by Claude Opus 5.5
In what kinds of tasks does AI most often substitute for human labour versus augment it, and what conditions predict each outcome?
AI tends to substitute where a task is high-volume, digital, well specified and cheap to check, and where a mistake is cheap to fix. It tends to augment where quality is hard to verify, errors are costly, the context is tacit, or someone must be accountable for the result. Technology does not settle the outcome alone. The cost of labour, how elastic demand is for the output, and whether the firm redesigns the workflow all decide whether the same task goes one way or the other. Augmenting senior staff can also substitute for junior ones.
Tasks that tend to be substituted
The clearest substitution is in work that is mostly text or data in and text or data out, done at volume to a known standard:
First-line customer contact and routine administration. Bank of England staff analysis on the Bank Underground blog (August 2026) found vacancies in customer-service and administrative occupations down 23% and 22% over about three years, the steepest falls of any group. It is staff research, not Bank policy.
Back-office processing, such as reconciliations, document extraction, standard compliance checks and data entry. Much of this was offshored in the 2000s for the same reason it is now automatable: it can be specified and checked remotely. Standard Chartered’s planned cut of about 7,800 back-office roles by the end of the decade is the largest UK-linked example in 2026 where the company itself framed the change as replacement.
Commodity content. Hui, Reshef and Zhou studied a large online freelance platform. After ChatGPT’s release, writing freelancers saw monthly jobs fall 2% and earnings fall 5.2%. After image-generation tools arrived, image freelancers saw jobs fall 3.7% and earnings 9.4%. Better-rated freelancers were not protected.
These tasks share one feature: the output can be judged quickly and cheaply. A drafted standard reply, an extracted field or a stock image is either acceptable or it isn’t, and someone can tell in seconds.
Tasks that tend to be augmented
Augmentation dominates where a person still has to decide what “good” means, and be answerable for it:
Professional analysis and advice. In Dell’Acqua and colleagues’ field experiment with 758 BCG consultants, those using GPT-4 completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. On a task deliberately set just outside the model’s competence, however, they were 19 percentage points less likely to get the right answer. The authors called this the “jagged frontier”. People who can tell which side of the frontier they are on are the ones who stay valuable.
Software development. In a controlled trial, developers using GitHub Copilot completed a programming task 55.8% faster (Peng et al., 2023). The person still specifies requirements, integrates, tests and owns security.
Clinical work. NHS adoption of ambient scribes and AI-read imaging changes how clinicians spend their time, but the diagnosis, consent and liability stay with them.
Five conditions that predict the outcome
Cost of verifying output. Leans towards substitution: Cheap, fast, objective. Leans towards augmentation: Expensive, slow or a matter of judgement.
Cost of an error and who bears it. Leans towards substitution: Low, reversible, internal. Leans towards augmentation: High, irreversible, legal or reputational.
Separability of the task. Leans towards substitution: Can be cut out of the workflow cleanly. Leans towards augmentation: Tangled with other tasks, context and hand-offs.
Demand for the output. Leans towards substitution: Fixed (inelastic). Leans towards augmentation: Expands when cost falls (elastic).
Relative cost of labour. Leans towards substitution: Labour expensive relative to software. Leans towards augmentation: Labour cheap, or tasks need physical presence.
Some practical notes on the conditions:
Verification is the hinge. If checking an AI output costs nearly as much as producing it, the human stays. This is why first-draft legal research is augmented and document classification is substituted.
Regulation keeps people in some loops. Since 5 February 2026, the Data (Use and Access) Act’s new automated decision-making framework has allowed significant decisions based solely on automated processing, but only with safeguards. People must be told, must be able to contest the decision and must be able to get human intervention. Special-category data stays more tightly restricted. In hiring and HR, that pushes employers towards a “human reviews AI” design even where full automation is technically feasible. (This is general information, not legal advice. Take advice on specific deployments.)
Labour costs tip marginal cases. The Bank of England’s February 2026 Monetary Policy Report said higher employer National Insurance contributions and the National Living Wage had weakened employment relative to output. Its Agents reported firms planning to meet demand “by investing in automation and AI, rather than by raising headcount”. When a job costs more, a task that was borderline for automation becomes worth automating. Acemoglu and Restrepo warn that this kind of “so-so automation” displaces workers without much productivity gain.
Workflow redesign turns augmentation into substitution. A tool that helps every agent handle more cases is augmentation at the task level. If the firm then redesigns the workflow so that AI handles routine cases end to end and people handle only exceptions, it becomes substitution at the job level. Which path firms take is a management decision, not a technical one.
Augmentation can still reduce jobs
Two cautions matter for interpreting 2026 evidence.
First, if four people with AI can do the work of five and demand for the output doesn’t grow, the fifth job goes, even though every remaining worker was “augmented”.
Second, augmentation is often unequal across seniority. If AI makes an experienced professional much faster at the drafting and checking a junior used to do, the senior is augmented and the junior is substituted. The Stanford “Canaries in the Coal Mine” update (August 2026) finds US employment of 22–25-year-olds in highly AI-exposed occupations about 19% below comparable peers. The declines are concentrated where AI automates tasks, while employment is flat or rising where it complements workers, particularly experienced ones. The adjustment works mainly through reduced hiring of young workers, not layoffs. The authors stress this is descriptive, not proof of causation. The UK equivalent, the DSIT and LinkedIn snapshot, finds employers “favouring experienced candidates” in declining entry-level occupations.
Bottom line
To predict which way a task will go, ask whether its output can be checked cheaply, what a mistake would cost and who would pay, whether demand would grow if it got cheaper, and whether the firm intends to redesign the workflow around it. The early UK pattern fits these conditions, though the Bank’s own staff say confident attribution to AI is premature. Customer-service and admin vacancies are falling fastest, and professional work is being augmented. Some of that professional augmentation, though, comes at the expense of the junior rungs.