AI Agents For Legal Workflows

What an AI-Agent System Should Do

For lawyers, the appropriate model is not a single general-purpose AI tool that “runs the litigation”. It is a suite of specialised agents, each assigned a defined document task and operating over a shared, controlled matter record. That record should contain approved client instructions, correspondence, pleadings, evidence, discovered documents, orders, procedural timetables and legal research. Every factual proposition generated by an agent should remain linked to its source document or identified instruction.

1. Matter-Intake and Chronology Agent

The first agent should convert instructions and source documents into a structured matter map. It identifies parties, dates, transactions, agreements, communications, potential witnesses, relevant documents, deadlines, asserted wrongs, potential remedies and information gaps. Its principal outputs are a working chronology, a document register, a draft issues list and a list of questions requiring instructions.

This is the foundation for every later document. A Statement of Claim agent should not have to infer the key events from thousands of emails; it should receive an approved chronology in which each event is linked to a source. Similarly, a discovery agent should not begin with an undefined document universe; it should receive identified custodians, repositories, date ranges and pleaded issues.

2. Originating Application and Statement of Claim Agent

A Statement of Claim agent should work from an approved claim matrix containing: the parties; the relief sought; each proposed cause of action; its legal elements; the material facts said to satisfy those elements; the available supporting documents or instructions; and any known gaps. The agent can produce a first draft of the Originating Application and Statement of Claim in the appropriate format, including proposed relief and separately numbered material allegations.

Its most important output is not merely the draft pleading. It should generate an internal allegation-to-source schedule, identifying whether each paragraph is supported by a contemporaneous document, a client instruction, a witness account, or requires confirmation. It should flag potentially missing particulars, unexplained loss, unclear causation, inconsistent dates, unsupported allegations and limitation issues. The supervising lawyer decides which facts are material, settles the legal characterisation and ensures the pleading is accurate, sustainable and consistent with the client’s instructions.

A related Genuine Steps agent can prepare a draft statement from the pre-action record. It should extract relevant correspondence, offers, meetings, mediation proposals and explanations for why pre-action steps were impracticable or inappropriate. It should not overstate settlement efforts or imply steps that were not actually taken.

3. Defence, Reply and Amended-Pleadings Agent

A Defence agent should read the filed Statement of Claim paragraph by paragraph and create a response matrix. For each allegation, it should identify the client’s proposed position—admission, denial, non-admission, inability to admit, or positive response—together with the instructions and documents supporting that position. It can then draft the Defence in the same sequence as the Statement of Claim, making the response process more reliable and reducing the risk that an allegation is missed.

The agent should also identify where a bare denial may be insufficient because the respondent has a positive case that needs to be pleaded; where a request for particulars is appropriate; where a limitation, estoppel, contractual, statutory or other defence may need consideration; and where instructions are incomplete. A Reply agent should identify genuinely new matters raised in the Defence and draft a focused response, rather than simply restating the applicant’s case.

An amendment agent can compare successive versions of pleadings and produce a change report. That report should identify which allegations have been added, removed or altered, and flag consequential review of discovery categories, affidavit evidence, expert instructions, chronologies, submissions and the court book.

4. Interlocutory-Application Agent

An interlocutory agent should be configured around a specific procedural objective: for example, further and better particulars, an extension of time, a stay, a set-aside application, a discovery order or a case-management direction. The lawyer defines the order sought and the procedural basis. The agent then draws together the procedural history, relevant correspondence, existing orders, supporting evidence and applicable rules or practice-note requirements.

Its outputs can include a draft interlocutory application, proposed orders, a supporting affidavit structure, a procedural chronology and an Outline of Submissions. It should also run consistency checks: whether the affidavit proves the facts needed for the order; whether the outline accurately states the relief; whether the proposed orders are sufficiently precise; and whether the application is consistent with the existing timetable. The lawyer retains responsibility for strategy, procedural judgment and the final articulation of the relief.

5. Discovery, List of Documents and Subpoena Agent

A discovery agent should translate the pleaded issues into a document-management plan. It can propose custodians, systems, date ranges, search terms, document categories and collection priorities. Once material is collected, it can process the data, identify duplicates and email families, extract metadata, group related documents, identify missing attachments and create review queues.

For the List of Documents, the agent can prepare a draft schedule from documents that human reviewers have classified as responsive, privileged, confidential or requiring further review. It can help identify document descriptions, document dates, custodians and production status. However, final relevance, privilege and confidentiality decisions must remain with lawyers or appropriately supervised reviewers.

