Criminal disclosure AI: a source-linked chronology workflow

A chronology is useful when a solicitor can check each entry against the material supplied. AI can help prepare an index and propose events, but an apparently fluent summary can hide a wrong date, an omitted page or a conclusion that no source supports.

A practical document workflow should make those problems easier to find. Its output is a review queue with source references, unresolved questions and corrections. It is not a decision about the case.

This guide describes a proposed workflow using wholly synthetic material. It is not an Ampliflow client result or a demonstration of a deployed system.

Establish what arrived

Preserve the files as received. Create a register containing a stable document identifier, filename, version, page count, receipt details and processing state. Keep the original separate from any extracted text.

Optical character recognition, or OCR, converts scanned text into machine-readable text. It can misread handwriting, dates and poor scans. A page that fails extraction must remain visible in the register as an exception.

The register should distinguish three statements: a document was received; its text was processed; a person checked its content. None establishes that the material supplied is complete.

Extract what the source actually says

Ask for a narrow output: the event wording, the date as written, the source document and page, and any ambiguity. Preserve an uncertain date as uncertain. Do not resolve disagreements by asking the model to choose the more plausible account.

The synthetic review pack accompanying this draft includes these examples:

These examples are evaluation inputs. They have no connection to a real person or matter. The source pack and answer key let a reviewer see precisely what the workflow should preserve and where it should stop.

Specify a chronology row a solicitor can check

Use a consistent output record: document ID, version, page, event date as written, account or speaker, proposed summary, uncertainty and review decision. Keep receipt dates separate from event dates. For a long email chain, preserve the message and attachment relationship rather than giving the whole file one undifferentiated reference.

For example, an invented witness note saying “I was told the meeting was on Friday” supports an attributed account, not confirmation that the meeting occurred. The proposed chronology should identify the speaker and the qualification. The reviewer decides its significance.

This guide concerns preparation from supplied material. It does not claim to discover missing prosecution material or determine whether disclosure obligations have been met. The evaluation checklist supplies the separate method for checking extraction quality and access controls.

Put the original beside the draft

For each proposed chronology row, the reviewer should be able to open the correct document version and page. A link that merely opens the folder is too weak for checking a disputed fact.

The review screen should distinguish source text from model interpretation and staff corrections. If a fact cannot be traced, flag it and prevent it being treated as approved. Similar documents can be marked as possible duplicates, but preserve both until a person has checked them.

Solicitors remain accountable for AI-assisted legal work. The SRA’s effective-supervision guidance addresses human scrutiny and professional judgement when AI is involved. A review button alone does not establish that suitable supervision occurred. SRA: effective supervision.

Check access before using a live matter

Criminal-offence information has a separate data-protection condition layer. The ICO explains the Article 10 rules and relevant Data Protection Act 2018 provisions. A general lawful basis does not settle every condition for this category of information. ICO: criminal-offence data.

The firm needs to determine the applicable conditions, approve the provider and establish who can access each matter. The proposed implementation must enforce those decisions in search, document storage, exports and logs. A synthetic trial can demonstrate the review process without sending real criminal-case material to a new supplier.

Measure the finished work

Compare manual preparation with assisted preparation plus review and correction. Count missing entries, invented facts, wrong source links and unresolved exceptions. Include a deliberate conflict and an unreadable page in the evaluation; an easy sample will not show how the workflow handles failure.

For legal-aid work, the Ministry of Justice guidance requires claims to reflect actual, reasonable work following AI use, including appropriate checking. A faster workflow is not a basis for billing the hypothetical time a manual process would have taken. Legal Aid Learning: using AI.

The legal-aid time-recording guide shows how to separate staff work from automated processing in a proposed time record.

The first deliverable to assess is therefore modest and inspectable: a document register, a proposed chronology, an exception list and a reviewer record. Use the legal workflow hub to place it alongside intake and firm operations; use the introductory guide to decide whether it is the right first task.