An anonymous worker studies a training notebook while a large orange AI workflow tile approaches the desk beneath the words AI USE, TRAINING GAP.

AI-generated editorial illustration.

On November 19, staffing company Adecco published a survey with two lines that would sit awkwardly beside each other in an annual training plan. The share of respondents who said they used AI agents at work rose from 64% in 2025 to 74% in 2026. The share who said they had participated in AI training fell from 77% to 67%. Both figures come from Adecco’s Global Workforce of the Future release, not from a census of workers or an audit of company training records.

The release draws on an online survey of 40,500 workers in 36 countries and 21 industries, fielded in June and July. It does not say that the ten-point training decline caused an increase in unprepared workers. Nor does it follow the same people from one year to the next. The samples and questions need more scrutiny than the public summary permits. But a manager buying licenses now has a problem to price: paid learning time and job redesign may arrive months after access does.

Adecco also reports that 66% of surveyed workers say AI changed the skills their jobs require. Among workers whose employers involved them in redesigning work, 96% said they were willing to adapt to change. The corresponding figure among those not involved was 29%. That 67-point difference is striking. It is an association between groups, not a measured effect of inviting someone to a workshop. A better-resourced employer might both involve workers and offer training, while people who expect to remain with their employer might be more likely to participate.

In a November 18 study release, The Conference Board says only 11% of surveyed leaders had integrated AI’s expected effects into workforce and financial plans across their enterprise. Seventy percent expected job, task or workflow redesign to rank among their most consequential AI-related workforce actions. The Board surveyed leaders; Adecco surveyed workers. Subtracting one percentage from the other would invent a gap neither study measured. Put the findings beside a real budget decision, though, and the problem is plain: an employer can fund access this quarter while the cost of changing the work remains unassigned.

Licenses arrive before learning time

What, exactly, did Adecco count? Its release calls the respondents’ tools “AI agents,” but the publicly accessible release does not publish a detailed taxonomy of agent autonomy or the underlying questionnaire. A worker may understand the term differently from a product engineer. The 74% figure is a survey response under Adecco’s category. It cannot establish that three-quarters of all employees operate autonomous systems.

The public summary also gives the direction of the training series without a full account of which activities counted as training. A one-hour introduction, coached practice on a live workflow and a formal certification may all be described as training by different respondents. The full report is linked from the release but was not publicly readable in this review. That limits any precise claim about training quality or the comparability of the annual samples.

Within those bounds, the numbers describe a procurement problem. Software access can be bought centrally and allocated quickly. Learning takes scheduled employee hours, a manager who can release those hours, examples from actual work and someone qualified to correct mistakes. A license approval records an expense; a training plan needs to record time away from production as well. If the latter sits outside the AI budget, the apparent cost of adoption is artificially low.

The result can show up in one job before it reaches a headcount plan. Consider an illustrative, not reported support team that adds an agent to draft responses. A manager can count licenses and response times immediately. The worker still has to learn which cases require a human answer, how to verify a draft against policy and when a faster response is a worse response. Those tasks consume time that a seat price does not include. This is not an Adecco client case or an observed deployment.

Adecco’s own figures suggest why an access-only rollout may fail to deliver what finance expects. Workers using AI that they described as well integrated into workflows reported 20% more time saved per day than workers who described their AI as less integrated. Among AI users, 93% of respondents who completed training said AI made them more productive, against 47% of those who had not. Both comparisons are self-reported. Employers that offer training may also choose better tools or give staff more useful tasks. Neither comparison isolates the incremental return on a training dollar.

That makes the unit of analysis important. A training slide deck does not guarantee an employee can use an agent safely in a live queue. A claim of minutes saved does not say whether a customer received a better answer, a worker had a lighter day, or a manager raised the target. The study offers a reason to investigate those differences inside a firm, not a universal conversion rate for a spreadsheet.

A job changes before its title does

The Conference Board’s sample of 775 international leaders and 35 interviews puts numbers on the planning lag. Thirty-five percent said AI considerations had entered workforce and financial plans for some functions, roles, business units or geographies; 34% were still discussing the implications without incorporating them into plans; 13% were not yet assessing them. Only 11% reported enterprise-wide integration. The reported categories do not add exactly to 100%, likely because of rounding or the release’s reporting scope; they should not be forced into a false precision.

