AI Slop Lands in the Manager Review File
On July 5, 2026, SHRM put a blunt phrase into the workplace AI file.
In its report, Navigating AI in the Workplace: 2026, 41% of U.S.-based workers said they use AI for work. Among those users, 44% said their output includes “AI slop.” SHRM did not define the term as a technical defect. It was closer to a worker’s admission that AI-assisted output can look finished while still being low quality, generic, incorrect, misaligned or hard for someone else to trust.
That makes it a manager problem.
AI slop can be bad writing on the public internet. Inside a company, it can be a sales note that sounds confident but misses the customer’s actual objection. It can be a performance-summary draft that turns a nuanced coaching issue into bland praise. It can be an internal analysis that cites the wrong policy, a support reply that answers the wrong question, a project plan that hides dependencies, or a recruiting summary that compresses a candidate into a sentence the hiring manager cannot defend.
The old review process was built for human work. A manager could ask who wrote the memo, which data source supported the chart, which colleague reviewed the account note, and why a recommendation changed. AI changes that chain. The person still owns the work, but the draft may have been shaped by a model, a prompt, a plug-in, a summarizer, an agent or an enterprise assistant connected to private files.
The manager has to review the output without turning every employee into a suspect.
That is harder than a policy memo makes it sound. SHRM’s report found the pressure is not evenly distributed. Entry-level and early-career workers reported the highest pressure to use AI tools, at 45%. Directors and above used AI more than other workers: 63% of them reported using AI, compared with 50% of managers and 34% of individual contributors. Senior leaders also carried the more worrying ethics signal. When an AI recommendation conflicted with what they believed was the most ethical choice, 55% of directors and above said they would follow the AI’s advice.
The company cannot treat AI slop as a junior-worker problem. Leaders are using the tools more. Managers are expected to catch the quality issues. Employees are trying to learn what counts as good work while the standard is moving under them.
The review file now needs to answer four questions: what work used AI, what quality risk the AI introduced, what evidence supports the final output, and what the manager should coach rather than punish.
July 5 put bad output into the open
The first workplace AI wave was measured by adoption. The second is being measured by friction.
Gallup reported in April that, for the first time in its measurement, half of employed American adults used AI in their role at least a few times a year in the first quarter of 2026. Daily use reached 13%. Daily or weekly use reached 28%. Within organizations that had adopted AI, 65% of employees said it had improved productivity and efficiency.
That should have been the easy story: more people use AI, more people feel faster, productivity improves.
Gallup’s harder finding was that the productivity gain mostly stayed at the task level. Only about one in 10 employees in AI-adopting organizations strongly agreed that AI had transformed how work gets done. The gap between individual speed and organizational change is where quality problems collect. If the workflow is unchanged, AI output drops into the same approval path, the same performance review, the same customer handoff and the same manager calendar.
SHRM’s July report adds the missing quality language. Workers are asking whether AI saves time and whether the faster output leaves a lower-quality residue behind.
That matters because most organizations have no shared label for this kind of defect. A factual error is visible when someone checks the source. A hallucinated citation can be caught by a link review. A weak customer email is harder. A generic project plan may not be wrong, but it may still waste the next team meeting. A manager who says “this sounds like AI” may be correct and still unfair if the comment is vague.
The review problem starts when the company cannot distinguish between four different cases.
One employee uses AI to improve a rough draft, checks the sources, rewrites the argument and submits better work. Another asks a tool to summarize a complex meeting and sends the summary without reading it closely. A third uses AI because the manager said the team should be more efficient but never explained what good AI-assisted work looks like. A fourth hides AI use because disclosure feels like admitting a shortcut.
The output may look similar. The coaching should not.
Without a review file, managers will improvise. Some will ban tools because they do not want the quality burden. Some will reward fast output because it clears the queue. Some will punish junior workers for using the same tools executives use openly. Some will treat every mistake as an AI mistake and every AI-assisted paragraph as less authentic.
That path turns a quality problem into a trust problem.
