monday.com AI survey exposes the management cost of faster output

The monday.com AI survey finds that 53% of senior leaders struggle more to assess employees’ original contributions after adopting AI. The finding puts a management problem inside the productivity pitch: producing finished work and understanding someone’s contribution are different tasks.

Published October 21, the Nielsen-conducted research surveyed 1,701 directors and senior managers across seven countries. Its findings suggest that employers face a harder measurement problem as AI-assisted work becomes ordinary. A polished deliverable offers less evidence about the judgment behind it.

That tension sits alongside evidence that AI can improve performance in specific settings. Research by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found measurable gains in customer support. The question for employers is whether those gains survive the full journey from individual output to reviewed, usable work.

The monday.com AI survey puts evaluation under pressure

According to the company’s AI management findings, 47% of respondents spend more time reviewing and guiding team output. Meanwhile, 78% have changed their employee evaluation approach since adopting AI.

Those findings describe related pressures, although they do not establish that the same respondents experienced every problem. They also do not measure how much additional review time managers spend. The distinction matters when translating employee evaluation difficulties into claims about business performance.

The immediate change is a weaker connection between a finished deliverable and the individual contribution it appears to represent.

Before AI assistance, presentation quality could offer some evidence of writing ability, preparation, or familiarity with a subject. It never provided a complete picture. Colleagues, templates, editors, and existing materials have always influenced workplace output.

AI makes that inference more complicated because it can supply much of the visible presentation. A clear explanation might reflect deep understanding, effective editing, or acceptance of a generated answer. The final document alone cannot reliably distinguish those possibilities.

For managers, the resulting question extends beyond whether an employee used AI. They need to understand whether the employee selected appropriate evidence, caught mistakes, and made defensible decisions. Those contributions can remain invisible in the finished text.

The research also complicates a simple resistance narrative. In the same survey, 94% expressed excitement about improving work with AI. Enthusiasm and difficulty coexist, suggesting that the problem concerns implementation and evaluation rather than straightforward rejection.

The headline finding received attention through TechCrunch’s coverage of employee evaluation difficulties. The primary report supplies the more useful basis for interpretation: these are senior leaders describing their experiences.

That is meaningful evidence about management friction. It is not a controlled comparison of organizations using AI against organizations avoiding it. Nor does it establish that employers have lost more time than employees have saved.

The monday.com AI survey therefore identifies a pressure point rather than settling the productivity debate. Organizations adopting AI need ways to evaluate contribution without assuming that fluency demonstrates expertise. Otherwise, the apparent improvement in output leaves an unresolved question about who deserves credit.

This changes the management conversation. The useful question becomes whether employers can recognize effective AI-assisted work while preserving meaningful standards for competence. That requires looking beyond how convincing a deliverable appears when it reaches someone’s inbox.

Managers inherit work that output metrics miss

The people under immediate pressure are managers responsible for both delivery and employee development. They must accept useful assistance while deciding whether their teams can explain and defend the resulting work. Those responsibilities overlap, but they are not interchangeable.

A manager can approve a correct document without learning much about the author’s underlying capability. That might be sufficient for a routine communication. It becomes less sufficient when the same document informs promotion, staffing, or readiness for a more demanding assignment.

Faster production creates business value only when the resulting work remains economical to review and dependable to use.

Consider an illustrative product-planning memo. An employee uses AI to organize customer observations into a recommendation about which feature deserves attention. The memo reads clearly and presents an orderly argument.

The manager still needs to establish whether the observations support the recommendation. Were conflicting customer requests preserved? Did the employee distinguish direct evidence from interpretation? Did the draft quietly turn an assumption into an apparent fact?

These are hypothetical questions, not documented outcomes from the survey. They show why presentation quality cannot carry the entire evaluation burden. A readable memo can reduce communication effort while leaving the underlying decision just as difficult.

The employee might have completed the task faster and performed excellent judgment throughout. Alternatively, the manager might need to reconstruct the analysis because the supporting evidence is missing. The same polished surface can conceal very different working processes.

