A useful strategy for legal AI begins with a service the firm wants to improve and evidence about what clients need. It should not depend on a claim that every boutique will win, every large firm will lose or a fixed adoption window is closing.
Use a forecast to frame questions
The third edition of Richard Susskind's Tomorrow's Lawyers, published in 2023, explores changing legal services, technology, access to justice and new work and training roles. Its forecasts can help frame a strategy discussion. They do not establish the current structure of the Swiss market or a guaranteed return from a particular product.
Treat specialisation, templates, monitoring services and different pricing arrangements as options to evaluate. For each, identify a client need, a feasible delivery process and the evidence required before investment or expansion.
Pair productivity tests with learning tests
Schwarcz and colleagues' 2026 randomised study found benefits from specified tools on bounded tasks performed by advanced law students. It does not show that an entire firm can reduce staffing or safely expand into unfamiliar work.
A 2026 preprint by Chen and Bao studied training for AI-assisted issue spotting among law students. Training improved some measured outcomes, while other comparisons remained uncertain. It supports investigating how training changes tool use, not assuming that any short course establishes professional competence.
For a practice pilot, assess whether people can verify sources, detect omissions and explain their decisions as well as complete work. Keep opportunities to practise those skills, including without AI where appropriate. Faster output is not a substitute for evidence of learning.
Plan around decisions, not promised deadlines
Ada Studio proposes the sequence below. Progress depends on the evidence and risk of the task, not a universal number of months.
Define
Choose the service
Identify the client need, task boundary, baseline and responsible owner.
Test
Compare complete work
Check quality, omissions, full effort, data handling and staff understanding against the existing process.
Decide
Extend, revise or stop
Assess evidence against agreed criteria, record unresolved issues and set monitoring for any permitted use.
Begin with representative material that is permitted in the test environment. A polished demonstration or a handful of favourable examples cannot establish suitability for a whole practice area. Include difficult, ambiguous and out-of-scope tasks.
Check governance in the actual service
The Swiss Bar Association (SAV) AI guidance of February 2025, published in German, French and Italian, informs assessment of secrecy, supplier arrangements and independent review. Swiss hosting and an AI label do not certify compliance. Examine access, retention, subprocessors, contracts and the intended matter before introducing client data.
Likewise, a precedent library is not automatically available for model training or reuse across clients. Check confidentiality, permissions, source quality and who can access or change the material. Record model and service changes that can affect prior evaluation results.
Make the business case explicit
Evidence for a proposed service
0/6These are planning questions, not a scored readiness certification. A decision to continue the existing process can be reasonable when a proposed change fails its quality or economic tests. A successful pilot supports its tested scope, with monitoring as that scope changes.
Key Takeaway
Get in touch to design an evaluation for your practice.