The opposite failure is possible too. A company can make the review process so heavy that employees stop admitting how they work. They will rewrite prompts by hand, avoid useful internal tools, or keep AI use in personal accounts because the approved process feels slower than the task. A review file works only when it is proportional to the risk of the work. A customer escalation, hiring recommendation or performance note deserves more evidence than a meeting agenda.
SHRM found slop in ordinary work
The phrase “AI slop” is useful because it is unglamorous. It does not sound like transformation, augmentation or a productivity breakthrough. It sounds like what a coworker sees when they open a draft and realize that the next hour will be spent cleaning up someone else’s automated output.
Inside a company, slop has a different cost than it does online.
On a public feed, a low-quality AI post can be ignored. In a workplace, the same quality problem can enter a decision. It can shape a customer’s answer, an employee’s record, a hiring scorecard, a finance note, a policy summary or a manager’s view of someone’s judgment. The company may never call it AI slop. It may call it rework, review burden, poor writing, weak analysis, low ownership, bad judgment or lack of attention to detail.
The label changes by function. The pattern is the same.
SHRM’s numbers show why this cannot be left to informal taste. Forty-one percent of workers using AI is already too large for case-by-case suspicion. The 44% AI-slop admission among users suggests the problem is broad enough to require an operating standard, not a lecture. The leadership-use gap raises the stakes. If senior leaders use AI more often than individual contributors, they help set the acceptable norm for the company.
The easiest mistake is to make disclosure the whole standard. Disclosure matters, but it is not enough. An employee can disclose AI use and still submit weak work. Another can use AI invisibly and submit work that is well checked, well reasoned and better than the unaided version. The manager needs to evaluate the work, the evidence and the judgment around it.
Three signals matter more than the mere fact of AI use.
First, did the employee know which part of the work required human judgment? Summarizing ten customer calls may be a reasonable AI task. Deciding which customer risk deserves escalation is not the same task.
Second, did the employee preserve the evidence chain? A useful memo should point to the call, ticket, dataset, policy, interview note or source behind the claim. AI can help assemble that material, but it cannot become the source of record for a workplace decision.
Third, did the employee improve the output after the model produced it? Microsoft found in its 2026 Work Trend Index that 86% of surveyed AI users treat AI output as a starting point rather than a final answer. That is the habit managers should be trying to see. The problem is not the tool. The problem is the unreviewed handoff.
The manager review file should therefore record what the employee did after the model answered.
That record does not have to be bureaucratic. A manager does not need a five-page form for every sales email or design brief. But high-stakes work needs a small evidence trail. If AI helped produce a customer recommendation, the reviewer should see the customer facts that support it. If AI helped draft a performance note, the manager should see the observed behavior and dates. If AI helped prepare a hiring summary, the interviewer should see the structured evidence behind the summary.
AI slop becomes dangerous when it crosses from draft space into decision space.
Senior leaders use the tool before the rules catch up
Workplace AI policies often sound like they were written for employees at the bottom of the org chart. Do not paste confidential data. Do not rely on AI for final decisions. Check sources. Use approved tools. Escalate uncertain cases.
Those rules are sensible. They are also incomplete.
SHRM’s report found that leaders use AI more than workers below them. Directors and above were the most likely group to use AI at work. That means senior leaders set policy and model behavior at the same time. If executives ask for AI-assisted speed, they create pressure downstream. If they reward AI-polished answers without asking for evidence, they teach the organization that polish matters more than traceability. If they follow AI advice against their own ethical judgment, they turn an assistant into a shield.
The manager in the middle absorbs the contradiction.
A manager may hear one message from leadership: use AI to move faster. The same manager may hear another from legal, security or HR: keep people accountable, protect data and avoid automated decisions that cannot be explained. The team then brings in work that has been drafted, summarized, scored or rephrased by tools the manager may not have chosen.
This is where AI quality becomes a management competency.