This makes narrowly defined output metrics less informative. Counting completed documents rewards production without establishing whether colleagues can act on them. Measuring drafting time excludes whatever happens during clarification, correction, approval, and implementation.

The economic unit that matters is the completed workflow. A draft contributes to that workflow, but so do the people who validate it and resolve its consequences. Moving effort between those participants does not automatically eliminate effort.

Employee development adds another complication. Managers need to know whether someone can recognize when an answer is weak, incomplete, or inappropriate. A successful deliverable provides limited evidence if the employee cannot explain the choices that shaped it.

That does not mean employees should perform every task without assistance. Work already depends on tools and collaboration. The evaluation problem is identifying which forms of judgment remain essential when assistance handles more of the visible execution.

For an engineering manager, that might involve understanding why a proposed change fits an existing system. For a sales manager, it might involve checking whether a recommendation reflects the customer’s actual constraints. The relevant evidence varies with the job.

The pressure also reaches employees whose strongest contribution involves rejecting bad suggestions. Their work can disappear from view when an evaluation emphasizes only the final artifact. Preventing an error often leaves less visible evidence than generating another page.

These are analytical implications of the findings, rather than additional survey results. They explain why organizations cannot resolve employee evaluation difficulties simply by demanding more AI usage. Adoption and competence answer different questions.

Managers consequently face a difficult allocation decision. They need enough visibility to assess judgment without turning every assignment into a detailed audit. If evaluation becomes excessively burdensome, the process consumes the efficiency that assistance was supposed to create.

Productivity gains depend on where the work lands

There is credible evidence that AI assistance can improve workplace performance. The strongest version of the productivity argument does not depend on every employee becoming faster. It depends on measurable improvements in particular tasks, under particular conditions.

Brynjolfsson, Li, and Raymond’s customer support study examined 5,172 agents and found a 15% average increase in issues resolved per hour. Benefits varied across workers, with less experienced employees gaining more.

The researchers also reported differences in quality and learning. Those findings matter because they move beyond whether participants enjoyed using a tool. They examine an operational outcome with a recognizable connection to the work.

AI assistance and managerial friction can exist together because they describe different parts of the production process.

A support assistant that helps an employee resolve a customer’s problem operates within a relatively specific workflow. A general writing assistant used to prepare an internal strategy document faces different expectations. The evidence needed to assess success also differs.

The comparison does not imply that support work is simple. It shows why a productivity claim needs a defined task and outcome. “More useful resolutions per hour” conveys information that “more AI-assisted content” does not.

A contrasting result came from METR, the AI evaluation organization. Its developer productivity experiment involved 16 experienced developers completing 246 tasks in familiar open-source projects. Access to early-2025 AI tools increased completion time by 19%.

Participants nevertheless believed that AI had made them faster. That gap illustrates a measurement problem: a workflow can feel easier or more productive without producing the expected time savings. Perceived assistance and measured acceleration require separate evidence.

The result has clear boundaries. It concerned particular developers, repositories, tasks, and tools during a specific period. It does not establish that current coding assistants slow every developer, or that gains reported elsewhere are imaginary.

METR itself later stressed those limitations. In its February 2026 productivity measurement update, the organization said selection effects made its newer estimates unreliable. Some developers did not want to participate under conditions requiring work without AI.

The update also described difficulties measuring time when developers used multiple agents concurrently. METR considered newer tools likely more helpful than its earlier estimates suggested. It cautioned that the available data provided weak evidence about the size of that improvement.

Together, these studies support a restrained conclusion. The effects of AI depend on context, worker experience, task selection, and the definition of success. Neither enthusiasm nor an isolated negative result settles the wider question.

The product comparison follows the same logic. ChatGPT and Claude can help create text, while coding assistants operate inside software development workflows. A work-management platform such as monday.com addresses coordination around tasks and teams.

Those categories overlap, but they do not answer identical questions. An assistant can improve a draft without establishing who verified its claims. A coordination system can record activity without establishing whether that activity produced a sound decision.

The primary conflict therefore remains productivity promises versus the work required to realize them. Competing brands provide context, but switching assistants does not automatically resolve ambiguous evaluation standards. The employer still needs to define what counts as good work.

