by Claude Opus 5.5

What’s the most useful way to distinguish jobs, tasks, and workflows when discussing AI’s impact on employment?

Treat them as three nested levels, each with its own question. Tasks are what AI acts on. Workflows are where an organisation decides whether a task-level gain becomes more output, better output or fewer people. Jobs are what employers contract and pay for, and what official statistics count. Most bad arguments about AI and work go straight from “this task is exposed” to “this job will go” and skip the workflow in between.

Three levels, three different questions

  • Task. Working definition: A unit of work with an identifiable output: drafting a letter, coding an invoice, summarising a case file. The question to ask: Can AI do this, or do it faster, to an acceptable standard? How it is usually measured: Exposure studies that score task lists.

  • Workflow. Working definition: The sequence of tasks, hand-offs, checks and approvals that delivers a result to a customer, patient or regulator. The question to ask: When this task gets cheaper, what happens to the steps around it? How it is usually measured: Hardly at all, outside firm case studies and trials.

  • Job. Working definition: A bundle of tasks plus accountability, pay, a contract and a career identity. The question to ask: Does the remaining bundle still justify the role, the headcount and the pay? How it is usually measured: ONS labour-market surveys and payroll data by occupation, plus job adverts.

The last point in each item, how it is usually measured, is the useful part. Researchers measure tasks, and national statistics measure jobs. Almost nobody systematically measures the workflow level, which is where the outcome is actually decided. That gap explains why exposure estimates and employment data so often seem to disagree.

Tasks: the right starting point, but not the answer

The task approach comes from Autor, Levy and Murnane (2003). They showed that computers replaced routine, rules-based tasks and complemented non-routine problem-solving, and that much of the resulting shift happened within occupations rather than between them. Generative AI is analysed the same way. Eloundou and colleagues estimated that around 80% of the US workforce could have at least 10% of their tasks affected by large language models, and about 19% could have at least half their tasks affected. For the UK, Coface’s April 2026 scenario map puts nearly 20% of UK tasks in the exposed category, among the highest of the advanced economies (as reported).

These are measures of where change is possible. They don’t predict job losses, because a job is rarely one task, and because what an organisation does with a cheaper task is a separate choice.

Workflows: where the outcome is decided

Three UK examples show why the middle level matters.

GP documentation. Ambient voice technology listens to a consultation and drafts the clinical note. The government’s 10-year health plan estimates that “saving just 90 seconds on each appointment would generate over 2,000 full time equivalent worth of GP capacity”. That is a task-level saving, and an assumed one. What happens next depends on the workflow. If practices book more appointments, output rises. If GPs take longer over complex patients, quality rises. If nothing is redesigned, the time disappears into the day. The tool also adds tasks: checking the note, recording consent and governing the vendor. With unmet demand in general practice, a fall in GP employment is close to the least likely outcome.

Time saved is not output. In a 2024 Department for Business and Trade trial of Microsoft 365 Copilot, staff saved time on some tasks, especially written work, but the evaluation “did not find evidence that time savings have led to improved productivity”. The task improved, but the workflow didn’t change, so nothing measurable followed. The cross-government trial produced the more widely quoted figure of 26 minutes a day, but that figure was self-reported.

Work that disappears upstream. Centrica’s July 2026 decision to cut 1,300 British Gas call-centre roles was linked by the company to a 20% fall in call volumes. Its chief executive said “AI isn’t driving these particular job reductions; that’s mainly due to changing customer behaviour”, and unions dispute that. Either way, the lesson for analysis holds. Nobody automated the task “answer a call”. The workflow changed so that fewer calls happened at all. Job losses that come from redesigned workflows are much harder to attribute than direct substitution, which is one reason the debate about AI’s role in 2026’s weak hiring remains unresolved.

Jobs: bundles held together by accountability

A job survives losing tasks if what remains still needs a person, and above all a person who can be held responsible. Autor and Thompson (2025) add a useful twist. When automation removes the less expert tasks in an occupation, the remaining work becomes more specialised, so wages tend to rise and employment to fall. When it removes the more expert tasks, the job becomes easier to enter, so wages fall and employment can rise. The same exposure score can therefore point in opposite directions for pay and headcount.

UK entry-level data illustrate the bundle problem. The DSIT and LinkedIn snapshot (June 2026) found entry-level accountant hiring down 29% on the year to April 2026, against 14% for hiring overall. The tasks that make up junior accountancy, such as reconciliations, first-pass analysis and document review, are highly exposed. The partner’s sign-off is not. The job title “accountant” survives, but the bottom of the bundle is thinning. The snapshot itself cautions that “further research is needed” before attributing this to AI.

A three-question test for any claim

When you meet a claim about AI and employment, push it through the three levels:

  1. Task: which specific tasks are affected, and what share of the job’s time and value do they make up?

  2. Workflow: when those tasks get cheaper, does the organisation remove steps, increase throughput, add checking and compliance, or move work to another team or to the customer?

  3. Job: after the change, who is accountable, and does the remaining bundle still add up to the same role, a different role or no role?

Most confident predictions fail at step 2, because they assume the organisation responds in one particular way.

Bottom line

Use tasks to locate where change is possible, workflows to judge whether it will create value, and jobs to read the employment result. In 2026, the clearest UK signals show up at the job level, mainly as slower hiring into junior roles. The evidence that would explain them sits at the workflow level, and it is the least measured of the three.

Sources

From AI and Jobs: UK, October 2026