Updated October 7, 2026. General information, not legal advice.

An employee can be pleased with what AI helps them produce and still feel that their job has become harder.

PwC's UK findings, published on October 1, capture that tension. Among respondents who used AI at work, 70% said it had improved the quality of their work, while 44% reported an increased workload. The survey included 2,023 UK workers. These are employees' assessments, not independently measured changes in output, but they are worth listening to. 1

For HR, that raises a practical question: after introducing AI, what has improved for the team, and what extra work has landed elsewhere? A useful adoption plan needs to answer both. It also needs to take employees' concerns about their roles seriously, without pretending that anyone can promise exactly how the next few years will unfold.

Start with tasks, not job titles

The International Labour Organization's 2025 assessment, covering nearly 30,000 occupational tasks, estimated that one in four workers worldwide held a job with some exposure to generative AI. Its central conclusion was that transformation was more likely than complete replacement. Exposure describes work that the technology could affect; it is not a count of jobs that will disappear. 2

The ILO's June 2026 review of empirical research adds some caution. It found uneven productivity gains and limited evidence of large-scale displacement in the studies reviewed. It also identified concerns about inequality, younger workers' opportunities, and changes to job quality. Limited evidence of widespread losses so far is no guarantee that a particular occupation or employer will avoid them. 3

Broader forecasts need the same care. The World Economic Forum's Future of Jobs Report 2025 projected 170 million jobs created and 92 million displaced by 2030, a net increase of 78 million. Those figures cover several economic and technological forces, not AI alone. Surveyed employers also expected 39% of workers' existing skills to change or become outdated, and 63% identified skills gaps as a major barrier to transformation. These were employer expectations, not observed outcomes. 4

Within a company, a task-level discussion is more useful than assigning an entire job an "automation percentage." Consider a recruiter. Drafting an advert, coordinating interviews, evaluating a candidate, and discussing an offer involve different information and different consequences if something goes wrong. Even a substantial improvement in one activity does not establish that the whole role can be automated.

Ask employees where time goes, what causes rework, and which decisions depend on knowledge that never makes it into the job description. Then identify where AI assistance might help and where a person needs to stay closely involved.

The productivity gains are real, but specific

Some of the strongest evidence comes from a field study of 5,172 customer-support agents, published in The Quarterly Journal of Economics in 2025. Brynjolfsson, Li, and Raymond found that access to an AI assistant increased issues resolved per hour by about 15% on average. Less experienced workers gained more; the most experienced and highest-skilled workers saw small improvements in speed alongside small declines in quality. The researchers also found improvements in aspects of the customer interaction. 5

That is a substantial result for the company and workflow studied. It is not a sensible percentage to apply to every department's staffing plan.

A consultant experiment by Dell'Acqua and colleagues illustrates why. The research, first circulated in 2023 and published in Organization Science in March 2026, found that AI improved performance on tasks within its capabilities but reduced the likelihood of a correct answer on a task outside them. The researchers called this uneven boundary the "jagged technological frontier." Apparently similar assignments could sit on opposite sides of it. 6

For an HR team, the practical response is to test the work with the tools available now. Check whether the result is useful after verification and revision. An assistant that drafts a polished policy summary has not necessarily understood an exception in the underlying policy. Review needs to reach the substance, including facts, omissions, bias, and the consequences of acting on the answer.

Find out what is getting in the way

Before booking another training session, ask what employees are struggling to do.

Deloitte's UK GenAI Workforce Survey, released in September 2026, surveyed 25,000 workers. Nearly a third of generative-AI users said they used it without their employer's knowledge. The responses were collected in May and June, so this is a snapshot from earlier in the year, not a live measure of October behavior. Nor does an employer's lack of knowledge, on its own, prove that an employee breached a policy. 7

It does give HR and IT a reason to ask about work that approved tools or existing guidance may not support. The following checklist separates six problems that can otherwise get bundled together as "low adoption."

Gap What an employee encounters A useful response
Tool gap The approved tool cannot do an important part of the task. Document the missing capability and evaluate alternatives.
Access gap A suitable tool exists, but access is missing or slow to obtain. Fix eligibility, provisioning, and support.
Skill gap The employee can use the tool but struggles to guide it or assess the answer. Practice the actual task with feedback and verification exercises.
Integration gap The tool cannot reach approved information or fit into existing systems. Work with IT and process owners to remove unnecessary handoffs.
Policy gap The employee cannot tell which information or uses are allowed. Give clear examples and a reliable route for questions.
Workflow gap AI has changed one step, but ownership, review, and handoffs remain unclear. Redesign the process with the people doing the work.

Treat these as questions to investigate, not a validated diagnosis or a judgment about the employee. Several gaps may coexist. Low usage can also be appropriate when there is little relevant work for the tool to improve.

For this playbook, useful adoption means repeated, appropriate use that creates value within a supported workflow. Login counts tell you only a small part of that story.

