AI Workslop: The Hidden Data Quality Risk for Enterprise AI

Unchecked AI drafts can become the information your next AI system relies on. Here's how ownership and source controls keep workslop out of enterprise decisions.

Alec Merle — Senior Forward Deployed Engineer, Foxtrot · 9 min read

You've probably received one recently. A status update or a vendor assessment arrives on time and runs longer than you expected, and somewhere around the third page you notice it never answers the question you asked. The thinking you expected to receive is now yours to do.

That's workslop, and inside an enterprise it is also becoming a data quality problem. Once an AI assistant reads an unchecked document, a claim nobody verified can become the basis for its next answer.

I see the writing side of this from the inside. AI drafting is part of how our engineers work now, and agentic tools have pushed the volume of writing up sharply over the past year. The same tools make a document faster to produce and easier to send before anyone has fully worked through it, and the gap usually shows when someone tries to act on the recommendation.

What is workslop, and what does it cost?

Researchers at BetterUp Labs and the Stanford Social Media Lab named the problem in Harvard Business Review in September 2025, defining workslop as "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task." In their online survey of 1,150 full-time U.S. desk workers, 40 percent had received workslop in the previous month.

Each incident took about two hours to resolve. The researchers put the cost at $186 per employee per month, or more than $9 million a year for a company of 10,000 people. Those figures measure cleanup time between coworkers, and the cost to enterprise AI arrives later, where it is much harder to see.

THE REPORTED COST OF WORKSLOP

Source: BetterUp Labs and Stanford Social Media Lab survey, September 2025

Why does workslop become a data quality problem?

When AI assistants connect to shared drives and wikis, a document gains another reader. An employee may recognize that a status report is incomplete, while an assistant can retrieve it as evidence of what happened without noticing what's missing.

Matthias Holweg and Thomas Davenport made the same point about whole organizations in Harvard Business Review in June 2026, warning of "decay in the accuracy and quality of organizational knowledge."

Consider an illustrative example. An engineer's notes say deployment testing is still underway, but an AI-generated summary calls the system "ready for deployment," and nobody corrects it. Weeks later, an assistant retrieves that summary when an executive asks which systems are ready to launch, and the unsupported statement now reads as evidence.

The failure starts well before the executive asks, at the moment an unchecked statement enters the company's shared record. Retrieval carries the mistake forward even when the assistant summarizes the document accurately, which is why decision-grade AI depends on the documents as much as on the model.

A document can become evidence long after its author stops thinking about it.

Why is delivery work so exposed to workslop?

In delivery work, writing a status report or an architecture recommendation is usually the moment an engineer has to work out what actually happened and what it means for the next decision. That working-out is the point.

Hand the reasoning to a model and only skim what comes back, and the document arrives with gaps its author can't answer questions about later. A client executive often makes decisions about their own system from that report.

Length can be a tell of its own. Someone who understands the system they're documenting can usually describe it in a few sentences, because they know which parts matter. A model defaults to breadth, explaining things nobody asked about and padding depth with coverage, so a document that keeps growing is often a sign the author let the model decide what needed to be said.

How can you spot workslop?

Style can be a clue, but the stronger signals show up in the substance of a document. Look for these four:

• Unsourced authority. "Studies have shown" or "industry best practice" appears without a source the reader can inspect. If there's a real source, name it, and if there isn't, cut the claim.
• Adjectives where evidence should be. Compare "significantly improved onboarding" with "cut onboarding time from six weeks to nine days." Only the second gives a reader something to act on.
• Assumptions in place of decisions. The draft stops to list the assumptions it made, where an author who understood the problem would have made the call and explained it.
• Manufactured conclusions. A closing paragraph restates what you just read without a decision or a next step, so it can go.

A useful report works the other way. It leads with status, and you can trace its recommendation back to the evidence and decisions behind it.

How do we keep workslop out of client work?

We wrote our standard down, because vague expectations are part of the cause. A January 2026 follow-up from the BetterUp and Stanford researchers traced workslop to unclear AI mandates and overwhelmed teams. The core rule is simple: the named author owns every sentence, whether they typed it or a model did, and rewrites anywhere it doesn't sound like something they'd actually say.

If you wouldn't defend a sentence's wording or claim in front of the client, it doesn't ship.

Review is calibrated to how a document will be used, because a chat message and an architecture recommendation shouldn't get the same scrutiny. The goal is for AI-assisted writing to get more intentional as it gets more common.

