The AI Productivity Divide Is Not Adoption
PA Consulting's 9 November 2026 report says senior leaders report more AI value than individual contributors. Daily use is a different cut.
Christian Kosiek•Published •18 min read

A person can open an AI tool every day and still not be the person who gets more from it. That distinction is the point of PA Consulting's report The capability paradox, announced on 9 November 2026. The newsroom page calls PA Consulting a Jacobs company. The report's production line says October 2026. The fieldwork, the report says, took place largely through the first half of 2026. Those are three dates, not one.12
The research combined a survey of 2,000 employees in the United Kingdom, a two-week online diary study with 20 participants, and 13 interviews with senior leaders, AI specialists and subject-matter experts. The questionnaire focused on tools such as ChatGPT, Claude, Gemini, Microsoft Copilot and other generative AI applications. The report says AI-generated synthetic personas were used only to look for gaps in the perspectives represented, and that the findings, insights and quotations come from real participants. This article does not assign each percentage below to the diary or to the 13 interviews. The report does not print a separate base for each figure.2
Two labels that are not interchangeable
For most of the findings, the report uses two cohorts. Individual contributors are people in entry-level, junior and mid-level non-managerial roles. Senior leaders are senior managers and heads of department, directors and senior leadership, and executive/C-suite roles. The report says these groupings are not used where it names a particular role.2
The 9 November newsroom page prints headline percentages against a narrower label: employees in entry-level and junior roles, and in one line "entry-level or junior". The report landing page prints "non-leadership roles". This article does not treat those three phrases as one group. The figures below follow the report unless a sentence says they are the newsroom's or the landing page's.13
What senior leaders report
Seventy percent of senior leaders say AI enables them to accomplish things they could not previously have done, compared with 38 percent of individual contributors. Seventy-two percent of senior leaders report quality improvements in their outputs with AI, compared with 43 percent of individual contributors. The report adds that it is not until management level that more than 50 percent of respondents report an improvement in quality because of AI. "More than 50 percent" is that threshold. It is not the 72 percent, and it is not a rate for every manager.2
These are reported experiences. They are not an experiment showing that seniority causes better AI outcomes. They do not show that AI has made junior employees less capable. The foreword says senior leaders are reporting substantially greater gains in productivity, capability and quality than people in junior, entry-level and non-managerial roles. That sentence uses "reporting".2
The newsroom page reprints 70 percent against 38 percent of employees in entry-level and junior roles, and 72 percent against 43 percent of entry-level and junior employees. It says senior leaders report higher-quality work from AI. The report says "quality improvements in their outputs with AI" and compares them with individual contributors. Pasting the newsroom's "entry-level and junior" onto the report's 38 percent would drop the mid-level non-managerial roles that the methodology includes in individual contributors.12
Usage names particular roles, so the senior-leader grouping is not the line to use. The report says AI usage rises with seniority, from 1.8x a day for entry-level and junior roles to 6.9x a day for executive/C-suite levels. The newsroom says senior leaders use AI more than three times as often daily as employees in entry-level and junior roles, averaging 6.9 uses a day. This article does not treat "senior leaders" and "executive/C-suite" as the same sentence, and it does not divide 6.9 by 1.8. The 1.8x figure is entry-level and junior roles. It is not the whole individual-contributor cohort.12
| What is printed | Figure | What it is not |
|---|---|---|
| Senior leaders who say AI lets them do things they could not previously have done, versus individual contributors | 70 percent versus 38 percent | Proof that seniority causes the result. The newsroom label "entry-level and junior" |
| Senior leaders who report quality improvements in outputs, versus individual contributors | 72 percent versus 43 percent | The separate line that quality improvement passes 50 percent only from management level |
| Uses a day: entry-level and junior roles, then executive/C-suite | 1.8x, then 6.9x | A senior-leader average. The newsroom's "6.9 uses a day" is a different sentence |
| Newsroom: senior leaders versus entry-level and junior | More than three times as often daily, averaging 6.9 uses a day | The report's 1.8x to 6.9x line |
A chart on confidence and on the standard of work excludes responses marked neither or do not know. The bar values did not come through as text, so they are not printed here.2
Kate Spencer is named on the newsroom page as a PA people and change expert. The quotation there uses contractions, so this article paraphrases it. She says AI is often seen as a leveller, and that the research suggests the opposite may currently be happening: people who already have greater expertise and judgement are deriving the greatest benefits, while people earlier in their careers risk losing some opportunities to develop their skills. The report page calls her a workforce transformation expert. This article does not merge the two titles. The sentence is her reading. It is not a measured loss of skill.12
Faster is not the same as value
The report says 60 percent of workers across all cohorts believe AI is helping them work faster, and 79 percent say AI makes them more efficient. Those are two answers. This article does not average them.2
Thirty-seven percent say they spend the same amount of time reviewing and refining work as they would have spent producing it themselves. Thirty-four percent said AI made no difference to their workload. Fourteen percent said AI increased their workload. Almost one third, 32 percent, of workers are yet to report meaningful benefits from AI. "Yet to report" is the report's phrase. It is not "received no benefit". These four lines are not added together. They can overlap, and the report does not say they partition the sample.2
