
A useful legal assistant can make a document collection easier to inspect. It can propose a chronology, identify missing materials, compare versions, or prepare a draft around evidence the lawyer has already selected. The work becomes more useful when every claim leads back to its source and unresolved questions remain visible.
The first design decision is the work product. An intake summary, a research memorandum, and a filing each require different sources, review, and permissions. This guide develops a fictional commercial-matter chronology for a United States practice. It is a proposed implementation pattern, not a report of a law firm’s deployment or advice about a particular case.
On this page
- Separate the jobs before connecting a matter
- First build: a chronology with evidence links
- Build the first version
- Walk through a fictional disagreement
- Evaluate more than whether the draft reads well
- What changes with practice size
- Professional obligations need jurisdiction-specific review
- Keep the operating model visible
Separate the jobs before connecting a matter
| Job | Useful assistance | Decision retained by the practice |
|---|---|---|
| Intake | Organize supplied facts and identify missing fields | Conflicts, engagement, scope, and suitability |
| Document organization | Classify files and suggest event entries | Relevance, completeness, and evidentiary significance |
| Research | Find candidate authorities and compare passages | Applicable law, authority, current validity, and argument |
| Drafting | Prepare an editable text grounded in approved material | Legal strategy and the final work product |
| Deadline handling | Extract a date exactly as stated | Applicable rule, calculation, exceptions, and docket entry |
Keep these stages independently reviewable. A factual chronology does not establish that a statement is true, that a document is admissible, or that a claim will succeed.
First build: a chronology with evidence links
Use a demonstration matter containing a purchase order, three fictional emails, an invoice, and a delivery note. The assistant’s job is to propose dated events, identify conflicting accounts, and produce questions for a reviewer. The output remains a working draft.
A prototype can use local fictional files and a table. A connected version needs matter-scoped document access, reliable extraction, stable document identifiers, a review interface, and a way to retain the approved result with its source references. Use the practice’s supported document-management interfaces. A connector that can search every matter is broader than this task requires.
The architecture is: authorized matter selection → permitted source retrieval → event extraction → source validation → reviewer correction → approved chronology version. Keep source documents immutable. Record which document versions produced each draft.
Build the first version
- Create the matter boundary. The application selects
MATTER-DEMO-017using an authorized user’s access. Apply the same boundary to search results, extracted text, generated files, and logs. A prompt that says “this matter only” cannot replace an access check. - Create source records. Give each document an identifier, file hash or version, title, and page references. For email, retain sender, recipients, sent time, and attachments. Keep received date and event date distinct.
- Specify the event schema. Include
event_date_as_written,normalized_date,date_basis,description,source_id,page_or_message,short_supporting_passage, andreview_status. An uncertain date remains uncertain. - Extract candidate events. Ask for one event per entry and a short supporting passage. Preserve allegations as attributed statements. “The supplier wrote that delivery occurred” is a different event from “Delivery occurred.”
- Validate mechanically. Check document identifiers, page ranges, duplicate entries, and required fields. Confirm that quoted support appears in the extracted source, allowing only documented OCR normalization. Flag conflicting dates instead of choosing a convenient version.
- Review the original. A lawyer or appropriately supervised reviewer inspects the source image or message. Corrections should remain traceable. A visually plausible table can still be incomplete.
- Save an approved version. Record reviewer, time, source versions, and unresolved questions. Export an accessible table or document. Create a new version after a source changes rather than silently rewriting a reviewed chronology.
Use a deliberately constrained extraction instruction:
Create candidate chronology entries from the supplied demo matter.
Use only the provided documents. Cite source_id and page/message.
Separate an asserted event from the fact that somebody asserted it.
Preserve ambiguous dates and contradictory accounts.
Do not calculate legal deadlines or decide legal significance.
Return unresolved questions for the reviewer.
Instructions define the task; the application must enforce access, allowed operations, and output validation.
Walk through a fictional disagreement
Email E03, sent by the supplier on May 7, says an order was delivered on May 6. Delivery note D02 is dated May 8. The assistant should not merge those into a single “delivery on May 6” event.
| Candidate entry | Source | Review result |
|---|---|---|
| Supplier stated on May 7 that delivery occurred May 6 | E03, message 1 | Verify statement and retain attribution |
| Delivery note bears May 8 date | D02, page 1 | Verify document and identify what its date means |
| Actual delivery date remains unresolved | E03 and D02 | Ask for receiving record or other evidence |
If an invoice displays 04/05/26, preserve that text and ask which date convention applies. Do not let locale settings decide the matter’s chronology without evidence.
For a research assistant, add a separate authority-verification stage. The reviewer opens the actual authority, confirms the quoted language, checks jurisdiction and weight, and uses the practice’s approved process to determine current validity. A link that opens is only the beginning of that review. A general model’s recollection of a case is not a source document.
Evaluate more than whether the draft reads well
Build a small evaluation set that includes duplicates, missing attachments, a forwarded message with multiple dates, an illegible page, two matters containing similar names, and a source containing instructions to reveal another matter. Include a document deliberately outside the user’s permissions.
Measure whether proposed entries preserve attribution, identify uncertainty, point to the correct original, and omit unsupported claims. Separately measure missed events: excellent precision on a small subset can conceal a poor chronology. Record reviewer correction time and compare the final reviewed result with the existing process.
Proposed release gates include no cross-matter disclosure in the evaluation set, no invented source references, and successful handling of unavailable documents. These are project criteria to validate, not a claim that a single test establishes ongoing confidentiality or correctness.
Maintain a test set after launch. A change to OCR, retrieval, model, prompt, or access mapping can change the result. Pause the workflow if permissions cannot be verified or source provenance is lost.
What changes with practice size
A solo practice can start with fictional files and a manual review table. The practical questions are whether the approved software terms fit the work, how the lawyer will verify every result, and whether the tool reduces total work after corrections. Keep subscription and administration demands proportionate.
A medium firm needs matter-access mapping, supervisory ownership, standardized review records, and a consistent way to handle new or departed staff. Different practice groups may need different schemas and source collections. Build those differences deliberately rather than placing every file into one searchable index.
A large firm or corporate legal department adds ethical walls, business-unit boundaries, outside counsel, retention and hold requirements, and more complex document systems. Treat exports and generated summaries as controlled matter information too. An approved model endpoint does not by itself resolve the lifecycle of the files sent to it.
In each setting, the person approving use needs to understand where information goes, who can access it, and what the service contract permits.
Professional obligations need jurisdiction-specific review
North Carolina’s 2024 Formal Ethics Opinion 1 addresses competence, protection of client information, supervision, and the lawyer’s independent responsibility when using AI. Its discussion reinforces reviewing and evaluating another person’s or a system’s work before relying on it. The appropriate application depends on the representation and the tool. North Carolina State Bar opinion
For work elsewhere, check the relevant jurisdiction’s professional rules and court requirements. Do not assume that an AI disclosure or citation rule is identical in every court. Have the responsible lawyer decide what client communication, consent, supervision, billing treatment, and records are required for the specific use.
Confidentiality and privilege are related but distinct questions. Do not label every generated file privileged simply because a lawyer requested it. Design permissions, storage, and sharing around the practice’s actual requirements and the material’s purpose.
Keep the operating model visible
Assign ownership for the source collection, access mapping, reviewer training, and technology. Budget for OCR, search or document processing, model use, review, support, and exports. The measure that matters is a defensible reviewed work product and its total cost, not the number of pages generated.
Begin with workflow design, use agent anatomy to understand tool access, and compare small-business and enterprise implementation patterns.