Illustrated AI job cards rise beside an anonymous worker at a desk beneath the words AI JOBS, LOW TRUST.

AI-generated editorial illustration.

On November 18, Indeed Hiring Lab and Glassdoor Economic Research published two measures of AI at work that seem to be moving in opposite directions. The share of US job postings on Indeed that mention AI or generative AI rose from 3.6% in September 2025 to 7.2% in September 2026. Among Glassdoor reviews that mentioned AI, the share classed as positive fell to 45% in 2026, from 82% in 2019. The first number is about what employers advertise. The second is about what a self-selected group of employees chose to say about their workplaces. Neither counts jobs created, jobs lost or employees whose opinions changed.

Illustrative scenario, not a reported employer case: An employer approves a vacancy that asks applicants to use AI. The ad goes live immediately. Whether the tool improves the work will be known only after a hire, training, supervision and a run of ordinary cases. Meanwhile, an existing employee might already spend an extra hour each day checking the tool’s output. An ad and a review can describe different stages of that rollout. Indeed’s new labor-market chapter places job ads, worker use, productivity and reviews beside one another, while its methodology makes their limits visible.

The release arrived a day after BSI reported that 54% of 1,287 surveyed business leaders across nine countries said AI was helping reduce headcount, up from 41% in its previous survey. Nearly half said entry-level opportunities had already fallen. The BSI findings sound more severe than Indeed’s assessment that there is little evidence of widespread job losses driven solely by AI. Yet the studies measure different people, places and outcomes. BSI asked executives what they believed had happened in their organizations. Indeed examined a US posting platform, a worker survey, employee reviews and a national productivity series. Neither can be subtracted from the other to calculate an AI job-loss rate.

For a company deciding its 2027 hiring and AI budgets, the question is narrower: what changed in its openings, workload, employee experience and output, and what evidence connects those changes to AI? Recruitment copy does not answer it. A critical review cannot answer it either. Both make the cost of leaving it unanswered harder to ignore.

A doubled label is not a doubled workforce

Indeed’s AI Tracker looks for AI-related terms in US job postings on its platform. The 3.6% and 7.2% values are monthly averages for September of two different years. They mean that AI appeared in twice the share of ads, not that the number of AI jobs doubled. The total number of postings, the mix of occupations and employers’ wording can all change. One opening might be relisted, and one employer may advertise several vacancies. The tracker is valuable as a measure of advertised demand for AI-related work, with the denominator and platform kept in view.

The share of companies with at least one AI-mentioning posting increased by 44% over the same period. In the report’s chart, 7.6% of firms with an Indeed posting had at least one flagged ad by September 2026. A firm with one AI-marked listing and a firm rebuilding every function count the same in that measure. It says the vocabulary has reached more employers; it cannot tell whether the work was redesigned after a person started.

Indeed makes a second, smaller classification that should not be confused with the tracker. It calls a title “AI-touched” when AI appears in the job title itself. Almost two-thirds of those titles now sit outside traditional technology occupations. The examples include truck drivers, retail sales associates and physical therapists. AI has entered the names of jobs far beyond software. That does not establish that most drivers or therapists use AI every day. The title subset, the broader ad-text tracker and the whole labor force have different denominators.

An employer might write AI into a job ad because a worker will operate an existing tool, because a new hire is supposed to introduce one, or because the wording helps sell the role. A posting cannot distinguish those motives. The organization would need to connect the listing to a filled role, an assigned workflow and an observed outcome. Indeed’s series does not follow that chain. Calling every flagged ad an “AI job” turns a search term into a claim about the work itself.

This distinction matters outside a human resources office. A founder reading rising AI-ad figures may assume rivals have already rebuilt their operations. A finance chief may assume AI skills command an immediate productivity return. A candidate may accept a job expecting effective tools and discover that the employer has bought licenses without changing the work. Each reader is using a different interpretation of the same ad. Indeed’s data supports the observation that AI language is spreading; the more ambitious claims require evidence from inside companies.

For an applicant, the wording also changes the interview. “Experience with AI” might mean writing prompts, checking model output against a professional standard, building a workflow, or maintaining the systems that run a data center. Indeed says almost a quarter of data-center-related ads in the first half of 2026 were for installation and maintenance workers, whose advertised pay was 42% higher for comparable work at a data center than elsewhere. That is infrastructure demand created around AI, not proof that those tradespeople use a chatbot to do their core job. A candidate who asks which part of the role is genuinely AI-related can learn more than the job title supplies. The employer should be able to answer without treating the question as resistance to technology.