The newsroom page says more than a third, 37 percent, admit to then spending as much time reviewing and refining AI-generated work as they would have spent producing it themselves. The report's bullet does not say "then" and does not say "AI-generated". This article does not treat the 37 percent as a subset of the 60 percent.1
Sixty-eight percent of senior leaders say AI enables them to focus on higher-value tasks, compared with 38 percent of individual contributors. That 38 percent answers a different question from the 38 percent beside the 70 percent. The landing page says 68 percent of senior leaders can focus on higher-value work, compared with 38 percent of employees in non-leadership roles. "Tasks" is not "work", and "individual contributors" is not "non-leadership roles".23
Imagine a small charity preparing a grant application. In this hypothetical example, a junior coordinator drafts a proposal quickly and then spends the afternoon checking eligibility rules, beneficiary figures and funding conditions. A director who already knows the funder can see a weak point sooner. Counting the draft would flatter the tool. Counting the checked application would not. This scene is not a diary participant, not one of the 13 interviews and not a sector statistic. The checks are an editorial illustration.
A different article already separates time saved from who receives the hours, with other sources: AI saved time. Who gets the hours back?. Those sources are not this report.
Trust when the answers disagree
The report's figure caption says one in eight workers trust AI's judgement over their own. The newsroom page prints one in eight employees (12 percent) who say they trust AI's judgement over their own. This article does not turn "one in eight" into a figure it recomputed.12
When AI output disagreed with their own judgement, 18 percent of frequent users said they opted to trust AI. Frequent means five or more AI usages a day. The comparison is 11 percent of medium users, one to four times a day, and 7 percent of infrequent users, less than daily. The newsroom says frequent users were 1.5x more likely to trust AI over their own judgement than medium users. That multiplier is the newsroom's. This article does not calculate it from 18 and 11.12
A second cut is seniority, not frequency. Senior leaders were more likely than individual contributors to trust AI over their own judgement when the two disagreed: 17 percent versus 9 percent. That pair is not the 18, 11 and 7, and it is not the one in eight.2
Dependence is a third question. Twenty percent of respondents admit to feeling overly dependent on AI. Frequent users report dependency rates of 28 percent, compared with 10 percent of infrequent users. The 20 percent is not the 28 percent. Neither one is a trust rate.2
The same chart shows shares who trust their own judgement and shares marked neutral. Those shares did not extract cleanly as text, so this article does not print them.
Who still owns the output
Sixty-six percent of workers believe accountability remains with the output owner, whatever AI's involvement. Thirteen percent feel less responsible for quality when AI produces the bulk of the work. Twenty-one percent are unsure where accountability lies. Twenty-six percent feel uncomfortable signing off work that has been shaped by AI.2
The newsroom page says accountability remains with the person responsible for an output, regardless of AI's involvement. "Output owner" and "person responsible for an output" are not the same phrase. The newsroom also places "a further" before the 13 percent. The report lists the 13 percent as its own bullet. This article does not add 66, 21 and 13, and it does not say they sum to 100.12
The findings show uncertainty about ownership. They are not evidence that workers routinely avoid responsibility.
The report separates a system owner, a workflow owner, an output owner and a decision owner. It says telling employees they remain accountable is necessary and not sufficient, because people may have to use systems they did not select, configure or train. That is the report's argument. It is not a tested compliance method.2
One interviewee suggested that senior leaders should "dedicate roughly 25% of their time to intentional mentoring and coaching to ensure knowledge transfer and judgement skills are passed down". That is one suggestion. It is not a measured share of leaders' calendars, and it is not one of the survey percentages.2
The report includes client stories, among them BT Enterprise, ABN AMRO, National Highways and Schiphol. This article does not use those stories as survey findings and does not import their user counts. A count of users in a client story is not the survey of 2,000.2
Where reported benefit is wider
The newsroom page prints sector percentages that the report body, as read for this article, does not repeat in prose. They may sit in a chart. This article attributes them to the 9 November page.12
Employees who report seeing meaningful benefits: technology, media and telecommunications, 87 percent; financial services, 77 percent; energy and utilities, 75 percent; healthcare and life sciences, 55 percent; the public sector, 53 percent; charities and not-for-profit organizations, 39 percent.1
The report's sector list is wider than those six lines. It names the public sector; professional services; financial services; tech, media and telecoms; retail, hospitality and leisure; energy, utilities and infrastructure; manufacturing, engineering and construction; healthcare and life sciences; and the third sector, non-profit and charity. The newsroom page does not print a percentage for retail, manufacturing or professional services. This article does not invent one. It also does not treat "charities and not-for-profit organizations" as proof that the report's third-sector label is the same cut, even though both point at non-profit work. The survey does not show that nonprofit workers are less capable or that their tools are worse. Budget, work design, data sensitivity, task mix and training access are possible explanations. They are not conclusions the survey establishes.12
Four shifts, and a list the report did not test
The report sets out four shifts: rebuild the pathways to expertise; turn faster work into better outcomes; decide where human judgement belongs; align accountability with control. The newsroom page states four priorities in the same order: opportunities to learn, interrogate outputs, practise judgement and develop skills; measure business value rather than adoption or frequency alone; keep oversight where errors carry greater risk; define responsibility so it sits with the right people. "Practise" is the newsroom's spelling.12
Those are the firm's recommendations. The report does not present them as an intervention it has shown will close the measured gap.