Leaders are not waiting for a complete plan to imagine the job changes. In that same survey, 70% expected work redesign to be a high-impact workforce action, 68% named upskilling or reskilling as a priority, and 37% anticipated investing in AI capabilities instead of adding headcount. Seventeen percent identified reducing full-time headcount in AI-affected areas as a top action. These are responses to different questions, not mutually exclusive choices. A company can plan to train one team and stop hiring another.

Removing a recurring draft may free an hour in a role. It does not remove the position. Someone may still review exceptions, explain decisions to customers or teach a junior colleague how to spot an error. The Conference Board’s researchers make the same point in their release: task removal can create responsibilities elsewhere, so saved time need not lower costs. A budget that treats every automated hour as a payroll saving quietly assumes away that remaining work.

An earlier Digidai analysis of task maps and organization charts examined how a job description can hide work that crosses teams. Adecco’s new worker survey adds a different question. If an employee sees the hidden work and the employer asks that employee to help redesign it, is there paid time to do so? A company can draw a more accurate task map and still leave the people who perform the tasks with no capacity to change them.

That is where the worker voice figure earns attention. The reported 96% versus 29% willingness to adapt is not a recipe that will transfer at the same rate to another firm. It is a warning against treating worker participation as a communications step after managers have chosen the new target. Workers know which exceptions remain after a tool handles the easy cases. Their involvement can reveal that the work left behind is more difficult even if the queue is smaller. Without that information, the staffing plan may record a reduction in volume and miss an increase in judgment required per case.

Saved minutes need an owner

An AI tool can save time without deciding what the organization does with it. Adecco’s release says well-integrated users were more likely to reinvest saved time in higher-value work. It does not publish a verified count of hours actually reassigned, the value of the new work, or the wage outcome. The difference matters because three budget stories can begin with the same time-saving claim.

An overhead editorial collage shows an AI tile, a clock, training pages and a cost sheet as alternative uses of saved work time.

AI-generated editorial illustration. The paths represent choices about saved time, not measured results.

One employer keeps headcount steady and raises throughput. Another uses the time for paid learning, quality checks or customer problems that previously waited. A third leaves open vacancies unfilled. All may call the change “productivity.” Their effects on workload, skills, service and employment differ. An organization needs to say which path it chose before it can claim a financial return or tell workers what their jobs now require.

The International Labour Organization’s June 2026 review offers a useful counterweight to survey optimism. It finds that productivity gains can be real yet uneven and often unverified. Worker-reported time savings of a few percent of working hours have not reliably appeared as higher measured output, earnings or employment. The review is not an evaluation of Adecco respondents; it explains why an internal time-saved claim needs an outcome measure before it becomes an investment case.

Adecco’s result about worker involvement also has a cost side. Asking employees to inspect the changed workflow, test examples and explain failure cases requires paid hours. If a manager has to meet the old volume target during that exercise, participation becomes extra work. The company may then observe that only employees with spare capacity took part, making the apparent association between involvement and willingness to adapt even harder to interpret.

The arithmetic can be modest enough to hide. In an illustrative, not reported, planning example, a department with 100 employees that assigns each person eight paid hours of guided practice commits 800 working hours before counting a trainer, backfill or the time managers spend reviewing mistakes. If finance budgets only 100 software seats, those hours surface later as overtime, a slower queue or missed development work. The choice is not automatically to train everyone for eight hours. It is to place learning time somewhere visible, then test whether that amount lets people do the changed work. A team with seasonal demand may have to schedule the practice across several weeks; a shift operation may need paid coverage for people who cannot leave a live station.

Workers also differ in what they can safely delegate. The person who answers a standard question may benefit from an agent’s draft; the person who explains an adverse decision may need to reconstruct every source and judgment. A single course completion rate would miss that distinction. Managers can ask employees to identify the cases where an AI suggestion saves effort and the cases where checking it adds effort. Those answers should alter the workflow before the team sets a universal speed target.