The distinction also changes workforce planning. If a business adds AI wording to existing vacancies, its recruiter may report a surge in AI hiring without adding a single position. If it changes a junior opening to a senior one, the total vacancy count may stay flat while the path into the occupation worsens. If a data-center builder posts more electrician jobs, the AI economy may create demand in an occupation that is itself hard to automate. One dashboard labeled “AI jobs” would flatten all three decisions. They require different budgets, training pathways and ways of judging success.

The review is a signal with a selection problem

Glassdoor’s measure has its own boundary. The researchers searched US employee reviews for “AI,” “artificial intelligence,” “ChatGPT,” “GPT” and “LLM.” They excluded names such as Claude and Gemini because those terms produced many false positives. For sentiment, a review is positive if AI appears only in its pros section, critical if it appears only in cons, and mixed if both sections include it. An algorithm did not read every comment and decide whether the worker liked automation in general. The classification follows where the person placed a keyword in a structured review.

By that rule, 45% of AI mentions were positive and 51% were critical in 2026. The article compares the positive share with 82% in 2019. It also says AI mentions in Glassdoor reviews were 2.6 times as common in September 2026 as a year earlier. The sentiment analysis covers current employees’ reviews; the broader count of AI mentions includes both current employees and people who left within the past year. These details in the consolidated methodology matter because the population who mentions AI in a review is not a random sample of all workers.

The early group mentioning AI in 2019 may have been concentrated in firms and roles that valued the technology. The 2026 group is larger and likely contains a wider range of workplaces. A fall in the positive share therefore does not show that the same employees changed their minds. Glassdoor users also choose whether to leave a review at all. Workers with unusually good or bad experiences may be more likely to post, and the site cannot see the silent majority at an employer. What changed is the visible review conversation. The 51% figure cannot be read as the share of all US employees who oppose AI.

Still, the complaints are not merely abstract fear of future layoffs. Among AI-critical reviews, references to burnout were 3.5 times as common as in all reviews; job insecurity, 3.8 times; layoffs, 7.1 times. Twenty percent of workers who put AI in the cons section also mentioned layoffs. Those are associations in review text, not proof that AI caused a layoff or burnout. Glassdoor reports that the differences remained after adjusting for each review’s overall star rating. That reduces one obvious explanation, a generally unhappy reviewer, but does not make the groups comparable in every other way.

The chapter also reports that only 37% of AI-critical reviewers approved of their CEO, against 89% of AI-positive reviewers. A single underlying management problem might lead to both the poor CEO rating and the AI complaint. A badly managed rollout might also damage trust in leadership. The data cannot assign direction. A leadership team can, however, read the complaints themselves, separate tool failures from fears about staffing, and ask whether managers are being judged on a metric that makes employee review time invisible.

Employers and workers describe different workplaces

Indeed’s separate Workforce Insights Survey offers a third vantage point. Among employed US respondents, weekly AI use at work rose from 36% in 2025 to 42% in 2026. Weekly personal use rose from 44% to 53%. More people are using AI, but personal and professional use remain separate questions. An employer cannot claim the 53% as an internal adoption rate, and an employee who uses a chatbot at home may still have no approved tool at work.

The same survey found that 31% described their employer’s AI adoption as low or nonexistent; 25% described high adoption and integration. The methodology defines these categories by respondents’ perceptions of whether AI is core, encouraged, absent or discouraged. They are not audits of software logs or procurement records. That matters when an employer says a tool is deployed but the people meant to use it experience only a login and a policy memo.

The report says 56% of employed US respondents did not agree that AI makes, or could make, them more efficient. “Did not agree” includes people who did not select either agreement option; it must not be rewritten as “56% said AI makes them less efficient.” Among regular AI users at work, estimated time saved per day slipped from 2.3 hours to 2.1 hours across the two survey waves. Both are self-reported estimates, not stopwatch observations. A changing mix of users, tasks and expectations may explain some of the movement. The point is that more use does not guarantee that the reported return per user rises.

Other sources make the same measurement problem visible from the opposite side. Adecco’s November 19 worker survey reported more respondents saying they used AI agents at work while fewer said they had participated in AI training. Its workers, countries and definitions differ from Indeed’s US sample. There is no sound way to combine the percentages into one adoption curve. But a buyer can recognize the shared operational risk: access may spread faster than the paid time and assistance needed to use the tool well.

In BSI’s leader survey, 81% said they personally used AI to become more efficient, while 42% said they would be less likely to promote someone who sent them AI-generated content. The respondents were leaders, not a matched sample of the employees whose work they assessed. That apparent double standard is a reported attitude, not evidence that a particular promotion was denied. It is nevertheless an uncomfortable policy question. When managers endorse AI use in general but penalize visible AI work in practice, a worker has to infer the rule from inconsistent signals.