For a small team, an editorial list sits beside that, and it is not a PA finding. Measure a whole workflow, including review and corrections, not only prompts or licences. Pair a less experienced person with someone who can explain why an answer is wrong. Keep some practice on the core task so judgment is not only approval of a draft. Give supervised ownership at a level of risk the person can check. Ask what support people need before treating reluctance as a skills gap. PA Consulting did not test this list.
Neighboring questions already have their own sources on this site: AI training reaches bosses, not beginners, why training time becomes a career gap, who teaches judgment after the first draft and entry-level ads that ask for senior judgment. This article does not reuse their figures.
Job advertisements are a different study
PwC's Global AI Jobs Barometer of 15 June 2026 analysed more than one billion job advertisements in 27 countries and territories. The release says AI is increasing the emphasis on human skills such as judgement, creativity and leadership. The spelling "judgement" is the release's. That is a statement about skills employers describe in advertisements. It is not evidence that a particular employee acquired those skills, and it is not this PA survey.4
The same release has a US entry-level comparison, a wage premium, and figures on company headcount and productivity. They are not reprinted here. The entry-level comparison is already unpacked in the entry-level job that asks for senior judgment. The inbox draft's phrase "stronger demand for human-intensive skills in AI-exposed work, including judgment and leadership" is looser than the release, so it is not this article's finding.
A literacy duty is not this UK survey
For an organization in the European Union, the closest written duty is about learning, and it is not a finding from the 2,000 UK employees. The Commission AI Act Service Desk displays a text it says is based on the EUR-Lex consolidated version as at 27 July 2026. Article 4 is marked there as amended. Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used. The same paragraph says this obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual. A duty to support learning is not a promise that every worker reaches the same level.5
This is not legal advice. EUR-Lex was not opened for this article, and neither was the Digital Omnibus regulation. The wording is the service desk's, read on 22 November 2026.
Human oversight in Article 14, and the deployer duties in Article 26, are written for high-risk AI systems. Article 26(2) says deployers shall assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support. Article 14(4)(b) says the people assigned to oversee a high-risk system are to be enabled to remain aware of the possible tendency to rely or over-rely on its output, which that article calls automation bias, in particular where the system provides information or recommendations for decisions. A general writing tool in the PA survey is not automatically a high-risk system. This article does not apply those two articles to every draft. It also does not state when each article starts to apply. The enforcement FAQ was not opened on this pass.67
How hilighter can help
hilighter does not read the PA file, score who benefits from AI, or decide who is accountable.
What a team can do is write the work down. Map skills by role, including the review a person still does. Compare people with those requirements in skills and talent insights. Record a correction or a finished check in skills evidence and recognition. Set personalized growth paths so practice is not left to chance. A manager's review of that evidence sits in manager assessments. The AI job impact view is indicative. It is not the 70 percent, and it is not the 6.9x line. The Free plan covers up to three job roles. The number of employees is not capped.
Sources
- PA Consulting, AI risks widening the workplace divide, new report finds, newsroom, 9 November 2026. PA Consulting, a Jacobs company. Survey of 2,000 U.K. employees. 70 percent of senior leaders versus 38 percent of entry-level and junior employees on accomplishing things they could not have done before. 72 percent versus 43 percent of entry-level and junior employees on higher-quality work. Senior leaders average 6.9 uses a day, more than three times as often daily as entry-level and junior employees. 60 percent think AI helps them work faster. 37 percent then spend as much time reviewing and refining AI-generated work. 32 percent are yet to report meaningful benefits. 20 percent overly dependent. One in eight (12 percent) trust AI's judgement over their own. Frequent users 1.5x more likely than medium users. 66 percent, person responsible for an output; 21 percent unsure; a further 13 percent less responsible when AI has produced most of the work; 26 percent uncomfortable signing off. Sector benefits: technology, media and telecommunications 87 percent; financial services 77 percent; energy and utilities 75 percent; healthcare and life sciences 55 percent; public sector 53 percent; charities and not-for-profit organizations 39 percent. Four priorities. Kate Spencer paraphrased because the quotation uses contractions. Read 22 November 2026.