A practical planning file can separate four decisions that an AI business case often compresses into one. This is a proposed operating table for a team, not a model or requirement published by Adecco:

Decision before expansionEvidence to collect in one workflowBudget or staffing ownerResult to check later
Which task changes, and which exceptions remain?Baseline cases, error types and time spent on reviewService or product managerQuality and cycle time by case type
Who gets paid time to learn and redesign?Hours scheduled, coached practice and participation across rolesFunction leader with finance and learning teamDemonstrated skill in live work, not course attendance alone
Where does released time go?Output target, backlog, learning allocation or vacancy decisionFinance and function leader togetherActual capacity, workload and retention
Who can contest a bad workflow?Escalations, employee corrections and customer complaintsAccountable managerCorrected cases and repeated failures

The table is deliberately narrower than a company-wide “AI readiness” score. It asks for a baseline in one workflow, then links a change in work to the person who pays for it and the outcome to be checked. A service team may value faster accurate resolutions. A design team may value fewer rounds of rework. A public agency may have service and due-process requirements that dominate speed. Counting identical minutes across these settings would hide the different work.

Using it should begin with one disputed claim, such as “the agent cuts first-response work in half.” The manager first separates routine replies from cases that require investigation. Workers then record a sample of corrections: did the draft save time, move the time into verification, or create a complaint later? Finance attaches paid training and backfill to the same period, not to a different learning budget. At the end of the period, the group compares service quality and total paid hours, including review. If the work mix changed, it says so in the result. A percentage reduction in draft time may survive that test and still leave the case unresolved if customer satisfaction fell or employees lost time for difficult cases.

There is no virtue in collecting every possible metric. The file should answer the staffing decision at hand. If the proposed action is to leave a vacancy unfilled, measure the work a colleague must absorb and how long the arrangement lasts. If the proposal is to redeploy a worker, record the receiving job, paid learning hours and whether the new work exists. If the proposal is a higher output target, observe the exceptions and overtime after the target changes. Each route has a different denominator. Calling them all “hours saved” makes the decision look more certain than it is.

Workers see a different risk than leaders

The worker and executive surveys were built for different questions. Adecco asks people about use, training, comfort and willingness. BSI asks leaders about employment decisions. In BSI’s November 17 release, 54% of 1,287 surveyed business leaders across nine countries said AI was helping reduce headcount, up from 41% in the 2025 comparison. Forty-nine percent said AI had already reduced entry-level opportunities. These are leaders’ accounts, not payroll records or a causal estimate of jobs removed by AI. The sample, countries and questions differ from Adecco’s worker sample.

The two accounts can belong to the same organization. A leader limits junior hiring; an employee trains an agent and checks its output. The potential entrant loses routine work that once served as practice. A senior colleague keeps the review and mentoring load. The ILO review flags uneven gains and erosion of opportunities for younger workers as risks; it does not report one global jobs result.

Adecco’s own report names a leadership problem. Only 19% of surveyed workers said leaders were empathetic to the effects of change, 21% felt concerns about new technologies were taken seriously, and 39% felt supported by their manager to adapt. Workers whose leaders explained how AI would change their role were much more likely to say they were comfortable with AI-related change, 92% versus 15% among those without that communication. Again, that is a cross-group association. A clear explanation may travel with a better-designed rollout, a stronger manager and more paid learning time.

There is also a retention dimension that does not fit neatly into an AI license budget. Adecco says work-life balance overtook salary as the leading reason respondents would consider leaving their employer, with schedule flexibility and remote-work access next. The release does not show that AI caused this ranking. It does suggest that extracting every saved minute as a new output target may collide with a reason employees say they stay or leave. Training that happens at night can have the same effect even when the formal course is free.

This is a more specific question than whether workers “support AI.” An employee may welcome a tool that eliminates tedious copying while objecting to a higher daily quota and less time to check answers. A manager may appreciate faster drafts and still need a staffing buffer for unusual cases. A worker voice process earns its cost if it identifies those differences before the new target becomes permanent. It is less useful as a survey of sentiment after the target has been announced.

The ILO’s November 2026 policy brief on AI and developing economies argues that exposure, adoption and realized labor outcomes are separate stages. Its recommendations include worker consultation and social dialogue when AI enters decisions about recruitment, scheduling, monitoring, pay, performance or dismissal. Those are policy recommendations across varied economies, not proof that every firm in Adecco’s sample follows them. They identify a sharper boundary for participation: an employee who may be evaluated by the system has a stake in its design that goes beyond a request for usability feedback.

That boundary matters most when the same tool both assists and measures. A drafting aid may report how often an employee accepts a suggestion. The employer could use that count to decide who is “adapting” without checking whether the suggestions were good. Workers who reject poor outputs might then look less productive than colleagues who accept them. A participation process can specify which measurements are appropriate for improvement and which should not become shortcuts for performance judgment. The employer still has to make and document the decision; the survey cannot make it for them.

