Article

AI and investigations: from a helping hand to a source of evidence

AI and investigations: from a helping hand to a source of evidence

Part of our report

Cross-border white-collar crime and investigations review 2026
Published Date
Jan 27 2026
Related people
Image of Thomas Declerck
Thomas DeclerckPartner, Brussels
Image of Christina Zachariasen
Christina ZachariasenExecutive Director, London
Image of David Schmid
Dr. David SchmidPartner, Frankfurt
Daniel David Smith
Daniel SmithGlobal eDiscovery Consultant, London
Image of Amy Edwards
Amy EdwardsSenior Knowledge Lawyer, London

Artificial intelligence is already embedded in how many organizations operate and investigate. It speeds up document review, detects anomalies, drafts notes and interview questionnaires, and surfaces patterns. It now also creates new categories of evidence. For in-house teams and investigations leads, the question is no longer whether AI matters, but how to run reliable investigations where AI is both a tool and a source of evidence. 

The modern investigation dataset

Modern investigations involve far more data than before. In-house teams need to find relevant facts quickly, defensibly, and at a reasonable cost. The data is no longer largely limited to email and shared drives. It includes chat, collaboration tools, and mobile content. The variety and scale make both collection and review harder, and auditability key.

Hybrid work has made Microsoft Teams central to daily communication. Informal exchanges that once happened in hallways now appear as chats, channel posts, reactions, and meeting records. Employees frequently use consumer messaging apps such as WhatsApp for work-related conversations, which can contain candid, contemporaneous messages and presents its own auditability challenges because of user features such as disappearing messages.

With Microsoft 365, content often spans Exchange, SharePoint, OneDrive, OneNote and Teams. Modern “attachments” are cloud links, and features like co-authoring, version histories, and permissions introduce new preservation and evidential challenges. As a result, the starting point is no longer a tidy email set—it’s a broader digital workspace that offers richer evidence but requires the right methods and tools to manage effectively.

Working with AI-generated evidence

As AI copilots and workplace tools become integrated into daily work in the broader business, those conducting investigations must consider the evidence they produce. Relevant material may include user prompts, generated outputs, and interaction metadata associated with AI systems, whether general-purpose models such as ChatGPT or enterprise tools like Harvey. Audit logs, access controls, and versioning in AI platforms can become central to establishing timelines, intent, and knowledge.

“Some review platforms are already adding features to catch human-to-AI conversations, making these exchanges reviewable alongside other chat data.”

Preservation and discovery requests may soon reach prompt strings, system instructions, retrieved context, and answer citations, as well as audit logs showing who asked what, when, and with which data. Consistent naming, retention, and export practices will make it easier to collect and review these materials without disrupting business operations.

Using GenAI to accelerate review

Collecting data broadly, limited by dates rather than narrow keywords, can be essential to ensure that important evidence is not missed. This increases downstream processing and review, but fortunately modern AI-enabled review can reduce that burden.

Modern eDiscovery platforms are now able to leverage GenAI to evaluate documents based on matter-specific guidance. This allows for the review of large volumes of material at speeds far beyond human capacity while providing concise rationales for each relevance determination. It reduces the traditional link between cost and raw document volume and minimizes reliance on keyword searches. 

The speed at which GenAI can complete its review can prove invaluable in determining early case strategy. In a recent investigation, GenAI’s review surfaced roughly one thousand relevant documents within hours, including firsthand narrative accounts by the subject that keyword-only methods likely would have missed. By surfacing insights early, it enabled a more informed and effective case strategy from the outset.

 

This investigation challenged the traditional assumption that AI review can never match human review. The technology identified a document hundreds of pages long that appeared irrelevant at first glance but contained a few sentences critical to the case, buried deep within.

Pairing technology with well-designed and tested workflows will achieve the best results. Teams are encouraged to allocate resources to prompt development at an early stage, ensuring that legal considerations and relevant hypotheses are appropriately addressed. GenAI accuracy improves with targeted guidance and iterative refinement. Continuous learning workflows (e.g., CAL) can complement GenAI by surfacing subtle, high-heat materials that sit at the margins of relevance scoring. Where in-scope documents are multilingual, language-agnostic models avoid premature translation and allow language services to be reserved for the most relevant materials. 

