AI adoption starts with a decision about work: which task is worth improving, what evidence would justify the change, and who remains responsible? A small firm can answer those questions without committing to a firm-wide platform.
A 2026 preprint by Chen and Bao reports a randomised legal issue-spotting experiment with law students. The trained group used the tool more and performed better than the untrained-access group on specified measures; several other differences were uncertain. This supports evaluating instruction as part of adoption, without treating a student examination as evidence of a firm's complete productivity or client outcomes.
Start with a process review
Choose a recurring task, such as an internal briefing, an intake draft or clause extraction. Record the current inputs, handovers, review effort and recurring errors. Compare a shared template, a manual improvement and an AI-assisted workflow. Treat administrative work with the same care: it can contain confidential data and consequential errors.
Review the vendor and the complete service
Check processing locations, retention, training use, subprocessors, support access, deletion and exit arrangements. GDPR does not require all processing to stay in the EU. Where GDPR applies, transfers to third countries require an applicable Chapter V mechanism and any necessary additional safeguards. EU hosting alone does not establish compliance. Professional confidentiality, security and the client's mandate also need separate assessment. GDPR, Articles 28 and 44 to 49. For processing subject to Swiss law, the FADP sets its own principles for disclosing personal data abroad (FADP, Article 16) and its own rules on processors (Article 9). For lawyers, professional secrecy under Article 13 of the Lawyers Act and Article 321 of the Swiss Criminal Code must also be assessed.
Evaluate the vendor's legal-source coverage, languages, citations and update practices against your own tasks. Neither a legal model label nor a general accuracy figure proves reliability for Swiss law. Ask what can be exported and how a reviewer can reconstruct the answer.
Apply the AI Act by purpose and role
A firm established outside the EU, for example in Switzerland, falls within the AI Act only in the cases set out in Article 2, such as placing an AI system on the EU market or providing or deploying an AI system whose output is used in the EU. Ordinary legal research or drafting for a private practice is not automatically an Annex III high-risk use. Assess the intended purpose against the listed categories, such as employment decisions, specified uses by or on behalf of judicial authorities, or similar uses in alternative dispute resolution. Identify whether the firm is a deployer or takes on provider duties. Using a general-purpose AI model does not itself make the firm its provider.
Article 4 on AI literacy has applied since 2 February 2025. Until 26 July 2026 it required providers and deployers to take measures to ensure, to their best extent, a sufficient level of AI literacy of their staff and of other persons dealing with the operation and use of AI systems on their behalf. Since 27 July 2026, as amended by Regulation (EU) 2026/1744, it requires measures to support the development of AI literacy and does not require providers or deployers to guarantee a specific level of AI literacy of any individual. Article 50 transparency duties generally apply from 2 August 2026, with a separate transition for relevant existing provider systems. The relevant Annex III high-risk duties apply from 2 December 2027 and the corresponding product-related duties from 2 August 2028. Existing-system transitions, exceptions and each party's role require individual review. Consolidated AI Act, Articles 2, 4, 6, 25, 50, 111 and 113.
Begin with an approved internal workflow
An illustrative starting point is a briefing based on public supplier notices. A subject specialist checks every material statement against the source. If the draft changes a permission into an obligation, record the defect, correct it and test whether the same error appears elsewhere. The manual template remains available.
Measure preparation, drafting, verification, correction and documentation together. Faster generation alone is not a saving. Capacity released by a pilot does not automatically become revenue or justify reducing review.
Make the change part of everyday work
Name the task owner, permitted inputs, access rights, review responsibilities and escalation route. Let colleagues demonstrate their checking process and record difficulties. Training should include a defective output and an explanation of the correction, rather than attendance alone.
Use a phased adoption roadmap
The following sequence is illustrative, not a promised timetable or a validated acceptance threshold.
Adoption decision checklist
0/7Agree expectations with clients
Clarify the expected deliverable, authorised data processing, quality checks and communication about AI use. Explain who signs off the work. Discuss billing and outcomes using measured evidence from the complete workflow, without promising untested savings.
Key Takeaway
Start with an approved task and a manual comparison. Make business value and responsibility visible in the same decision record. Extend use only when the evidence supports it.
Get in touch to discuss an adoption decision and its review conditions.