Microsoft’s Work Trend Index points to the new skill set. Among surveyed AI users, 50% said quality control of AI output becomes more important as AI takes on more work. Forty-six percent cited critical thinking. Microsoft also described a small group of advanced users, Frontier Professionals, who routinely rethink workflows, create shared AI standards and decide when work should be done by AI versus a human.
That sounds like a specialist profile. It is becoming a manager profile.
Managers need to know when AI is suitable for a draft, when it is suitable for analysis, when it should be used only for synthesis, and when it should stay out of the decision. They need to coach employees on what counts as source support. They need to identify where the AI made the work sound more certain than the evidence permits. They need to catch when a model flattens the most important exception.
They also need to protect employees from arbitrary review.
If a manager dislikes AI and marks down any visible AI use, employees will hide it. If a manager loves AI and rewards speed over quality, employees will ship low-review work. If a manager cannot explain the standard, performance conversations become personal taste disguised as quality control.
The standard has to be observable.
For a customer-facing note, the observable standard may be whether the note reflects the actual customer facts, names the open risk and gives a next action the account owner can defend. For an internal analysis, it may be whether claims trace back to data, assumptions are named and the recommendation changes if the most uncertain assumption changes. For a performance review, it may be whether feedback ties to observed behavior rather than AI-generated adjectives.
AI raises the bar for review because it lowers the effort required to create a plausible answer.
That is why senior leader behavior matters. When leaders use AI, they should show their evidence habits along with their efficiency habits. A leader can say: I used AI to summarize the material, then checked the claims against the source notes and rewrote the recommendation. That sentence teaches more than a policy link.
BCG found saved hours without a destination
AI slop is also a time-allocation issue.
BCG’s June 2026 report, AI at Work: Why Strategy Matters More Than Tools, found that 74% of frontline employees are regular AI users, up 23 percentage points from 2025. Among regular frontline users, 42% reported saving at least eight hours per week. In human resources, BCG reported even higher time savings among regular users.
The missing piece was direction. BCG found that 66% of frontline employees received limited or no guidance on what to do with the time AI saved, and more than half were not reinvesting the time into more strategic work.
That finding explains a lot of AI slop.
When a company tells employees to use AI but does not redesign the work, the tool often becomes a faster way to produce the same queue. People write more drafts, generate more summaries, create more variants, answer more messages and fill more internal templates. Some of the extra output is useful. Some becomes noise. The review burden shifts to managers and peers.
The saved time has to go somewhere.
It can go into customer understanding, deeper analysis, coaching, quality review, documentation, better handoffs, skill development or a smaller workload. It can also go into more output. If the company chooses the last option without a quality standard, AI slop is a predictable result.
This is why the manager review file should not be limited to individual mistakes. It should also expose workflow design.
If ten people on a team are using AI to create three times as many status notes, the problem may not be the employees. The problem may be that the reporting system now rewards volume because output became cheaper. If recruiters are using AI to summarize candidates but hiring managers still ask for full context later, the summary may be saving time in one place and adding rework in another. If customer support agents use AI to draft replies but supervisors must rewrite them before they leave the queue, the organization has not saved labor. It has moved labor.
The review file should therefore ask where rework appears.
| Work type | AI role | Quality risk | Evidence needed | Manager action |
|---|---|---|---|---|
| Customer account note | Summarizes calls, tickets and next steps | Generic risk language misses the customer’s real blocker | Call excerpt, ticket history, owner confirmation | Coach on source-backed account judgment |
| Performance feedback | Drafts examples and tone | Inflated praise or vague criticism enters the employee record | Date, behavior, impact and prior coaching note | Require observed facts before wording polish |
| Hiring summary | Condenses interview notes | Candidate signal gets flattened or biased language survives | Structured scorecard and interviewer notes | Review evidence before moving candidate status |
| Policy or compliance memo | Summarizes rules | Old or wrong policy source enters an official answer | Current policy link and owner signoff | Escalate if source is uncertain |
| Product or strategy analysis | Builds first-pass argument | Confident recommendation hides weak assumptions | Dataset, assumptions, sensitivity check | Ask employee to state what would change the decision |
| Support reply | Drafts response to user issue | Tone improves while answer misses the actual issue | Customer issue, product state, known limits | Separate tone edit from factual review |
| Internal project plan | Generates tasks and timeline | Dependencies and owners are invented or blurred | Named owner, dependency and decision date | Convert draft into accountable plan |
This table is not meant to slow every task. It is meant to make the review proportional.