- PA Consulting, The capability paradox (PDF, 32 pages), production line October 2026, code 9_571400. Landing page: The capability paradox. Survey of 2,000 UK employees, diary study of 20 over two weeks, 13 interviews. Fieldwork largely through the first half of 2026. Tools named: ChatGPT, Claude, Gemini, Microsoft Copilot and other generative AI applications. Synthetic personas were a research aid only. Individual contributors: entry-level, junior and mid-level non-managerial. Senior leaders: senior managers and heads of department, directors and senior leadership, executive/C-suite. Groupings are not used where a particular role is named. 70 percent versus 38 percent of individual contributors. 72 percent versus 43 percent of individual contributors. Quality improvement above 50 percent only from management level. Usage 1.8x a day for entry-level and junior roles to 6.9x a day for executive/C-suite. 60 percent faster and 79 percent more efficient. 37 percent same review time, without "then" or "AI-generated". 34 percent no workload difference. 14 percent increased workload. 32 percent yet to report meaningful benefits. 68 percent of senior leaders on higher-value tasks versus 38 percent of individual contributors. One in eight workers in the figure caption. When outputs disagreed: 18 percent of frequent users (five or more usages a day), 11 percent of medium users (one to four times a day), 7 percent of infrequent users (less than daily). Senior leaders 17 percent versus individual contributors 9 percent. Dependence 20 percent; frequent users 28 percent versus infrequent users 10 percent. Accountability: output owner 66 percent; 13 percent less responsible; 21 percent unsure; 26 percent uncomfortable signing off. One interviewee on roughly 25 percent of time for mentoring. Client stories not used as findings. Chart bar values for confidence and work standard, and the trust-own and neutral shares, were not printed because they did not extract as clean text. Read 22 November 2026.
- PA Consulting, The capability paradox, landing page. Cards, read from the page markup on 22 November 2026: 70 percent of senior leaders versus 38 percent of employees in non-leadership roles, on things they could not previously have done; 68 percent of senior leaders versus 38 percent in non-leadership roles, on higher-value work; one in eight trust AI over their judgement; 21 percent unclear where accountability lies, while over a quarter feel uncomfortable signing off. "Non-leadership roles" and "higher-value work" are this page's labels, not the report's. The page also says the report draws on a survey of 2,000 workers, interviews and diary research.
- PwC, AI reshapes global labour market into two distinct paths, press release, London, 15 June 2026. More than one billion job advertisements in 27 countries and territories. Increasing emphasis on human skills such as judgement, creativity and leadership. The US entry-level comparison, the wage premium and the headcount and productivity figures are not used here. Read 22 November 2026.
- European Commission, AI Act Service Desk, Article 4: AI literacy. The service desk says the text is based on the EUR-Lex consolidated version as at 27 July 2026 and marks Article 4 as amended. The obligation does not require a guaranteed level of AI literacy for any individual. Read 22 November 2026. EUR-Lex was not opened.
- European Commission, AI Act Service Desk, Article 14: Human oversight, paragraph 4(b), automation bias, for high-risk systems. Read 22 November 2026. EUR-Lex was not opened.
- European Commission, AI Act Service Desk, Article 26: Obligations of deployers of high-risk AI systems, paragraph 2. Read 22 November 2026. The enforcement FAQ was not opened, so this article does not state application dates. EUR-Lex was not opened.
Sources last reviewed: 22 November 2026. The report's individual contributors are not the newsroom's entry-level and junior employees, and they are not the landing page's non-leadership roles. The 6.9x line is executive/C-suite, not the newsroom's senior-leader average of 6.9 uses a day. The 1.8x line is not divided into that 6.9. The 37 percent is not treated as a subset of the 60 percent. The 79, 34 and 14 percent lines are not merged with the 32 percent. The 18, 11 and 7 percent trust split is not the one-in-eight line, not the 17 versus 9 seniority split, and not the 1.5x multiplier. Dependence at 20, 28 and 10 percent is not a trust rate. The 66, 21 and 13 percent lines are not added. Sector percentages are the newsroom's. The 25 percent mentoring line is one interviewee. Client stories are not survey findings. The charity scene and the shop image are illustrations. Neither an AI-detector score nor the absence of Google penalties can be guaranteed.