The design must also fit the employment setting. A salaried professional may have some discretion over when to practice, although a crowded calendar can erase it. An hourly worker cannot simply stretch a shift to absorb the same lesson. A contractor may have no access to the employer’s internal training at all while being judged against a new AI-enabled service target. Adecco’s 36-country survey does not provide a common legal rule for these arrangements. The practical budget question is who has authority to schedule the learning, who pays for it and whether the person expected to use the tool can challenge a bad instruction without losing work. The answer will differ by contract, jurisdiction and task. That is a reason to examine the local workforce, not to apply a global percentage as a staffing formula.

Services firms can count specialists, not the result

Accenture’s fiscal 2026 annual report says it had nearly 110,000 AI and data practitioners, exceeding its original goal to double from 40,000 to 80,000. It says more than 400 clients began advanced AI work with the firm for the first time during the fiscal year. Accenture also describes a $3 billion, three-year AI investment announced in 2023. Those are real commitments and market signals. They do not tell a buyer how much of each worker’s day went to training, whether clients redesigned jobs with employees, or what return an individual deployment earned. A large specialist pool can be necessary for delivery without being a measure of customer outcomes.

Adecco itself has a commercial interest in the answer. The company sells staffing, skilling and workforce services, and its release argues that “hybrid workforce orchestration” is the path to value. Its survey is useful evidence about what respondents report. The interpretation is also part of a vendor’s case for services. The public release gives methodology at the level of sample size, fieldwork dates, countries, industries and online mode, but not a full public questionnaire or firm-level outcomes. A reader should separate the observed answers from the company’s proposed solution.

Adecco has also moved beyond a research claim: its Future Skills Programme, developed with Microsoft, offers free AI learning for jobseekers. That is a company action, but it addresses access to learning outside an employer as much as redesign inside one. The public description does not show how many of Adecco’s surveyed workers took it, whether their employers gave them paid practice, or whether their jobs improved afterward. It cannot be used to explain the survey’s training decline. Its relevance is simpler: the company that reports the gap is selling and building ways to close parts of it, which makes independent outcome checks more important.

The Conference Board recommends continuous workforce planning and decisions shared across functions. Its 11% figure is self-reported by leaders; it is not an audit showing that firms with enterprise plans earned a higher return. BSI describes executives’ perceptions of headcount effects; it cannot settle whether training or participation would have prevented a cut. None of these sources supplies a controlled test of a complete AI redesign program. Each points to a decision an employer needs to measure. None picks a winning consultant.

A function leader need not begin with a 36-country benchmark. Two comparable queues may tell more if the manager records their case mix and baseline quality before changing one. Put paid training hours, worker corrections and escalation rates beside output and cost. If easy cases move to the tool, the human queue may become harder even as its volume falls. And if faster handling simply raises the daily quota, workers may finish later. The result depends on what management did with the minutes.

Every team need not run a formal experiment before using a tool. It does need enough evidence to distinguish a license deployment from a change in work. A firm with a production budget but no learning time may discover that its bottleneck is the manager teaching exceptions. Another can fund courses without giving workers a way to change the workflow they return to. Course completion rises; the old job remains. Both failures cost money beyond the price of a seat.

Put learning time on the agenda

Adecco published its findings after the summer fieldwork, as companies prepared another cycle of AI spending and staffing decisions. A CFO can use the 74% agent-use figure to ask whether employees already have tools. A CHRO can point to the 67% training figure and ask who has been given paid time to learn. A manager can bring the exception list. A worker can say which step now requires more judgment. None of those answers, alone, says whether a job should be removed or a productivity target raised.

At that meeting, the first line item may be unglamorous: hours taken from current production so workers and supervisors can test the new route together. If the organization declines to buy those hours, it should describe the rollout as a tool purchase and measure the consequences. If it pays for them, it can test whether the new workflow actually changes service quality, staffing needs and the worker’s day. Those are different investments with different claims attached.

The November surveys cannot be assembled into one global workforce score. They sample different people, use different questions and record attitudes or reported plans rather than shared outcomes. The next useful number is not another count of agent users. It is a result from a specific workflow: what changed after employees helped redesign it, how much learning time the employer paid for, and who received the hours the tool freed. Without that last answer, a faster task can conceal a harder job.