AI legal research for Swiss law is useful when it makes the path from question to verifiable authority shorter without hiding the limits of the search. A reliable workflow defines the issue and relevant source universe first, then checks every important proposition against the original, current and applicable source before it enters legal work.
Start with the legal question, not the search box
A vague prompt produces a vague research path. Before using any tool, write down the decision the research must support. Separate facts that are fixed from assumptions that still need evidence. Identify the jurisdiction, relevant date, procedural posture and the type of authority that would answer the question.
This matters particularly for Swiss law. A federal statute, a Federal Supreme Court decision, cantonal case law and administrative guidance do not have the same role. A result can be textually close to the question and still be legally irrelevant because it concerns another canton, an earlier version of a rule or a different procedural setting.
A useful research instruction therefore contains four parts:
the precise legal issue;
the facts that may change the result;
the jurisdiction and time period;
the required output, including sources and uncertainty.
The instruction is not the legal analysis. It is a testable specification for the search.
Define the source universe before evaluating an answer
No research result can be more complete than the sources searched. Ask which collections the tool actually uses for the specific query. Marketing labels such as “Swiss law” or “legal database” are not enough. The relevant questions are whether the required level of court, canton, language and date range are covered and whether the full text or only summaries are searchable.
Do not infer coverage from one successful result. A tool may retrieve a well-known federal decision while missing less visible cantonal material. Conversely, a broad index does not guarantee that the ranking surfaces the authority that matters. Coverage, retrieval and answer generation are three different stages.
For an important matter, record the expected source classes before searching. The checklist may include legislation, case law, official materials and any licensed secondary sources available to the team. If a required class is unavailable, the research plan should contain a second search route.
Search in layers instead of asking one large question
One long prompt can compress several legal issues into an answer that looks complete but is difficult to audit. A layered workflow is easier to verify:
begin with terminology, applicable rules and likely leading authorities;
test the decisive factual variations separately;
search for contrary authority and exceptions;
check whether later decisions cite, distinguish or replace the initial result;
only then request a synthesis.
This sequence makes omissions visible. It also helps the responsible lawyer distinguish a failure to retrieve a source from a failure to interpret it correctly. The CASUS legal research workflow can support this process, but the method remains useful regardless of the chosen platform.
Verify the original source, not only the generated explanation
A citation is the beginning of verification, not its conclusion. Open the original decision or statute and check that the source exists, that the quoted proposition appears in it and that the surrounding reasoning supports the proposed use.
For every source that carries the conclusion, verify at least:
Check | Question |
|---|---|
Identity | Is this the correct decision, provision or official document? |
Context | Does the relevant passage concern the same issue and procedural setting? |
Authority | What weight does the source have for this question? |
Currency | Is the rule still in force and has later authority changed the position? |
Completeness | Are important exceptions, contrary decisions or factual distinctions missing? |
If the source cannot be opened or traced, the proposition is not ready for a memo, pleading or client advice. A plausible citation without a verifiable source is a warning signal.
Keep the research trail auditable
Good legal research should be reproducible by another member of the team. Save the final question, search date, source scope, decisive authorities and unresolved gaps. Record why a source was included and why apparently relevant material was rejected.
The research note does not need to reproduce every query. It should preserve the decisions that matter: which legal issue was tested, what evidence supports the answer, where uncertainty remains and what still requires human judgement. This trail is especially valuable when a matter is revisited months later or reviewed by a colleague.
Separate research quality from writing quality
A clear answer can be weakly sourced, while a dense list of citations can fail to answer the question. Evaluate the result on separate dimensions:
source support for each material proposition;
completeness of the relevant authority;
correct treatment of jurisdiction and time;
visibility of uncertainty and counterarguments;
usefulness of the final structure for the intended work product.
Only after the research passes these checks should the result move into drafting. The transition can still be efficient: a verified source table can become the backbone of an internal assessment, clause rationale or first draft. But the prose must not outrun the evidence.
Test a research tool with a fixed evaluation set
A live demo rewards the easiest available question. A proper pilot uses a small set of completed matters whose expected authorities and difficult distinctions are already known. Include routine and edge cases, more than one language where relevant, and at least one question where the correct answer is that the evidence remains insufficient.
Score the workflow, not only the answer. Useful criteria include time to the first relevant authority, share of decisive sources found, unsupported propositions, time needed for verification and the clarity of the audit trail. The separate evaluation framework for Swiss legal research tools explains how to compare products without relying on a generic feature table.
Common failure modes
Several patterns should stop a research result from moving downstream:
the answer gives no link or stable identifier for a decisive source;
a source supports a general statement but not the specific conclusion;
the result silently mixes jurisdictions or versions of a rule;
only confirming authority was searched;
an absence of results is presented as proof that no authority exists;
the summary is copied into a document before the original sources are checked.
These are process failures, not merely model failures. Clear review rules reduce them even when the underlying technology changes.
A practical handoff into legal work
At the end of the research, create a compact source table with the legal proposition, supporting authority, relevant passage, status of verification and open question. The responsible lawyer can then decide which propositions are safe to use and which require more work.
For teams evaluating broader Legal AI adoption, the same principle applies: usage is not impact, and speed is not quality. The guide to measuring AI time savings provides a separate method for testing effort and rework. Teams can also test CASUS with their own research questions and compare the output with their established source set.
Frequently asked questions
What makes AI legal research reliable?
Reliability comes from a defined question, a known source universe, layered searching, original-source verification and a documented review trail. The generated answer alone is not sufficient evidence.
Should every cited source be opened?
Every source carrying a material conclusion should be checked in the original. Background references may receive lighter review, but they must not substitute for the decisive authority.
How do I test source coverage?
Use completed matters with known authorities. Check whether the tool retrieves the correct source classes, languages, jurisdictions and dates, including contrary or less prominent material.
Is a missing result proof that no case exists?
No. It may reflect the query, ranking, language, collection or access level. Treat absence as a reason to use another search route, not as proof of non-existence.
When is a research result ready for drafting?
When the material propositions are supported by current, applicable and verified sources, important counterarguments are visible and remaining uncertainty is stated clearly.