That is the useful interpretation of the monday.com AI survey. It directs attention toward the organizational conditions around assistance. The next productivity gains depend partly on whether those conditions make useful output easier to recognize and approve.

A polished document can hide judgment in either direction

The most obvious concern is that AI-generated polish makes weak work appear stronger. An equally important possibility runs in the opposite direction. An employee can exercise excellent judgment while receiving less credit because the visible writing looks machine-assisted.

That creates a fairness problem before any automated performance rating enters the picture. If managers treat polished language as suspicious, capable employees can lose recognition. If they treat it as proof of competence, weak reasoning can receive undeserved approval.

Fair evaluation needs evidence of responsibility and judgment without making employees prove that they worked unaided.

An illustrative research summary shows the dilemma. An employee gathers original material, identifies conflicting findings, rejects unsupported claims, and uses AI to improve organization. The final prose reveals little about the difficult decisions that made the summary trustworthy.

Another employee might submit similarly fluent text after checking little of its supporting evidence. The documents could look comparable during a quick review. Their reliability, and the employees’ demonstrated competence, would differ substantially.

The answer cannot rest entirely on detecting generated language. Even a correct identification would establish assistance, rather than the quality of the employee’s decisions. Knowing that AI touched a document does not explain whether the author used it responsibly.

Nor does a record of prompts automatically solve the problem. It can show what someone asked a system to do. It cannot, by itself, establish whether the person recognized a faulty answer or selected appropriate evidence.

This is where the distinction between attribution and accountability becomes useful. Attribution concerns which person or system supplied a particular contribution. Accountability concerns who owns the result and must explain the decisions behind it.

Employers need both concepts, but the level of detail should match the task. A routine internal announcement does not require the same scrutiny as a recommendation affecting staffing. Treating every document as equally consequential creates unnecessary review work.

Performance standards also need consistency. An employee should not be praised for using AI on one assignment and penalized for equivalent assistance elsewhere without explanation. Inconsistent expectations make evaluation harder even when the underlying work is good.

Knowledge management becomes relevant because evidence often exists outside the final deliverable. Decisions, source material, and earlier discussions can help explain why a recommendation took its particular form. Their value comes from making reasoning recoverable.

A searchable knowledge base offers useful background on organizing that supporting material. It does not replace managerial assessment. Access to records helps only when reviewers understand which evidence matters.

The organizational challenge is maintaining that context without creating an indiscriminate surveillance process. Recording every interaction generates more material to review. A narrower record of consequential sources, decisions, and approvals can be more relevant to the evaluation question.

This is an analytical judgment, rather than a tested intervention from the Nielsen research. The survey does not establish which documentation approach works best. Employers would need their own evidence before claiming that a particular policy reduced review costs.

The same caution applies to AI-assisted evaluation. A system can organize feedback into readable language while leaving the underlying assessment incomplete. Better phrasing does not supply missing observations or resolve conflicting accounts of an employee’s performance.

Ultimately, a polished document can conceal either dependence or expertise. The management task is to distinguish those possibilities through evidence connected to the job. Otherwise, the evaluation process risks rewarding appearances while overlooking the decisions that make assistance useful.

The survey identifies friction, not a productivity verdict

The monday.com AI survey deserves attention because it identifies a specific organizational problem. Its findings do not justify declaring that AI has failed at work. They also do not establish that every reported difficulty is a temporary adjustment.

The research was commissioned by a company selling work-management software. That commercial context should remain visible when interpreting the company’s proposed solutions. A finding about coordination problems does not independently validate a platform designed to address them.

The evidence supports concern about evaluation and review burdens, while leaving their net economic effect unresolved.

Self-reported difficulty is relevant because managers’ experiences can influence adoption and employee relationships. However, it differs from observed time use or independently assessed performance. The measures answer related questions without becoming interchangeable.

The surveyed population also matters. Directors and senior managers are not a direct substitute for every frontline supervisor or individual contributor. Their perspective illuminates leadership experience without fully describing how employees experience the same arrangements.

