E-Discovery and Predictive Coding for Law Firms

Document-heavy work occurs in transactions, investigations, employment disputes, estates and arbitration. Technology-assisted review, or TAR, uses reviewer decisions and machine learning to help prioritise documents. Its usefulness depends on the collection, the question and the validation method. It should not be confused with a chatbot generating answers about an unverified set of documents.

Start with the actual production obligation

Determine the applicable procedure, requests, tribunal orders and confidentiality restrictions before choosing a review protocol. Evidence from US discovery research does not establish either an obligation to use TAR or the acceptability of a particular process in Switzerland or Germany.

The Swiss Arbitration Association's 2025 document-production initiative discusses the burden of production and proposes practical adjustments. It is a reform and practice discussion, not evidence that every Swiss tribunal has approved predictive coding. Record the process required or agreed for the particular matter.

What the research supports

In Cormack and Grossman's 2014 evaluation, active-learning protocols required less review effort than the compared passive-learning protocols at a given recall level across the studied tasks. Recall concerns how much of the relevant material is found. The study used a reviewer simulation and specific datasets; it is not a universal saving for today's products or a test of generative legal advice.

Their 2024 validation paper emphasises independent relevance assessments and proposes blind validation strategies. The practical lesson is to avoid assessing a system solely against the same judgments used to train it. A credible quality claim needs a defensible reference and a description of uncertainty.

Choose a bounded task

Potential uses include prioritising responsive material, suggesting document categories and preparing a chronology. Define which of those outputs will be relied on. A correctly copied date is not proof that an event occurred on that date; the document may describe a past event, a proposed deadline or an attachment from another period.

Preserve source documents, page references, versions and relationships. Check OCR quality and missing material. Report recall as a measured estimate with its uncertainty instead of promising that every date or relevant document will be found. Measure the effort needed to validate the complete result.

Check what the tool leaves out

Reviewing only the highest-ranked documents cannot reveal all omissions. An elusion check examines a sample of material treated as non-responsive. Its value depends on the sampling design, independent assessment and interpretation; a small convenient sample cannot prove there are no missed relevant documents.

Distinguish relevance from privilege, professional secrecy and permission to produce. A relevance classifier is not a final decision about those legal questions. Assign suitably qualified reviewers and escalation routes, and test for cross-matter access or confidentiality leaks.

For lawyers subject to the Swiss Lawyers Act (BGFA), Article 13 BGFA makes professional secrecy a statutory duty covering everything clients entrust to them in the course of their professional activities, and they must ensure that their auxiliaries preserve it. The SAV AI guidance supports examination of data handling and critical independent verification, and the CCBE guide on generative AI of 2 October 2025, hosted on the same page, asks lawyers not to enter client data without appropriate safeguards. Swiss or EU hosting alone does not establish compliance. Assess recipients, subcontractors, access, retention, reuse and the applicable legal permissions.

Compare the complete workflow

Count collection, preparation, review, correction, legal assessment and production alongside model processing. Compare equivalent tasks and quality standards. A tool can be fast while creating expensive verification work.

Do not assume that adding a human always produces the best result. A 2024 meta-analysis of human and AI experiments found varied outcomes and average performance below the better standalone participant across the included studies. That broad finding is not a TAR-specific result; it reinforces the need to evaluate the combined workflow directly.

Assessing a document-review pilot? Get in touch to discuss scope, validation and confidentiality.