AI and Work: What the Evidence Shows, Not What the Headlines Say
Coverage of AI at work runs to two templates: everything is about to be automated, or none of it is real. The published evidence supports neither, and it is worth reading carefully if you are trying to work out what to do about your own role.
Adoption Is Wide, But Shallow
The best evidence on how much AI is actually being used at work comes from national statistical offices rather than vendor surveys. The Office for National Statistics reports that self-reported AI use among UK businesses with ten or more employees rose from around 12 per cent in late 2023 to around 35 per cent, drawing on the Business Insights and Conditions Survey. That is a large increase over a short period. The more revealing figure sits beside it: the average number of distinct AI technologies used per adopting business rose only modestly over the same period, from around 1.4 to around 1.6.
In other words, adoption has spread across many more firms without deepening much within them. Most adopters are doing one or two things with AI, not rebuilding how they operate. The ONS also finds wide sector variation, with over half of businesses in information and communication reporting use against far lower levels in construction, and notes that improving existing operations is the most commonly reported purpose. Crucially, it records that this has not yet translated into widespread changes in overall workforce headcount.
Ireland shows the same shape at a lower level. The Central Statistics Office reported that just over 20 per cent of enterprises used AI in 2025, up from around 8 per cent in 2023, with large enterprises far ahead of small ones. Broad, fast, shallow, and heavily size-dependent.
Exposure Is Not Replacement
The most widely quoted research on which jobs are affected is the joint ILO and NASK work published in 2025, which built a refined global index of occupational exposure to generative AI. Its headline is that roughly one in four jobs worldwide is potentially exposed.
That number is routinely reported as though exposure meant replacement. The authors say the opposite. Their framing is that transformation of tasks, not elimination of jobs, is the most likely outcome for the great majority of exposed roles. The share of global employment falling into their highest exposure category is small, around 3 per cent, and clerical occupations dominate it: data entry, typing, bookkeeping and administrative support.
The study also finds a clear gender pattern, with the highest exposure category accounting for a noticeably larger share of female than male employment, and the gap widening in high-income countries. That is one of the more robust and least reported findings in the whole literature.
The distinction matters. Exposure is a statement about tasks within a job, not about the job itself, and it is entirely normal for several tasks to change while the role persists in a different shape.
The Field Experiments Point in More Than One Direction
Three studies are worth knowing, because they disagree in instructive ways.
Customer support. Brynjolfsson, Li and Raymond studied the staggered rollout of a generative AI assistant across 5,179 customer support agents, published in the Quarterly Journal of Economics. Access to the tool raised issues resolved per hour by around 14 per cent on average, but the average concealed the interesting result: novice and lower-skilled workers improved by roughly a third, while experienced high performers gained little. The tool was effectively distributing the practices of the best agents to everyone else.
Consulting. Dell'Acqua and colleagues, working with Boston Consulting Group, ran a pre-registered experiment with 758 consultants. On tasks within what the authors called the jagged technological frontier, AI-assisted consultants completed more tasks, considerably faster, at higher assessed quality. On a task deliberately designed to sit outside the model's capability, accuracy fell sharply relative to the control group, because participants trusted confident output that was wrong. The lesson is not that AI helps or hurts. It is that the boundary between the two is irregular and gives no warning signal.
Software development. METR ran a randomised controlled trial with sixteen experienced open-source developers across 246 real tasks in their own large repositories. Developers were 19 per cent slower when permitted to use AI tools, while estimating afterwards that they had been about 20 per cent faster. METR itself now labels the finding historical, noting it reflects the tools and workflows of the period studied.
Taken together these say something coherent: gains concentrate where the user is less expert than the tool, losses concentrate where the user is more expert than the tool but does not realise where the boundary lies, and self-perception is an unreliable guide to either.
Which Tasks Are Actually Changing
Across the adoption statistics and the experiments, the tasks consistently affected are those with high volume, defined structure and a checkable output: first drafts, summarisation, routine correspondence, data entry and reformatting, standard code, and routine customer queries. The least affected require accountability, physical presence, negotiation, or judgement under ambiguity.
Where the Evidence Is Genuinely Thin
Be honest about the limits, because a lot of confident commentary is not.
- Headcount effects are not established. The ONS explicitly reports that adoption has not yet produced widespread headcount change. Attributing any particular round of redundancies to AI rather than to trading conditions is, at present, mostly assertion.
- Adoption measures are self-reported and depend on how firms interpret the word AI. Cross-country comparisons are indicative only.
- Exposure indices are estimates of technical potential, not forecasts. They do not model cost, regulation, liability, workplace culture or plain organisational inertia, all of which slow things considerably.
- Field experiments date quickly. Each of the studies above tested particular tools at a particular moment, as METR itself points out, and publication effects favour striking results in both directions.
What This Means for Your Own Job
The reasonable reading is that the near-term risk to most professionals is not being replaced. It is being outperformed by a colleague who uses these tools well on the same tasks, combined with a slow reweighting of what a role is mostly made of. The defensible response is to identify which of your own tasks are high-volume, structured and checkable, learn to do those faster with assistance, and get better at the ones the evidence says are least exposed: judgement, accountability, and the parts of the work where somebody has to be answerable.
Our AI & Digital Skills series is built around exactly that: practical, task-level application for working professionals, across fifteen titles including dedicated guides to ChatGPT, Claude and AI for Teachers.