Investigations move fast. Regulatory deadlines can be measured in days rather than weeks. A dawn raid may leave you with seized devices and a requirement to respond to the SFO or FCA before you have had time to understand what you are dealing with. Internal investigations triggered by a whistleblower complaint or a board-level concern carry their own urgency. Whatever the context, you may find yourself facing tens or hundreds of thousands of documents with very little time and, sometimes, very little to go on.

Simply reviewing documents is not an option at that scale. A dataset of 100,000 documents would take a team of ten reviewers several weeks to work through at standard review rates. You probably have days.

Fortunately, the tools available on modern eDisclosure platforms are well suited to exactly this problem.

Start with what you know

If you have some knowledge of the matter — individuals involved, relevant date ranges, key terms or phrases — metadata filtering is a good first step. Filtering by custodian, date, file type, or search term can reduce the dataset quickly and exclude material that is unlikely to be relevant. The risk is over-exclusion: documents of interest can sit outside the obvious parameters, particularly where individuals have been careful about what they put in writing. Metadata filtering narrows the field; it rarely closes the case.

Ask the dataset questions

Most modern platforms allow users to run LLM-style queries across the entire uploaded dataset. In Everlaw, this functionality is called Deep Dive. For a one-off charge on upload, users can ask unlimited generative AI questions against the full document set. Answers are supported by referenced documents and a wider supporting document set, giving investigators an initial target for review.

Crucially, questions can be framed without prior knowledge of the specific issue. Searches built around concepts — concealment, acknowledgement of wrongdoing, awareness of risk, instructions to destroy — can surface relevant material even where you do not yet know what you are looking for.

Batch coding and extraction

Generative AI coding allows users to describe, in plain language, the kinds of documents or information they want to find, and then run that instruction as a batch action across thousands of documents in minutes. Rather than reading each document to decide whether it is relevant, the platform makes an initial classification at scale, leaving human reviewers to focus on the documents the platform has identified as most likely to matter.

Document clustering

Once an initial document set has been identified, clustering tools organise documents by conceptual similarity without any human input or prior assumptions. Groups of related documents are identified by the terms most prevalent within them and displayed visually. Clustering is particularly useful for understanding the shape of a dataset and identifying concentrations of potentially relevant material that a keyword search might not have surfaced.

Predictive coding

Where time allows for a slightly more structured approach, predictive coding can significantly accelerate review. Once a training set of around 200 human-reviewed documents has been built — with examples coded responsive and non-responsive — the platform generates a statistical model and uses it to rank the remaining documents by predicted relevance. Reviewers work through the highest-scoring documents first, and the model continues to refine its predictions as more documents are coded.

Putting it together

In practice, investigations rarely follow a linear path. The tools above are most powerful when used in combination: metadata filtering to reduce volume, Deep Dive to identify themes and targets, batch coding to classify at scale, clustering to find related material, and predictive coding to prioritise human review. The goal at the outset is not to review everything. It is to understand the dataset well enough to focus your effort where it matters most.

If you are at the start of an investigation and need to get on top of a large dataset quickly, we can help you design the right approach.