HOW WE REVIEW AI-ASSISTED WRITING

Scrutiny rises the moment a document is presented as someone's own thinking.

Stakes can also change with use. A quick email that becomes the basis for a contractual commitment deserves more scrutiny than its format suggests.

The step people skip is the first one. Read the whole draft once, start to finish, before editing anything, and notice where you stopped trusting it or stopped reading closely. A word-level cleanup leaves the reasoning gaps untouched, and every number and date still gets checked against its source, because a model's confidence is not a citation.

The last step is a read from the recipient's side. If a sentence would make you skim past it in someone else's document, rewrite it.

Skimming a draft before sending it moves the effort onto the reader, who now has to find the unsupported claims on their own time. When a reviewer spots a draft that reads as unedited AI, it goes back to its author to fix. A document that's factually correct but reads as obviously AI-generated has still failed.

What should happen before documents become AI sources?

Author ownership is the starting point, and an AI system also needs a way to tell working material apart from information the business has approved for use.

Start with one workflow your AI supports. Identify the documents it retrieves, then ask who owns each one and whether it's current and reliable enough for that decision. Then build those distinctions into the workflow:

• Make status explicit. Separate drafts and reviewed records from superseded versions. Where draft access is useful, carry that status into the context the model receives.
• Keep sources traceable. Let users open the document and version behind an answer. A source link helps people check the evidence, though it can't prove the evidence is right.
• Resolve conflicts. Decide which source governs when a report and an operational system disagree, and escalate unresolved conflicts so the model never picks one silently.
• Test the consequences. Check how the system handles an outdated document or a claim with no evidence behind it, and require human review wherever a wrong answer could trigger a consequential action.

Permissions decide who may access information, and review status and source authority decide what the system should rely on. A production AI workflow needs both.

How should teams use AI for writing?

Reading and editing a raw model draft is the minimum. Two habits produce better first drafts, because they keep the author's reasoning in the loop from the start.

The first is to dictate the thinking and let AI clean up the words. Talk through what happened and what you recommend before opening a drafting model at all, which is the same thinking-through-writing process a blank page forces, just faster. Then use AI for grammar and light restructuring. The claims stay yours, and the model's job is mechanical.

The second is a writing-style tool trained on your own past writing. I use one built from documents I wrote, so its drafts already sound closer to me and there's less to fix afterward. It still can't tell me whether a claim is true. Both habits front-load the author's own judgment instead of asking a model to guess at it, which is why they need less editing to reach a defensible final draft.

AI can contribute to the analysis itself, too, as long as people test its reasoning and own the recommendation. Sounding like the author and being work the author can defend are separate tests.

What should leaders ask of their teams and partners?

Two questions quickly reveal whether someone understands an AI-assisted document, and a third applies once AI is reading your documents.

Ask the same questions of partners and vendors, since their reports can end up in the systems your teams query.

There's also a simpler standard for documents nobody had time to read closely. If a draft isn't worth the author's own close read, send the reader the prompt or the raw notes and let them form their own view. Clayton Ramsey made this case in his May 2025 essay "I'd rather read the prompt", and the argument holds up in enterprise work. A prompt is honest about how much thinking went into it, and it costs the reader far less time than a polished draft they have to audit.

If a draft isn't worth the author's own close read, send the reader the prompt.

AI should speed up producing the words, while the thinking behind them stays the author's own. Clear writing and verified claims matter at delivery, and source ownership matters for as long as the information stays in use.

Frequently asked questions

What is workslop?

Workslop is AI-generated content that appears useful but lacks the substance to advance a task, so recipients have to fill in the missing reasoning or redo the work. Researchers at BetterUp Labs and the Stanford Social Media Lab introduced the term in Harvard Business Review in September 2025.

Is AI-assisted writing the problem?

AI-assisted writing can be very useful. The risk arises when people share output without checking its reasoning and accuracy, or let unchecked material become a source for later decisions.

How does workslop affect enterprise AI agents?

Agents that retrieve unchecked documents may carry unsupported claims into their answers or actions. Marking which sources are reviewed and current, and requiring human review where the stakes are high, both help reduce that risk.

What should a workslop policy include?

A policy should name an owner for every document and set review requirements according to consequence. It should also require verification of specific claims and say when a document is suitable for use as a shared record or an AI source.

Related reading

• The Agentic Enterprise: Moving AI from Assistance to Execution
• Can You Trust AI to Redact Sensitive Data?
• If You Had to Change Your AI Provider Tomorrow, What Would You Have to Rebuild?