How to know when AI is making your organization worse
An AI program can speed up every team's output and still degrade the organization. The warning signs show up at the receiving end of the work, not where it is produced.
BetterUp and the Stanford Social Media Lab surveyed US desk workers in 2025 and reported the results in Harvard Business Review. About 40% said they had received "workslop" in the previous month: polished-looking AI-generated work that does not move the task forward. Recipients put the cost at roughly two hours per instance. The survey is self-reported, and its write-ups disagree on the exact sample (about 1,000 to 1,150 respondents), so treat the numbers as an order of magnitude. It cannot tell us how much AI work is workslop. It does point at something a manager can use: the cost of a weak AI output is often paid by someone other than the person who produced it.
That asymmetry is the core of my argument. An organization can get worse because of AI while every producing team gets faster. Individual metrics improve. The damage lands on a colleague, a reviewer, a customer, or the person who will do this job in two years. If you only measure where the work is made, you will not see it.
My angle is narrow. Five AI metrics I would remove from an executive dashboard argues for better metrics. This essay is about where to look once a workflow is already live: at the receiving end.
Why the producer's dashboard cannot answer the question
Most AI measurement sits with the user: drafts generated, tasks completed, hours reported saved. Those are fair measures of activity and poor evidence of organizational health, for two reasons.
First, the producer is the person least exposed to the defects. Someone who sends an AI-assisted analysis saves the time of writing it. The recipient inherits the time of checking it, correcting it or quietly redoing it. Unless someone counts the second cost, the first looks like pure gain.
Second, people are unreliable judges of their own speed with a new tool. METR's 2025 trial of experienced open-source developers is the usual example, and I covered it in AI agents are exposing your org chart as fiction, so I will not repeat it. The relevant addition is METR's February 2026 follow-up, which calls its own newer data an unreliable signal, partly because developers increasingly refuse to work without AI. The lesson is not that AI slows people down or speeds them up. Felt productivity and measured outcomes can diverge, and measurement gets harder once a tool becomes the default.
Three drift signals
These are my working signals, not validated predictors. None proves harm alone. Together they help separate a healthy productivity gain from a transfer of cost.
1. Rework migrates downstream. The first sign is not that outputs are bad. It is that review, verification and correction move from the producer to the next person in the chain. Look for longer review queues, more "can you check this" messages, and reviewers who stopped trusting a category of document and now re-derive it. The workslop survey describes this pattern in perception data. In your own company you can measure it directly.
2. Unaided competence erodes. Some skills are maintained by doing the work. If a tool removes the practice, the skill may fade, and you will not notice until the tool is unavailable or wrong.
The clearest evidence I know of is from medicine, and it is an analogy, not a forecast for your finance or legal team. A multicentre observational study in The Lancet Gastroenterology & Hepatology (August 2025) followed 19 experienced endoscopists at four centres in Poland. Their adenoma detection rate on colonoscopies done without AI was 28.4% before regular AI use began and 22.4% afterward. The authors say factors other than AI may have contributed, and that results may not generalise beyond experienced clinicians. Two further limits matter for you. Medicine has a clear ground truth (a lesion was found or missed), and most office work does not. And this is one study in one specialty. I use it to show the mechanism is plausible, which is a reason to test for it, not evidence of how often it happens in your organization.
3. Perception outruns outcome. When people report large gains but end-to-end numbers (cycle time, defects that escape, customer recontacts, rework hours) have not moved, treat the gap as a finding. It may mean the gains are real but unmeasured. It may mean the gains are local and the costs are elsewhere. Do not assume either.
A four-week receiver check
This is a test I would propose before approving expansion of any live AI-assisted workflow. It is my design, not an established method, and it is deliberately small.
- Pick one boundary. Choose a place where AI-assisted work crosses from one person or team to another: a draft to a reviewer, an analysis to a decision-maker, a ticket summary to the next agent.
- Interview the receivers, not the producers. Ask how long it takes them to trust the output, what they now re-check, and what they no longer ask the producer to do. Log answers as time, not opinions.
- Sample outputs blind. Take a small, fixed sample of AI-assisted and non-AI-assisted outputs for the same task. Have a reviewer who does not know which is which score them against the standard the receiver actually needs.
- Run one unaided probe. Ask a few experienced people to do a representative task without the tool. You are not testing whether they are slower. You are checking whether they can still catch the tool's mistakes.
- Compare end to end. Put the producer's time saved next to the receiver's time added. If the second number is missing from the business case, the case is incomplete.
The decision it supports is concrete. If receiver cost is below the saving and the unaided probe is healthy, expand. If the saving is real but receiver cost is rising, hold and redesign the handoff. If quality or unaided competence is falling in a way you cannot explain, roll back part of the workflow while you find out why.
The strongest objection
The reasonable objection is that most evidence on AI points to real gains, and a company hunting for downsides will move slower than competitors who do not. That is a fair risk. A receiver check can be run so cautiously that it becomes a reason never to scale.
My answer is proportion. The check costs a few interviews and one sample. The failure it guards against is silent: costs moving to people who have no dashboard and no authority to complain. A related objection is that some rework is inevitable with any new tool. It is, so compare against the old process's rework, not against zero.
Two limits on my own claim. The evidence here is a perception survey, a narrow coding trial and one clinical study. None measures how often AI degrades a typical company. And the three signals are a way of noticing, not predictors. If your check finds nothing, that is a useful result: you will have evidence to scale with.
Before the next expansion decision, pick the receiver with the least say in the rollout and ask what changed in their week.
Where this connects
- The hidden metric in AI automation is supervision covers the cost of supervising an automated step; this essay looks at the cost that lands on the next person.
- Five AI metrics I would remove from an executive dashboard is the metrics-side companion.
- AI agents are exposing your org chart as fiction traces decision routes; the receiver check traces cost routes.
- Your AI roadmap is probably missing the only column that matters sets decision rules before launch; this is the monitoring step after it.
Sources
url: https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity title: AI-generated "workslop" is destroying productivity publisher: Harvard Business Review date: 2025-09
url: https://www.betterup.com/blog/hidden-costs-workslop title: The hidden costs of workslop publisher: BetterUp date: 2025-09
url: https://wrap.warwick.ac.uk/id/eprint/191005 title: Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study publisher: The Lancet Gastroenterology & Hepatology (record: University of Warwick repository) date: 2025-08