the Thinking Stopped

A rough human draft usually looks rough. An uncertain section trails off, a note asks for a source, or the writer tells you which part needs help. You can direct your attention to those places.

An AI draft can sound confident and complete when its argument is generic, its claims are unsupported or its context is wrong. An invented citation looks as carefully formatted as a real one. The reviewer has to find the gaps without any help from the surface.

GoTo and Workplace Intelligence asked 2,500 workers and IT leaders about this: 77 per cent said AI-generated work took longer to review than human work, and 66 per cent said reviewing other people’s workslop created extra work. The same survey found 43 per cent had used AI content despite suspecting it was low quality or contained something made up.

That last finding worries me. It suggests the suspicion can survive right up to the send button.

G.K. Chesterton wrote, “If a thing is worth doing, it is worth doing badly.” In What’s Wrong with the World, he was defending amateurs doing things themselves. An honest rough draft deserves some of that generosity: it gives a colleague something to work with and makes its limits visible.

Clean formatting still matters. We turn paragraphs into bullet points and numbers into charts so people can understand them. When making a deck, stop asking how to cram more into it and start asking how to organise complex information so someone else can follow it.

Slop Starts When the Task Beats the Goal

“Publish four blog posts” is a task. Building a reputation or generating pipeline gives those posts a reason to exist. This post is on my own list too, and I could satisfy that task with a long, well-sourced article that nobody finds useful enough to change anything.

The agency team I work with has plenty of tasks where that distinction matters:

• Someone preparing a pitch for a prospect needs to show that we understand their challenge and have the experience to solve it. Not the goal: an à la carte menu of services or a celebration of our team culture.
• Someone building a campaign landing page needs to get the people who clicked the ad to fill out the form by explaining what the offer is, who it’s for and what happens next. Not the goal: catchy headlines, cross-promoting other offers or a catch-all page for every visitor.
• Someone asking why a client’s site is down needs the team fixing it and an update for the client with next steps and a timeline. Not the goal: an exhaustive list of possible causes.
• Someone writing a conversion report needs to identify the changes most likely to improve conversions and get agreement to try them. Not the goal: demonstrating diligence by listing every possible optimisation.

Ask a model for any of those deliverables without explaining the goal and you can get the full service menu, every possible cause or every option. Exhaustive is its default, and nothing in the request told it which decision the reader needed to make.

Before drafting, try this question: “If I only had 30 seconds to speak, I should communicate…” Finish that sentence.

Nobody Is Measuring the Receiving End

Goldman Sachs reviewed fourth-quarter earnings calls and found that 70 per cent of S&P 500 management teams mentioned AI, 54 per cent framed it as productivity or efficiency, 10 per cent quantified a specific use case and 1 per cent quantified an earnings effect. There is a lot more talk about productivity than accounting for it.

Goldman’s senior US economist Ronnie Walker said: “We still do not find a meaningful relationship between productivity and AI adoption at the economy-wide level.”

The companies reporting quantified gains had a median improvement of about 30 per cent, in customer support and software development. Those areas have machine-checkable outcomes: code can run through tests, and a support system can record a ticket closing or reopening. Neither guarantees good work, but both give you a check beyond whether a document sounds convincing.

I have to hedge this, because Goldman does. Adoption is still early by their own tracker, and economy-wide gains may still arrive. These figures don’t prove that displaced work explains the missing relationship, but they leave room for costs that a count of completed drafts would miss.

Deloitte provides a less abstract warning. AP reported on October 7, 2025 that Deloitte Australia agreed to partially refund the AU$440,000 the Australian government paid for a report about automated welfare penalties after Sydney University researcher Chris Rudge found a fabricated federal court quote and references to nonexistent research papers. Deloitte agreed to repay the final instalment, whose amount wasn’t disclosed at the time.

The revised report disclosed Azure OpenAI use in writing it. The department said the recommendations hadn’t changed, which limits what we can claim about the consequences. The checking fell to an outside academic after the report was already published.

When you sell expertise, the document is part of the service. An agency demonstrates its judgement in proposals and emails as well as finished campaigns. If every person pastes a different kind of unchecked output into those documents, the inconsistency becomes part of what clients are buying.

The Second Reader Stops Checking

The first polished document gets a careful check. After a few that seem fine, the reader starts skimming, especially when they have their own work waiting. An error can survive simply because the person expected to catch it has learned to trust the format.

Matthias Holweg and Thomas Davenport describe knowledge decay: errors enter a process, survive review and become material for the next piece of work. A mistaken claim in a report can end up in the wiki, where the next writer treats it as something the organisation already knows.

The junior who used to write the draft also knew which parts they didn’t understand. As I argued in the post about entry-level work, automating that task affects the training pipeline. It also removes someone who could tell the senior reviewer where to look.

I Send This Stuff Too

This would be a dishonest post if it stopped at other people. I use these tools every day, on my own work and on client work, and I’ve handed people workslop.

In the old war-film move, someone throws themselves on a grenade to spare everyone else. At work, the receiver gets volunteered: they find the useful point, check it and rewrite the message the sender could have sent.

You Own What You Forward

If you can’t explain a paragraph, you can’t send it. You should be able to defend its claims to the person receiving it without reopening the tool to find out what it meant.

Whoever forwards the work owns it, including the sentences they didn’t read. Disclosure helps, but “AI wrote this” doesn’t give the recipient any help deciding which parts are safe to use.

Give them a specific request: “I’ve checked the figures. Can you check whether this recommendation fits the client’s setup?” That tells an expert where their judgement is needed. “Please review everything” hands them your unfinished first pass.

For a team, the practical measure is rework: time spent checking and repairing what arrived from elsewhere. Count that alongside the time saved producing drafts, so the receiving team’s hours have somewhere to appear.

A Thoughtful First Pass Is Enough for Monday

The goal is a thoughtful first pass. Nobody needs to become a writer or designer before asking for help, and perfection would make a fairly effective excuse for never sending anything.

If you’re sending work:

• Put the most important takeaway first and trim the wall of text. AI can help shorten it.
• Use the templates and fix obvious spacing and font problems before handing it over.
• Say what AI wrote and what you checked. Flag the specific parts that need an expert’s attention.
• Ask when you’re unsure, while the question is still small enough to answer.

If you run a team:

• Count rework hours for a month, including time spent fixing or verifying a colleague’s output.
• Put the ownership rule in writing: whoever forwards it owns it.
• Stop rewarding length. Ask whether the receiver could use the work to make the decision or take the next step.

For the next document you send, write the takeaway at the top and name the part you need help with. Give the person opening it somewhere to start.

If you’re trying to work out where AI fits in your business, and where it doesn’t, that’s the work I do. Start with the workflow, not the tool.