The workflow, not the model, decides how much AI does.
Anthropic’s June 2026 Economic Index report measures how much of a task people actually hand over, and concludes that the tool people work in probably matters more than the model behind it. For European and Swiss firms, whose most common uses of AI are the ones with the least to delegate, that moves the question from procurement to process design.
Die Insights erscheinen auf Englisch.
Producing one article in a conversational AI interface takes a median of thirteen rounds of back-and-forth. Producing the same output in an agentic session takes a single human prompt.
The comparison comes from Cadences, the June instalment of Anthropic’s Economic Index, published on 26 June 2026. It measures not model quality but how much of a task people actually hand over — and for the same finished output, that varies sharply with where the work is done.
Most management teams still ask which model to standardise on. For a company deciding what to buy, the more useful number describes something less visible and more controllable: the workflow built around the model.
Anthropic measured the same output on two surfaces and found two different divisions of labour
The report analyses conversations on Claude’s chat and Cowork surfaces sampled between 10 April and 10 June 2026, using a classifier that labels the concrete output of each session — a document, a script, an application. Claude Code sessions from the same period are included for the comparison.
Each conversation is also rated for AI autonomy on a scale of one to five, from none to extreme. The lowest-autonomy outputs are calculations, translations and question-and-answer exchanges, where the specification largely determines the result; the highest are applications and websites, games and presentations, which require selection among many possible choices.
Across 26 of the 31 output types measured, autonomy is higher on the agentic surface than in chat, by 0.37 points on average and 0.53 for scripts and code snippets. Blog posts and articles, where the underlying requests are similar on both surfaces, supply the thirteen-to-one contrast.
Two thirds of the gap is behaviour, and it survives holding the model constant
Anthropic decomposes the difference: roughly two thirds comes from the same tasks executed with more delegation on the agentic surface, the remaining third from the different mix of outputs produced there.
The obvious objection is model choice: 54% of agentic sessions ran on Opus, against 10% of chat and Cowork conversations. Among conversations served by Sonnet on both surfaces, the agentic one still shows 0.26 points more autonomy — from which the authors conclude that “the product used is likely more important than the underlying model”.
This is not a blanket rule: for data and spreadsheets, chat sessions score higher than agentic ones (3.09 against 2.74), some 70% of it compositional. The surface does not add autonomy by itself; it changes what people ask for and how completely they specify it.
Delegation also tracks cost. Across output types on chat and Cowork, mean autonomy and median token use rise together (r = 0.68): the outputs people hand over most fully consume the most compute. That is an association between output types rather than a price for delegating a given task, but it belongs in the budget.
European and Swiss firms are concentrated at the low-delegation end of the scale
Twenty percent of EU enterprises with ten or more employees used AI technologies in 2025, against 13.5% in 2024, according to Eurostat’s ICT usage survey published on 11 December 2025: 17.0% of small firms, 55.0% of large ones. Where firms point the technology matters more than how many use it: among AI-using enterprises, marketing or sales comes first at 34.7%, business administration and management at 31.1%, logistics far behind at 6.1%.
The OECD’s 2026 D4SME Survey is blunter about depth: among its respondents, 75% use off-the-shelf AI applications, 5% customised ones and 3.6% agentic ones, and 56.6% use AI for isolated tasks only. Where agentic systems are deployed, the OECD finds them running multi-step workflows: bookings, invoicing, campaign management.
Switzerland follows the same profile. The SME labour-market study published by AXA Switzerland on 8 October 2025 — Sotomo, 300 SMEs in German- and French-speaking Switzerland, fieldwork 3–10 March 2025 — found 34% of firms deliberately integrating AI into work processes, against 22% a year earlier. The leading applications were translation (52%) and correspondence (47%), ahead of workflow optimisation (34%) and data analysis (32%).
Read together, these figures make the picture precise rather than discouraging. Translation sits among the three lowest-autonomy outputs in Anthropic’s ranking. The most common Swiss use of AI is therefore the one in which delegation is structurally capped — not because the model cannot do more, but because the task leaves it almost nothing to decide.
The optimism finding comes with the report’s own caveats
The third chapter draws on the Anthropic Economic Index Survey, launched on 22 April 2026 and linked to usage data from mid-May to early June by a privacy-preserving method; the linked sample is about 9,700 people.
Across all six dimensions of job quality tested — pay, job security, ability to find a new job, meaning, autonomy and human interaction — respondents with a higher share of automated sessions were more optimistic about the coming year, with the largest effects on pay and on finding a new job. Large majorities report gains in speed (86%), scope (82%) and quality (69%); 57% say AI has made their skills more valuable, a share that rises with delegation.
The caveats are the authors’ own. The sample is representative of no workforce: it is drawn from Claude users, and computer and mathematical occupations account for roughly 30% of respondents against 4% of US employment. The results are associations, not causal effects, and the report states that its self-assessments do not rule out skill erosion. In the same survey, 10% rated losing their own job next year as likely or very likely, and more than a third put the probability above 60% for a junior colleague.
In Switzerland, the level of delegation is also a legal setting
Article 21 of the revised Federal Act on Data Protection, in force since 1 September 2023, applies where a decision is based exclusively on automated processing and has a legal consequence for the data subject or a considerable adverse effect on them. The controller must inform that person, who may express their point of view and may require review by a natural person, subject to two exceptions: a decision directly connected with a contract where the request is granted, and explicit consent.
The practical translation is narrow. Swiss law does not regulate how many rounds a workflow takes; it regulates the last one. A process may run end to end, but where its output is a decision about a person — a credit limit, a shortlist, a claim — a human review point must exist and be reachable. Designing the handover includes designing where it stops.
Governance lags: in the AXA study, only about a third of AI-using Swiss SMEs reported clear data-protection rules for AI applications, and 23% among firms with five to nine employees.
A management team can settle the delegation question before it buys anything
- Take one recurring output and count the rounds. Thirteen is the published median for an article; the figure for a quotation or a monthly reconciliation takes an afternoon to establish.
- Write down once what those rounds keep re-supplying: the brief, the context, the definition of a finished result, one acceptable example. Two thirds of the measured gap comes from the same task handed over differently.
- Decide the mode for each output type. Anthropic’s classifier distinguishes directive delegation, feedback loops, task iteration, learning and validation; only the first two count as automated use. That choice is managerial, not technical.
- Match ambition to output type. Translation and question-and-answer work will not become agentic work; the gain there is speed. The candidates are multi-step processes — order confirmations, invoicing, campaign preparation — where the OECD observes agentic systems in use.
- Budget compute, not only licences. Across output types, autonomy and token consumption move together, so a workflow meant to run with little supervision should be priced on consumption before it is adopted.
- Fix the review point before the first run, and check whether Article 21 makes it mandatory. Write the data rule at the same time: two thirds of AI-using Swiss SMEs still have none.
Anthropic runs its survey monthly; Eurostat’s enterprise ICT survey and the Swiss SME study are annual — three places where a shift in delegation would become visible. The measurement that matters inside a company is its own: human rounds per finished output, tracked over a quarter. If that number does not fall, changing the model is unlikely to move it.