Three paper evidence streams connect vacancy cards, a worker figure and a blank comment sheet to a decision desk.

AI-generated editorial illustration. The three streams represent different kinds of evidence about work, not measured data.

Employment changes can hide inside a stable headcount

The disagreement between “jobpocalypse” and “little evidence of widespread losses” is partly a disagreement about where to look. BSI’s 54% figure is the share of surveyed leaders saying AI helped reduce headcount in their organizations, not the share of all jobs eliminated by AI. Its 49% entry-level figure records leaders who say opportunities have already fallen, not a count of missing positions. The sample spans nine countries. Indeed’s chapter covers US job ads, review text, surveys and a national productivity series. A company trying to make a factual claim about its own jobs needs more specific records than either headline.

Gartner’s November 10 forecast adds another category: hiring that never happens. It estimates that 1.3 million positions were deferred in 2026 because executives postponed creating or filling them. It projects 10.6 million jobs created and 8.9 million removed worldwide by 2030, a net gain of 1.7 million. These are model forecasts, not observed payroll totals, and the public release does not provide enough of the model for an independent reconstruction. Gartner itself distinguishes a deferred position from a lost job. That distinction is important to a person hoping for a first role, even if it does not appear in a layoff report.

There is already a Digidai analysis of vacancies disappearing without mass layoffs. Indeed’s new evidence changes the buyer’s problem in a different way: the vacancy may survive with AI written into it, while the ladder beneath the role narrows. Indeed’s July study attributed 71% of the increase in software-development postings between May 2025 and May 2026 to senior roles. Its November chapter repeats that split and says 37% of the increase explicitly mentioned AI in the title. The rise in postings is real within that platform and window; it does not prove that an entry-level applicant has gained the same opportunity.

Anthropic provides a company-level example, though one at the frontier of model development. In its October employment study, junior employees’ share of technical departments fell from 21% to 14% since early 2025, while junior hires’ share fell from 27% to 9%. The company says technical hiring continued to grow in absolute terms as other functions grew faster. It also reports more internal movement across departments. That combination cannot be compressed into “AI eliminated jobs” or “AI created jobs.” Hiring composition, growth and reassignment changed together at one unusually AI-intensive employer.

Anthropic’s researchers acknowledge that its conditions may not transfer elsewhere. Its stock ownership, product growth and ability to redeploy people are not typical of a regional hospital or a small manufacturer. The study also concedes a limit of internal reassignment: it does nothing for a person never hired. A company that maintains a stable headcount while quietly stopping junior intake may face that problem years later, when there are too few experienced people to replace current staff.

That cost is easy to miss in a one-year productivity plan. Senior employees can review AI output because they accumulated judgment through earlier work, including the low-stakes tasks now handed to software. A smaller entry cohort can make this year’s payroll look efficient and next year’s promotion list thin. The mechanism is plausible, but the available studies do not estimate the size of a future skills shortage. A company should check its own progression data: how many juniors entered, which tasks they still practiced, how many reached independent work, and which supervisors had time to coach them. Those are slower measures than posting counts, but they test whether the workforce can reproduce its own expertise.

Productivity figures cannot settle the trust question

The national productivity number offers a check on sweeping claims, but it cannot adjudicate one employer’s rollout. Indeed cites the US Bureau of Labor Statistics’ nonfarm business output-per-hour series: annualized growth since late 2019 was 2.1%, above the 1.5% pace from 2008 to 2019 and close to the longer-run average. The researchers are careful not to assign that change to AI. It spans a business cycle, the pandemic recovery, investment and other technologies. A company might gain substantially from a tool while the national series changes little; a national improvement might have little to do with a particular AI product.

Nor should a firm count draft volume as finished work. Anthropic reports that code commits per engineer rose sharply as agents spread, while its own study warns that more code is not necessarily more business impact. It compared some agent-written and human-written code for later removal or reversion and found no difference on those measures. That is useful, but a revert rate cannot capture every maintenance cost or customer consequence. For nontechnical workers, the report estimates that reading all interactive agent output could occupy more than half a workweek. Human attention can become the scarce input even as generated output becomes cheap.

This is where the worker review connects to a finance decision. A team might process more drafts with the same headcount while spending additional hours checking them, repairing errors, meeting to coordinate exceptions or explaining a changed role. If those hours are invisible, the productivity report overstates the gain. The employee sees the new work. The manager sees the tool’s output count. Glassdoor cannot quantify that mechanism across the economy, but a company can measure it in its own workflows.