Low-stakes drafting can stay lightweight. High-stakes output needs evidence. The manager’s job is to know the difference and explain it before the work goes wrong.
Microsoft made quality control a core skill
Quality control used to sound like a back-office function. In AI-assisted work, it becomes a daily work skill.
Microsoft’s Work Trend Index is useful here because it separates AI use from human responsibility. The report found that most surveyed AI users say they treat AI output as a starting point and stay responsible for the thinking. It also found that quality control and critical thinking rose to the top of the human skills workers consider more important as AI takes on more work.
The implication is uncomfortable for companies that bought AI tools as a productivity shortcut. If quality control is a skill, it has to be taught, observed and rewarded. It cannot be assumed.
Most performance systems are not ready for that.
Traditional performance reviews ask whether someone delivered on goals, collaborated well, communicated clearly and showed good judgment. AI changes the evidence behind each category. An employee may deliver more output but with weaker source discipline. A manager may communicate faster but with less personal accountability. A team may collaborate through AI summaries but lose the shared context that used to come from reading the same source material. A leader may show confident judgment that originated as model advice.
The performance review has to separate three things: the quality of the final work, the employee’s handling of AI, and the work system that shaped the behavior.
Blending them creates bad incentives.
If the final work is poor because the employee pasted AI output without review, that is a coaching and performance issue. If the final work is poor because the organization pushed the employee to use a tool without training, that is also a management issue. If the final work is strong because the employee used AI well, the review should recognize the judgment, not pretend the work was less valuable because a tool helped.
AI quality control should become visible in leveling.
For an entry-level employee, the skill may be knowing when to use AI for a first draft, checking facts against assigned sources and asking for help when the output conflicts with instructions. For a mid-level employee, it may be choosing which parts of a workflow should be AI-assisted, explaining assumptions and catching subtle mismatches. For a manager, it may be setting team standards, designing review moments and coaching people without creating fear. For an executive, it may be modeling evidence-backed AI use and refusing to hide behind automated recommendations.
This is a better standard than a generic AI fluency badge.
AI fluency can become a vague credential. Quality control is observable. Did the person verify the source? Did they catch the missing exception? Did they improve the output? Did they know when to stop using the tool? Did they preserve accountability for the decision?
Deloitte’s 2026 Global Human Capital Trends gives the broader management frame. Seven in 10 business leaders said their primary competitive strategy over the next three years is to be fast and nimble. Deloitte also pointed to the orchestration of people and resources and workforce adaptability as central drivers of success.
AI output quality is part of that orchestration. A company cannot move faster if every AI-assisted draft creates another review queue. It cannot be more adaptable if employees do not know which judgment skills still matter. It cannot claim human advantage if managers have no language for the human work that remains after the model answers.
A review file for AI-assisted work
The simplest review file is a short record attached to high-stakes AI-assisted work. It should be easy enough for employees to use and concrete enough for managers to coach from.
It should not begin with accusation. It should begin with scope.
What part of the work used AI? Drafting, summarizing, analysis, translation, coding, research, classification, recommendation or formatting? Did AI touch only the language, or did it shape the substance? Did the employee use an approved internal tool, a general assistant, an agent connected to company systems, or a customer-facing feature?
The next question is risk.
Does this output affect a customer, candidate, employee, financial record, compliance response, security decision, product release or performance review? Does it use confidential data? Does it create a recommendation another person may rely on? Does it enter a system of record?
Only then should the manager ask for evidence.
