AI workslop is a handoff problem. It is AI-generated work that looks finished but leaves the next person to discover missing facts, make the decision or rewrite the result. Keep one team responsible from request to accepted outcome, with quality standards checked before handoff.

In a 2025 survey of 1,150 US desk workers, BetterUp Labs and Stanford Social Media Lab found that 40% reported receiving workslop in a month; 53% admitted sometimes sending it. Respondents estimated almost two hours of rework per incident. The $9 million annual cost for a 10,000-person company is modeled from those reports, not audited. Thomas Otter’s critique asks a fair question: how much sloppy work existed before AI, and has AI increased it or merely made it look more polished? The survey cannot tell.

The sender saves time. The recipient pays

A person asks an agent to draft a proposal, gets ten tidy pages, and sends them to another department. The recipient must find the recommendation and check the figures. The sender is faster; the company is not.

Shopify CEO Tobi Lütke’s “slop grenade” example describes exactly that failure with unreviewed code and long emails. It is an anecdote, not a measured estimate of its prevalence. Producing text is easy; accepting responsibility is harder. Even an end-to-end owner can believe a plausible answer.

Give a team the whole outcome, not a longer queue

In my transition guide for an existing company, product and service builders work in small, accountable teams. They can draw on agents and specialists, but one named human owns the result and its exceptions. That structure reduces the number of places where someone can say, “I made my part; quality is the next team’s problem.” A bad draft now comes back to its author’s own outcome.

Boundaries remain. Legal, security and service still review what crosses theirs. Map handoffs and returned deliverables; compare rework and customer errors before and after. Fewer internal returns with more customer errors would be a failure.

Managers manage standards, not every output

The second layer is a standard that can be checked. Managers and experts should publish good examples, required facts and human-review conditions. The same pattern can cover code, design, brand language or service. Put criteria in a versioned skill or test, not a manager’s memory.

This acceptance contract is my response to Shopify’s unreviewed handoffs, not a tested deployment. The service owner supplies the approved policy and good example; a check agent runs before handoff.

Check before handoffEvidence to returnOn failure
Customer request is restated accuratelyQuoted request and matching answer spanReturn to the author
Each policy claim matches the approved versionPolicy ID, version and cited paragraphReturn to the author
No unapproved price, refund or delivery promiseExact promise and approval record, if anyHuman review before sending
Tone follows an approved good exampleExample ID and the passage that differsReturn to the author

The audit agent returns pass, fail or uncertain with evidence. It cannot send or approve exceptions. The author fixes failures; the manager updates the standard. Legal and commercial calls stay human.

Start with one costly handoff. Track author and reviewer time, returns and accepted outcomes. A gate producing another polished report is workslop with a badge.

A BCG field experiment first released in 2023 found AI helped on tasks within its capabilities but made people less accurate on one task outside them. Test the result and its verification cost, not the output count.

A prediction to test

If teams adopt end-to-end ownership and pre-handoff checks, returns for missing evidence should fall at their boundaries without a rise in customer errors. If only draft volume improves, the work has moved, not disappeared. Cost per agent outcome is a better measure than tokens or drafts.

Further reading