A practical HR playbook

1. Find the work worth improving

Start with two or three role groups. Ask employees to walk through a recent piece of work, including the awkward parts: finding information, waiting for approval, correcting an earlier draft, or moving material between systems. Include people who rarely use AI as well as enthusiastic users.

Choose potential experiments by looking at value, feasibility, data sensitivity, and the consequence of a wrong answer. A task can have strong potential and still need controls that make it unsuitable for an early pilot.

Update role descriptions when the work changes. In a September 2026 BambooHR survey of 1,608 US salaried desk workers, 75% said AI had changed how they did their jobs. The HR-manager subgroup reported that, on average, 43% of employee job descriptions had been updated to reflect AI expectations. These are different survey measures, not an audited comparison within the same teams, but they give HR a useful question to ask locally. 8

A revised description should explain responsibilities and review expectations. Adding "AI proficiency" to every vacancy leaves too much unanswered.

2. Run a small experiment with a proper baseline

Pick a task with an owner, an approved tool, and an output someone competent can review. Before starting, record how long the existing process takes and what acceptable quality looks like. Include the time spent checking and correcting the work.

For example, an HR team could test drafting an internal onboarding checklist from approved, non-sensitive reference material. A reviewer checks every requirement against those references before anyone uses it. Compare the completed checklist with the existing approach, including omissions and revision time.

Agree on stop conditions, such as repeated factual errors or the need to enter information the tool is not approved to handle. Record the tool and workflow tested so a later change does not pass unnoticed. A successful trial justifies a next step; it does not establish that every related task is safe to automate.

3. Make policy usable during the working day

Employees need to know which tools and accounts are approved, what information they can enter, and when another person must review the output. They also need an answer to the question that comes up after a mistake: who should I tell?

Give examples close to their work. Drafting a generic induction email and uploading an employee's medical information are different uses, even inside the same product. Name the route for additional approval, and explain when AI assistance should be recorded or disclosed.

HR's own systems deserve early legal and privacy review. Recruitment, candidate evaluation, and certain worker-management uses fall within the EU AI Act's Annex III high-risk categories, subject to the classification rules and exceptions in Article 6. Having a person approve the result does not automatically remove a system from that category. 9

Under the amended timetable, the main Annex III high-risk requirements apply from December 2, 2027. That extension does not postpone the literacy obligation or existing prohibited-practice rules. AI systems used to infer emotions from biometric data in the workplace are already prohibited, except for medical or safety purposes. 10 11

When a system can act, such as sending messages or changing records, define its permissions and approval points before connecting it. Name someone who can stop it and deal with mistakes.

4. Teach people how to do their work with the tool

A recruiter and a payroll specialist need different examples. Build learning around the decisions each role faces: how to give useful instructions, when to avoid AI, how to verify an answer, and what information can be shared. Give people time to practice during working hours.

In Europe, this is also part of governance. The current Article 4 requires providers and deployers within the AI Act's scope to take measures supporting the development of AI literacy among staff and others operating or using AI on their behalf. Measures must account for people's background and the context of use. The amended article does not require a guarantee that every individual reaches a particular literacy level. 12

The Commission's guidance does not prescribe a universal course or require a certificate. It allows internal records of training and other guidance initiatives. Keep a proportionate record of what you provided, to whom, and how it relates to the work. Course completion alone should not be treated as proof that someone can safely handle every AI-assisted task. 13

Check who gets access to development. The ILO's March 2026 gender analysis found greater exposure in female-dominated occupations, reflecting concentration in clerical, administrative, and business-support work. Plan learning and career pathways with those patterns in mind rather than assuming the same invitation reaches everyone equally. 14

For junior employees, consider what experience they will lose if routine tasks are handed over completely. Retain supervised casework and conversations with experienced colleagues. Ask them to explain why a suggestion is sound and where they checked it.

5. Give managers time and permission to learn too

A manager who encourages experimentation needs to make room for it in the team's workload. Ask for examples of what worked and what needed correction, including cases where the team decided against using AI.

Microsoft's May 2026 Work Trend Index gives some support to this emphasis. Its survey of 20,000 AI-using knowledge workers across 10 markets found that organizational conditions, including manager support and culture, were more strongly associated with reported AI benefits than individual factors alone. These are associations from a vendor-sponsored survey of existing users, not proof that a particular management practice causes productivity gains. 15

In a team conversation, ask whether the completed work improved, how much review it needed, and what blocked progress. Show employees what happened to their feedback.

Be direct about job security. Do not promise that no role will change unless the organization can keep that promise. Explain what has been decided and how employees will be involved in decisions still open. For a listening exercise about unofficial tool use, agree on confidentiality and escalation limits in advance. Encourage good-faith reporting without promising an amnesty the company cannot honor.

6. Measure the whole result

Pair usage information with completed-work measures: quality, total cycle time, rework, and the effort required from reviewers. Ask employees about confidence, policy clarity, tool fit, and whether they can raise concerns. Include workload and development opportunities in the review.