The file can stay compact:
| Field | What the employee records | What the manager reviews |
|---|---|---|
| Work type | The task and the business context | Whether the task needed an AI review standard |
| AI role | Draft, summary, analysis, recommendation, code, classification or agent action | Whether AI shaped substance or only format |
| Source evidence | Links, notes, data, tickets, interview records or policy sources | Whether key claims trace to real material |
| Human revision | What changed after the AI output | Whether the employee added judgment |
| Quality risk | Possible error, bias, omission, stale source or tone issue | Whether risk matches the stakes |
| Escalation trigger | When legal, HR, security, product or another owner must review | Whether the trigger is clear enough before a mistake |
| Coaching note | What the employee should improve next time | Whether this is skill development or performance evidence |
| Decision owner | Person accountable for final output | Whether accountability stayed human |
This file creates a better conversation than “Did you use AI?”
It lets a manager say: the issue is not that you used AI to draft the customer note. The issue is that the note named a risk that was not in the call record and missed the customer’s renewal date. Next time, attach the source note and write the risk sentence yourself.
It also lets a manager say: the AI summary helped. You checked the policy link, added the missing exception and made the final answer more precise. That is good work.
The difference matters.
Workday’s June release of Agent-Ready Tools and Agent Passport points to the product direction. Enterprise systems are starting to add controlled guardrails for agents that access HR and finance data, plus verification signals for deployment. That helps with tool access and system trust. It does not replace manager review of the work that employees submit.
The tool can verify an agent. The manager still has to verify the work.
The file also protects employees. If a worker uses AI responsibly and improves the output, the record shows that. If a team receives contradictory guidance, the record can show where the process fails. If a manager wants to penalize AI use in general, the file redirects the conversation to quality, evidence and judgment.
The goal is not to make every task auditable. The goal is to keep high-stakes AI-assisted work from becoming invisible.
This is the employee side of the system. Workers need a way to show that AI helped them draft, compare, summarize or translate without making the final work careless. They also need a way to contest vague criticism. “This sounds automated” is not a coaching note. “The customer risk sentence is unsupported by the call record” is.
Performance talks need evidence, not suspicion
AI slop will enter performance conversations whether companies prepare for it or not.
A manager will receive a weak memo and suspect AI. A customer will complain that a response felt automated. A recruiter will see a candidate summary that sounds polished but thin. An employee will say a tool helped but the final work was theirs. A leader will ask why output rose while quality fell. HR will be asked whether AI use should appear in a review.
The wrong answer is to turn performance management into AI detection.
Detection tools will be unreliable in ordinary workplace writing, and suspicion will push employees to hide tool use. The better answer is evidence. What did the work require? What source supports it? What did the employee add? What risk did the manager ask the team to watch? What coaching was given before the review cycle?
That makes the conversation fairer and harder to dodge.
An employee should not be punished because a paragraph “sounds like AI.” The employee should be coached if the work lacks source support, misses the decision context, fails to meet a quality bar or sends unreviewed output into a high-stakes workflow. A manager should not be rewarded for pushing AI adoption if the team spends the saved time correcting slop. A senior leader should not be allowed to cite AI advice as the reason for an ethical choice they would not defend without the tool.
The standard should be simple: AI can assist the work, but it cannot absorb responsibility for the work.
That standard has consequences. Managers need time to review AI-assisted output. Teams need examples of good and bad work. Employees need permission to say when a tool is making the task worse. Performance systems need to reward source discipline rather than volume alone. Training needs to focus less on prompts and more on verification, judgment and escalation.
The review file is not glamorous. It is a way to keep quality visible after output becomes cheap.
At the end of a quarter, a manager should be able to look at an employee’s AI-assisted work and answer a concrete question: did this person use AI to strengthen judgment, or to avoid it? The answer should come from examples, source trails, revision habits and coaching notes, not from a hunch about writing style.
That is where SHRM’s July phrase changes the workplace conversation. AI slop is easy to mock when it shows up in public content. It is more expensive when it enters the manager review file. The company then has to decide whether it wants more output, better work, or a way to tell the difference.
Published July 8, 2026.