A cross-country sample broadens the research’s scope, but it does not guarantee identical conditions across workplaces. Jobs, management practices, and adoption stages differ. Aggregate findings can conceal variation that matters to an employer deciding how to respond.

Causality remains another limit. A respondent can associate greater evaluation difficulty with AI adoption without isolating AI from other changes. Expanding workloads, altered targets, or reorganized teams might also affect the experience.

These limitations do not erase the reported problem. They define what stronger evidence would need to establish. The next research step should connect reported friction with observed work, rather than assuming the connection is already measured.

One interpretation is that employers are paying an adjustment cost while redesigning their processes. Another is that cheap generation creates an enduring demand for expensive judgment. The current findings cannot determine which explanation dominates.

The first interpretation would gain support if review burdens declined as teams became more experienced. The second would gain support if increased generation continued to overwhelm approval capacity despite training and clearer expectations. Both require observation over time.

There is also a possible positive interpretation of additional review. Managers might spend more time checking work because employees now attempt more valuable assignments. In that case, higher review effort could accompany a useful increase in ambition.

Alternatively, extra review might consist mainly of correcting avoidable errors or recovering missing context. That would indicate displaced effort rather than a straightforward productivity gain. A time measure alone would still need a quality measure alongside it.

The defensible conclusion is narrower than the headline’s broad promise. AI changes the information managers can infer from a finished artifact. Employers need better evidence about the consequences before treating increased output as proof of improved performance.

Watch review costs, evaluation rules, and workflow results

Over the next several months, the most useful signals will concern review effort, evaluation practices, and completed work. These are more informative than another count of employees experimenting with AI. They connect adoption to the management pressure identified in the research.

The next test is whether employers can turn assisted output into accepted work while keeping contribution visible.

First, watch whether companies and researchers measure the work after the draft. Useful disclosures would include correction effort, approval delays, and the attention required from reviewers. Those measures would help distinguish saved effort from effort transferred to someone else.

Falling review burdens alongside stable quality would weaken the concern that AI simply moves work upstream. Persistent correction costs would strengthen it. The important comparison would follow similar assignments through completion, rather than stop when generation ends.

Second, watch how employers change performance criteria. A meaningful update would explain how they assess verification, judgment, and ownership of AI-assisted results. A requirement to use more tools would offer less evidence that evaluation difficulties are being resolved.

Clearer standards would strengthen the case that organizations are adapting successfully. Continued reliance on document volume or presentation quality would leave the attribution problem intact. Employees need to understand what earns recognition before managers can assess it consistently.

Third, watch for workflow-level results with credible comparisons. Stronger evidence would connect assistance to usable outcomes while accounting for quality and review. It would also acknowledge differences between jobs instead of presenting a single productivity claim for every worker.

A vendor case study can contribute useful detail, but its design matters. Readers should look for defined tasks, comparable conditions, and an explanation of what the measurement excludes. A persuasive narrative without those elements cannot settle the tradeoff.

These signals do not depend on a particular scheduled product launch. They are tests for the next round of claims from employers, researchers, and software companies. Each would help establish whether the management headache is shrinking or becoming embedded.

For knowledge workers, the immediate implication concerns how their contribution becomes visible. Producing a clear document remains useful, but explaining its evidence and decisions can matter just as much. The strongest work makes the result dependable without forcing colleagues to reconstruct it.

For enterprise buyers, the implication concerns what a purchase is expected to improve. Drafting assistance, coordination, and employee evaluation involve different requirements. A successful deployment needs evidence tied to the problem it is intended to solve.

Read the monday.com AI survey as an invitation to examine that whole chain. Look beyond whether a tool creates an impressive first draft. Ask whether the recipient can verify it, use it, and understand who exercised judgment. Compare those answers with the effort required before assistance entered the workflow. When the next productivity claim arrives, examine its stopping point: generation, approval, or a completed outcome. That boundary often determines whether the claim describes individual convenience or organizational value. For your team’s next AI-assisted assignment, can the reviewer recognize the employee’s contribution without repeating the underlying work?