An illustrative, unreported example makes the accounting problem concrete. Suppose a service team buys an AI reply assistant and asks every employee to use it. The queue dashboard shows faster first drafts. The workers now spend time checking policy, correcting wrong details and handling customers upset by a generic response. No real company or customer case is described here. The example identifies the fields a buyer would have to count: accepted outputs, corrections, repeat contacts, review minutes and the time saved after those costs. A license count or prompt count cannot substitute for them.

Four records for a credible AI work claim

The studies suggest a small evidence file that a business can build without pretending to estimate the national AI employment effect. Each record answers a separate question. They should be reviewed for the same role and time window, and the organization should say when it lacks a baseline.

RecordWhat to countWhat it can and cannot show
Hiring pathApproved requisitions, posted openings, level changes, fills, backfills and canceled rolesShows where a job opportunity changed; it does not by itself prove AI caused the change.
Work pathTasks moved to tools, human review time, corrections, repeat work and customer outcomesTests whether output became useful work; it needs a pre-rollout comparison and quality measures.
Worker pathPaid learning time, access, workload, mobility, complaints and retention by teamReveals who absorbed the transition cost; surveys and reviews still have selection limits.
Financial pathTool, integration, supervision and rework costs beside verified savingsTests the business case after hidden labor is included; it should not turn projected savings into realized cash.

The first record protects against a convenient ambiguity. A company can say it had no AI layoffs while eliminating a vacant role, delaying an approved opening, replacing two junior roles with one senior requisition or moving work to a contractor. Those are different decisions with different effects. A board or worker representative should be able to see them separately. The second record guards against counting every generated artifact as production. The third makes the experience of the people doing the checks visible. The fourth asks whether the organization actually saved money after paying for all of the extra work.

This file would also discipline interpretation of public studies. Indeed’s 7.2% is an ad-text share, not a filled-role count. BSI’s leader responses are perceptions of organizational change, not a census of eliminated positions. Glassdoor’s 51% critical share describes reviews that mention AI, not a referendum by the labor force. Gartner’s net gain is a forecast. Each measure belongs in the discussion, with its source and denominator attached. None should be used as a shortcut for another.

There is a practical test for an employer that has already announced an AI efficiency target. Choose one workflow and trace a month of cases before and after rollout. Record the time from intake to a satisfactory outcome, including the human corrections and returns that the tool did not count. Compare the same measures for workers with and without paid training, while noting that those groups may receive different work. Then reconcile the staffing plan: did vacancies disappear because demand fell, because the workflow genuinely needed fewer hours, or because managers assumed future savings? This will not isolate a clean causal effect without stronger design. It can expose a business case that counted the tool’s work and forgot the worker’s.

The comparison may reveal an ordinary, uncomfortable outcome. A firm could keep its current staff, ask them to produce more with AI, hire fewer beginners and receive worsening employee feedback. Another could use AI to remove routine work, expand a service and create new positions while employees still complain about poor tools. The same public data can accommodate both stories. The internal records would show which one the company is living.

A recruitment promise reaches the desk

The 2027 planning cycle gives the numbers a practical deadline. An employer writing AI into a job ad can specify which tool the hire will use, what judgment remains theirs, who teaches the workflow and how performance will be assessed. If a department claims AI will let it defer hiring, the same plan can record which tasks disappear, who inherits exceptions and what output improves. Those are ordinary management questions, but the public measures show how easily they are skipped when an AI label is treated as a result.

A manager may need to make the answer visible before an offer is accepted. “You will use AI” tells a candidate almost nothing. “You will review AI-prepared customer summaries, correct source errors and decide which cases go to a specialist” describes a job. The second version can be tested after a few months: are the summaries accurate, how many cases require correction, and has the specialist queue grown? It also tells the incumbent worker what the new hire is supposed to take over. If the employer has no answers yet, it can say that the workflow is still being tested and state who will decide whether to keep it.

The employee review belongs in that follow-up, even when it is inconvenient. A company cannot infer a representative workforce view from a public review site, but it can ask its own team whether the tool changed the quality of the work, whether it moved hidden checks onto them, and whether the promotion criteria match what managers actually value. The questions are more specific than “Do you like AI?” They connect a worker’s complaint to a decision the employer can change.

Indeed and Glassdoor do not show that an AI-labeled vacancy causes a critical review. They show that the advertising signal and the employee conversation can move in different directions at the same time. When the next AI-marked vacancy is approved, someone should be able to describe the work behind the label and the hours it will cost. Otherwise the ad may reach an applicant before the employer has answered the worker already sitting behind the tool.