AI is not a silver bullet. In Retrieval Augmented Generation workflows, answer quality depends on retrieving the most relevant passages, so deficient guidance, poorly constructed queries, or incomplete collections can degrade results, making rigorous validation against cited sources essential. 

GenAI’s speed can also pose challenges when new theories emerge during an investigation which require guidance to be updated. 

These risks highlight the importance of pairing advanced tools with experienced eDiscovery professionals who can design effective queries, calibrate guidance, monitor quality, and maintain defensible, adaptable workflows. Teams with deep, multi-matter experience are best positioned to ensure these technologies are applied correctly and to maximum effect.

Questions to ask

While GenAI offers significant advantages, its effective use depends on understanding both its limitations and associated risks. To ensure informed decision-making, in-house counsel should consider asking technology partners and GenAI providers questions such as:

  • Integration and security: How does the GenAI solution integrate with the review platform and security framework, and what measures are in place to prevent data leakage? Maintaining privilege and confidentiality is key: check how enterprise or third-party tools handle data. Contracts should prohibit training on client data, require data localization if needed, and set rules for deletion and audits. For cross-border work, consult privacy and employment counsel early to address local law requirements.
  • Transparency and explainability: What information supports each prediction (e.g., rationales and cited source passages) so reviewers can validate results efficiently?
  • Grounding and hallucination controls: How are outputs restricted to the collected corpus, and what safeguards exist to mitigate hallucinations?
  • Language coverage and testing: Which languages has the system been tested on, with what datasets, and how does performance vary across languages, scripts, and mixed-language content?
  • Calibration and quality management: What prompt refinement or iteration processes are necessary for accuracy?
  • Validation and defensibility: What built-in validation methods are available—such as statistically valid sampling, recall/precision measurement, confidence intervals, and error analysis—and how do audit trails document decisions to support defensibility before regulators and courts?
  • Workflow adaptability: How are workflows adapted when new investigative priorities emerge midstream?
  • Natural language querying with citations: Can natural language queries return answers from review sets, directly citing source documents for easy validation? 

Expectations of authorities regarding use of AI in investigations

Regulators’ and courts’ expectations for the use of AI in investigations and compliance are developing. There is a clear emphasis on risk-based controls, model transparency, documentation of training data sources, and post-deployment monitoring. Investigative teams should be prepared to explain why a particular AI tool was selected to assist with an investigation, how it was configured, what human oversight was applied, and which controls were in place to protect integrity, fairness, and defensibility.

Enforcement bodies are also adopting AI. This can speed up interactions, but it also introduces new issues around disclosure, validation, and potential error. In the UK, questions have been raised about the impact of shortcomings in AI assisted review on disclosure processes. Lawyers have been asking for clarity about how these issues may have affected current and past cases in terms of disclosure.

Where businesses encounter agency‑generated analytics, prioritization scores, and AI‑derived leads, in‑house teams should consider how and when to request underlying methodologies, validation summaries, and disclosure about human oversight. 

Conclusion

With investigative datasets growing exponentially, success depends on understanding the AI toolset and knowing when to apply each capability. This means recognizing the strengths and limitations of technologies, selecting the right combination to meet case objectives and achieving outcomes that are explainable before regulators and courts. 

Treat AI platforms as potential evidence sources in their own right, and update preservation and collection practices accordingly. 

About A&O Shearman’s eDiscovery capability

The A&O Shearman eDiscovery team is one of the most highly certified and experienced in the industry, with 11 Relativity Masters - the highest level of accreditation available. Our expertise is grounded in more than a decade of hands‑on use of Relativity, during which we have worked closely with the Relativity team to refine workflows, provide structured product feedback, and support the platform’s ongoing development. This longstanding relationship gives us early access to the latest AI capabilities, including aiR for Review, aiR for Case Strategy, and aiR Assist, enabling us to introduce new features to client investigations as soon as they are released. By pairing these innovations with rigorous quality controls and the judgment of seasoned specialists, we help clients surface key facts quickly, manage exponential data growth effectively, and navigate investigative challenges with confidence and strategic focus.

This article is part of the A&O Shearman cross-border white-collar crime and investigations review 2026.

Related capabilities