AI in customer service — the first high-ROI use case for small companies.

In May 2025 the Quarterly Journal of Economics published “Generative AI at Work” by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, a study of the staggered introduction of a generative-AI conversational assistant across 5,172 customer-support agents. Access to AI assistance raised productivity, measured as issues resolved per hour, by 15% on average.

That figure has travelled widely; the distribution behind it has not, and the distribution decides whether a small company gets a return or an expensive chat window.

The average gain of 15% conceals two opposite outcomes

In the study the agent still answers the customer; the system reads the conversation and proposes a reply in real time. Less experienced and lower-skilled agents improved both the speed and the quality of their output; the most experienced and highest-skilled recorded small gains in speed and small declines in quality. The authors offer suggestive evidence that the assistant disseminates the best practices of more able workers and helps newer workers move down the experience curve — a transfer of accumulated practice, not an increase in individual capability.

Two further results matter as much. Customers became more courteous and less likely to ask to speak to a manager, and the gains were largest for moderately rare problems, where agents have less baseline experience but the system still has adequate training data.

The gain is a transfer of experience, which is not the same thing as a smaller team

If the mechanism is knowledge transfer, the value is largest where experience is thin: a firm where two or three people answer everything, and a departure takes the answers with the person. It also sets a precondition no licence supplies — the assistant can only transfer what has already been resolved or written down.

Swiss evidence shows why the design choice deserves attention. In KOF Studien No. 186, published in October 2025, Jeremias Kläui and Michael Siegenthaler of the KOF Swiss Economic Institute at ETH Zürich ranked occupations by exposure to large language models, using SECO unemployment register data and near-complete job-advertisement data. After autumn 2022, registered jobseekers in the most exposed occupations rose by up to 27% more than in the least exposed. Contact-centre information clerks are among the thirty most exposed occupations the study charts individually, although the sharpest relative increases are in software, recruitment, journalism and marketing roles. The authors stress that this measures relative shifts between occupations, not total employment, which grew by 7.4% between the third quarter of 2020 and the second quarter of 2025. Assistance and substitution are not the same experiment, and a company chooses which one it is running.

Customer service pays off first because it is the function a company already counts

Most small firms point AI elsewhere. In the 2026 OECD D4SME Survey — 2,018 responses across twelve countries — 75% of respondents use off-the-shelf applications, and among those users 70% apply them to marketing content, 56% to drafting documents and 34% to translation or language support. Eurostat does not count customer service as a purpose at all; the closest category, marketing or sales, was reported by 34.70% of AI-using EU enterprises in 2025, ahead of every other.

Content generation is a poor place to prove anything: the output is judged subjectively and the counterfactual unknown. Customer service differs because the measurement already exists: the operational metrics are recorded before any AI arrives, so one period can be compared with the same period a year later.

Language adds a Swiss reason: serving customers in German, French, Italian and English costs coverage that scales with staff rather than volume. Klarna, describing its own deployment on 27 February 2024, reported 2.3 million conversations in the first month across 23 markets and more than 35 languages, average resolution falling from 11 minutes to under two, and that customers could still choose a human.

Disclosing the machine changes the result, and in the European Union it is now required

Disclosure is not a neutral formality. In a field experiment published in Marketing Science in 2019, Luo, Tong, Fang and Qu randomised more than 6,200 customers of a financial services company in Asia to outbound sales calls from chatbots or human workers. Undisclosed chatbots were as effective as proficient workers and four times more effective than inexperienced ones; disclosing the chatbot’s identity before the conversation reduced purchase rates by more than 79.7%, because customers judged the disclosed bot less knowledgeable and less empathetic. The setting is outbound sales on 2019 technology, but it isolates the variable European law has now fixed.

Article 50 of the Artificial Intelligence Act became applicable on 2 August 2026. Under Article 50(1), providers must design systems that interact directly with people so that those people are informed they are dealing with an AI system, unless that is obvious to a reasonably well-informed and observant person. The Commission’s guidance names chatbots, AI agents and avatars, requires the information from the start of the first interaction, and reads the “obvious” exception narrowly. Infringements carry fines of up to €15 million or 3% of worldwide annual turnover.

Two features bring this to Swiss companies. The duty falls on providers — those who develop a system, or have one developed, and put it into service under their own name or trademark — so a company running an assistant built for its own brand carries it directly, while one using a standard product depends on how its supplier designed it. And Article 2(1)(c) extends the Regulation to providers and deployers established in a third country where the system’s output is used in the Union: a Swiss company whose chat window answers a customer in Munich is in scope. A disclosed assistant cannot compete on being mistaken for a person. It competes on answering quickly, in the customer’s language, and handing over cleanly.

A Swiss deployment answers to a second rulebook: where the conversation data goes

Support conversations are personal data, and the revised Federal Act on Data Protection, in force since 1 September 2023, governs them whatever Brussels decides.

• Article 19 requires the controller to inform the data subject when collecting personal data: identity, purpose and, where applicable, recipients. If the data goes abroad, the notice must also name the state and the safeguard relied on.
• Article 16 allows cross-border disclosure where the Federal Council has recognised adequate protection, or under a safeguard such as standard clauses approved by the FDPIC. On 14 August 2024 the Federal Council amended the Data Protection Ordinance to recognise US companies certified under the Swiss–US Data Privacy Framework, in force from 15 September 2024.
• Article 21 applies where a decision is taken exclusively by automated processing and has a legal consequence or a considerable adverse effect. The controller must inform the person, allow them on request to state their point of view, and let them request review by a natural person. Answering a question is not such a decision; declining a refund end to end can be.

Together they describe from the data side what Article 50 describes from the customer side.

What a management team can decide before it buys anything

The evidence supports a short list of decisions, all prior to the choice of software.

• Record the baseline — response time, resolution rate, reopen rate and out-of-hours volume, on definitions written down beforehand.
• Assemble the corpus before the licence: resolved conversations, standard answers and product documentation are what the assistant has to work with.
• Write escalation as a list of triggers, not a sentiment threshold — money, complaints, contract changes, safety and any second failed attempt go to a person; publish how a customer reaches one.
• Decide the disclosure sentence, shown at the start of the first interaction, in the customer’s language.
• Settle the data path before the tool: which vendor, established where, under which Article 16 safeguard, with the privacy notice updated.
• Keep decisions with legal or adverse effect outside the automated loop, or build the review route Article 21 requires.

Two later dates belong in the calendar: the Act’s deferred high-risk obligations — 2 December 2027 for the stand-alone cases in Annex III, 2 August 2028 for AI in regulated products — and Switzerland’s consultation draft, due by the end of 2026.

If a customer writes tonight at ten, what happens, and what will the same measurement show a quarter from now?