A subpoena agent should operate on a narrower task. It should identify the proposed recipient, the documents or evidence sought, the connection to a pleaded issue, the required return date, and whether the request is appropriately framed. It can draft a subpoena or an application for non-party discovery, together with an internal note explaining why the material is relevant and why the selected procedure is suitable. For a Notice of Objection to a Subpoena, it can collate the grounds of objection and the documents affected, but the legal team must settle the objection and any privilege claim.

6. Evidence, Affidavit and Witness-Preparation Agent

An evidence agent should convert the pleadings and documentary record into an evidence map. It identifies each factual issue, which party bears the onus, the documents bearing on the issue, the available witnesses and any evidentiary gap. That map then supports preparation of witness outlines, affidavit structures and trial affidavits.

For an affidavit, the agent can produce a source-linked draft organised around what the witness personally saw, said, did or received, and the documents the witness can identify or explain. It should distinguish between personal knowledge, information from a document, and an inference. It should flag statements that appear to be argument, hearsay, opinion or unsupported assertion. The lawyer must conduct the witness conference, test the witness’s account, decide whether evidence is appropriate and ensure the final affidavit is in a form the witness can properly swear or affirm.

7. Trial-Preparation and Submissions Agent

A trial-preparation agent should maintain an issues matrix that links every disputed issue to the relevant pleading paragraphs, evidence, witnesses, documents, legal propositions and relief. From that matrix, it can draft a proposed Statement of Factual and Legal Issues in Dispute, a Statement of Agreed Facts, an agreed or party chronology, a tender list and a witness list.

A submissions agent should create a proposition map: each proposition of law or fact is linked to the relevant evidence, authority and relief sought. It can generate an Outline of Submissions and later a draft of closing submissions, while highlighting propositions that are not pleaded, not supported by evidence, unsupported by authority or unnecessary to the result. This allows lawyers to focus their work on the core advocacy tasks: selecting the best arguments, addressing adverse material, making forensic concessions where appropriate and presenting a coherent case theory.

8. Court-Book and Hearing-Bundle Agent

A court-book agent should be a document-control agent rather than an autonomous advocacy agent. It receives only the final, lawyer-approved set of pleadings, affidavits, exhibits, tender documents, agreed facts, chronologies, submissions and authorities. It then prepares a proposed index, applies consistent naming and ordering, identifies duplicates and superseded documents, creates bookmarks and pagination, and produces a cross-reference report.

The agent should test practical questions that are otherwise time-consuming and prone to error: Is every document cited in submissions included in the bundle? Does every chronology reference lead to the correct page? Is a cited affidavit annexure included? Is the document the final approved version? Is the material actually proposed to be used at hearing? Are there inconsistent dates or titles across the index, chronology and submissions?

The supervising solicitor must approve the final hearing bundle and eBook. That approval is essential because the court book is not merely an administrative compilation: it shapes how the Court can navigate and understand the evidence.

9. Judgment, Costs, Enforcement and Appeal Agent

A post-judgment agent should read the entered orders and reasons, extract operative obligations and dates, and create a compliance and decision register. It can prepare a chronology of post-judgment deadlines, identify matters reserved for later determination, assist with costs schedules and produce first drafts of costs submissions.

For enforcement, it can organise judgment details, debtor information and procedural options into an internal enforcement plan. For an appeal, it can create a decision-and-grounds matrix identifying proposed errors, relevant passages of the reasons, materiality, supporting record references and time limits. It may produce a draft Notice of Appeal, but the decision to appeal and the formulation of grounds remain legal and strategic judgments requiring senior lawyer or counsel review.

How the Agents Connect

The system’s value comes from the hand-off between agents. The approved chronology and issues matrix feed the Statement of Claim. The pleadings define discovery. Discovery and the evidence map feed affidavits. The settled pleadings and approved evidence feed submissions. The final pleadings, evidence and submissions feed the court book.

Each amendment or new document should trigger a controlled review. If the pleading changes, the system should flag discovery categories, witness evidence and submissions that may require revision. If a document is found to be privileged, unreliable or superseded, the system should identify every draft that refers to it. The litigation team thereby works from a connected and auditable record rather than a collection of disconnected drafts.

Core Safeguards

For this model to be useful in practice, it requires lawyer-controlled approval points, role-based access, secure handling of confidential material, privilege controls, version history and source citation. Each agent should state what it relied on, what it could not verify, and what requires legal review. The system should never generate unsupported facts or authorities merely to complete a document.