How to Audit AI-Assisted Work for Slop

Every new office technology has a period in which output rises before quality systems catch up. Email multiplied communication well before firms worked out how much of it was defensive noise. Slideware gave unfinished strategy a finished look, and dashboards lent weak measures an air of authority. Generative AI is doing the same to knowledge work.

The word for the failure mode is unpleasant: slop. It means output that looks like work, moves through the organization like work and consumes review time like work, yet is thin, generic, inaccurate, stale, unowned or beside the point of the real decision. A factually correct draft can still fail as work. A plausible one often survives the first glance, which is how it gets through.

The Society for Human Resource Management (SHRM) puts numbers on this in its 2026 workplace research. It reports that 41% of workers use AI in their work, and that just under half of those users describe their own output as "AI slop." The label is less interesting than the confession behind it: many workers know that some of what AI helps them produce falls below the standard their organization would want if anyone stopped to check.

The risk is work shaped to pass review, produced faster than managers can judge it.

Where slop sits in the workflow

Executives tend to discuss AI quality in two terms: model capability and human review. Better models remove some defects, and reviewers catch others. Both claims are partly true, yet together they fall short of a control system.

Slop appears in the gap between generation and decision. A fluent draft can be strategically empty. A tidy customer summary can miss the real objection, and a market memo can cite a respectable source while drawing a conclusion the source does not support. Performance reviews turn a hard personnel signal into polished mush, and project plans sound decisive while hiding dependencies, owners and dates. Anyone who has reviewed AI-assisted work at speed knows the sequence. The document arrives looking complete; the first minute goes on its structure, the next five on hunting for substance, and fifteen more on reconstructing the judgment that should have existed before anyone drafted it.

Why adoption metrics mislead

Stanford's Human-Centered Artificial Intelligence (HAI) institute reports in its 2026 AI Index that organizational AI adoption kept rising. It also finds that agent use is still early, and that productivity gains are strongest in structured, measurable work where outputs are easy to monitor. The caveat deserves more attention than the headline. Gains arrive when a task has a clear unit of output, a visible quality bar and fast feedback. Support tickets, code tasks, document operations and marketing variants can be measured that way. Strategy, negotiation, hiring judgment, regulatory interpretation and board advice cannot.

Executives who celebrate adoption without measuring the review burden are counting volume as operating gain. The company may produce more drafts, summaries and options while labor shifts from creation to inspection and from junior staff to managers, and some of it slips out of visible hours into a judgment debt that nobody books. Microsoft's 2026 Work Trend Index approaches the same gap from another side: many workers are moving faster than the systems around them. Leaders, it argues, need to rearchitect work, including the definition of finished work, which handing out more tools does not do on its own.

How an AI slop audit works

The audit is a sample-based quality review of AI-assisted work, aimed at the places where low-grade output has become an operating cost. It should show where AI improves throughput and where it creates a new layer of managerial cleanup.

What slop costs managers

Slop is expensive because it hides inside output that looks cheap. An AI-generated brief can take six minutes to produce and forty minutes to repair. A summary that saves an analyst time can cost a partner's trust, and a sales note can speed up the reply while flattening the customer's actual concern. Policy drafts look complete until legal finds three jurisdictions merged into one confident paragraph.

The cost is rarely booked. It turns up as meetings that drift, rework, vague feedback, tired managers, slower promotion readiness and a gradual decline in the standard of internal writing. The firm still feels busy, perhaps even more modern, while its work becomes harder to trust. So the audit should count accepted output before generated volume, weigh decision quality above document length, and track reviewer minutes alongside user satisfaction. How fast employees feel is a weak primary measure next to whether decisions are defensible and total waste has come down.

Who owns AI-assisted output

AI-assisted work needs a simple ownership rule: the worker may use AI and still owns the output. Drafts that make factual claims need a source trail. Workers should mark where AI was used, whether for synthesis, wording, calculation, code, research or decision support, so that reviewers know whether they are checking prose, facts, judgment, compliance or customer risk.

Without that rule, the model works as an unacknowledged junior employee with no training record. An organization that would stop an intern sending unsupported analysis into a board pack should stop a model-assisted draft on the same path, however much better its grammar. The rule should be light for low-risk work and strict where the consequences are serious. Brainstorming notes need no forensic record, but customer renewal recommendations do. Social captions can tolerate an ambiguity that an HR investigation summary cannot, and code comments need less control than production changes.

Where executives should start

Pick five workflows where AI-assisted output is now common, such as customer communication, internal analysis, performance management or software delivery. Pull a sample of recent AI-assisted work from each. Ignore tool-satisfaction scores and measure four things:

1. What share of generated output was accepted without major rework?
2. Which defects appeared most often: accuracy, specificity, judgment, tone, compliance or missing context?
3. How many review minutes did each accepted piece of work need?
4. What rule would have kept the worst output from reaching review?

Then change the workflow. Add source trails where facts matter and explicit human decision points where judgment matters. Use templates only when they make the work more specific. Set standards for managers before buying more seats, and reward the employees who use AI to improve the work over those who use it to polish more volume.

A well-run AI organization should be able to show, with evidence, which workflows AI has made faster, which it has made better, which it has made riskier, and which it should stay out of for now. Generated volume cannot answer any of those questions, so the sampled findings and the priced review burden should reach leadership before anyone sets a volume target.