The PwC workload findings are a reason to check what happens to any time released. Decide whether it will support better service, a smaller backlog, learning, or another stated priority. Do not count the same time as both extra productive capacity and a cash saving unless there is evidence for both. 1

Keep employee self-reports separate from observed operational measures. Neither is a substitute for the other. A rise in confidence is useful information, but it does not prove that errors declined. Likewise, more reported near misses may reflect improved reporting rather than more unsafe behavior.

Use aggregated reporting with safeguards against identifying people in small or overlapping groups. Explain what will be collected and who can see it. Avoid individual AI-use leaderboards or turning a support survey into a performance assessment.

NIST's voluntary AI Risk Management Framework provides a useful organizing structure: Govern, Map, Measure, and Manage. These are continuing, connected activities, not a one-time checklist. Review the workflow again when the tool, its permissions, or the work itself changes. 16

A 30-day starting point

In the first week, listen to a few role groups and map their work. Record current AI use, unmet needs, and policy questions. Check that employees understand how their feedback will be used.

In week two, choose one or two low-risk experiments. Set the baseline, name the reviewer, and agree on success measures and stop conditions. Resolve access and data-handling questions before the trial begins.

Use week three for practice and a limited test. Keep examples of errors as well as useful outputs. Ask reviewers about their workload, and give managers time to discuss what the team is learning.

In week four, compare the completed work with the baseline. Fix the most important gap before expanding. Update the relevant guidance or role description, assign an owner to follow-up actions, and set a review date.

A month will not establish long-term return on investment or settle every governance question. It should leave you with a clearer decision about a specific workflow, including a defensible decision to leave it alone when AI adds more trouble than value.

Turn employee feedback into a plan

Behaviture's AI Adoption Pulse helps HR and business leaders establish a survey-based picture of AI use, approved-tool fit, employee confidence, and policy or training gaps. Employees receive private guidance, while leaders receive aggregated findings to help prioritize support and follow-up actions.

The results are a starting point for conversations and experiments. They do not establish individual performance, certify legal compliance, or prove a productivity gain. Used alongside operational measures and the judgment of people who know the work, they can help a team decide what to improve next.

Learn more about Behaviture AI Adoption Pulse.

Sources and further reading

  1. PwC UK. "One in five UK employees are now using AI everyday at work but 44% report a heavier workload." October 1, 2026. UK findings from the 2026 Hopes and Fears Survey; 2,023 respondents.

  2. International Labour Organization and NASK. Generative AI and jobs: A 2025 update. May 20, 2025. Task-based estimates of occupational exposure, not a forecast of actual job losses.

  3. Rossana Merola and colleagues, International Labour Organization. The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence. June 1, 2026.

  4. World Economic Forum. The Future of Jobs Report 2025. January 7, 2025. Employer expectations covering multiple economic and technological forces through 2030.

  5. Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond. "Generative AI at Work." The Quarterly Journal of Economics, 140(2), 889-942, 2025. Field study of 5,172 customer-support agents.

  6. Fabrizio Dell'Acqua and colleagues. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality." Organization Science, published online March 11, 2026. The research first circulated as a working paper in 2023.

  7. Deloitte UK. GenAI Workforce Survey 2026, released September 2026. Ipsos surveyed 25,000 UK workers from May 7 to June 10, 2026. See also the September 16 research announcement.

  8. BambooHR. Redesigning Work: AI's Performance Review, research announcement, September 1, 2026. Online survey of 1,608 US salaried desk workers, including 520 HR professionals at manager level or above; fieldwork June 26 to July 15, 2026.

  9. European Commission, AI Act Service Desk. Annex III, employment and worker-management uses. Read with Article 6 on high-risk classification and its exceptions. Current text checked October 7, 2026.

  10. European Commission, AI Act Service Desk. Article 113: Entry into force and application, reflecting the amended timetable. See also the Commission's AI Act overview. Checked October 7, 2026.

  11. European Commission, AI Act Service Desk. Article 5: Prohibited AI practices, especially Article 5(1)(f). See Article 3 for the definition of emotion recognition. Checked October 7, 2026.

  12. European Commission, AI Act Service Desk. Article 4: AI literacy, as amended in 2026. Checked October 7, 2026.

  13. European Commission. AI Literacy: Questions & Answers. Current guidance on context-specific measures, documentation, and the absence of a mandatory certificate. Checked October 7, 2026.

  14. International Labour Organization. Gen AI, occupational segregation and gender equality in the world of work. March 5, 2026. Analysis using occupational exposure estimates and harmonized microdata covering 84 countries.

  15. Microsoft. 2026 Work Trend Index: Agents, human agency, and the opportunity for every organization. May 5, 2026. The main survey covered 20,000 AI-using knowledge workers in 10 markets; reported relationships are associations, not experimental estimates of causation.

  16. National Institute of Standards and Technology. AI Risk Management Framework: Core. The voluntary framework organizes risk management around Govern, Map, Measure, and Manage.