The groundwork between intake and review, done by agents.
Six specialised agents interview, structure, check and chase, so your team starts every matter from a sourced draft. Each one proposes; a person on your team decides.
What the record surfaces
Illustrative
- Timeline
- Evidence
- Conflicts
- Tasks
MatterGraph’s AI agents are specialised assistants that each handle one part of turning a client’s account into a case record: interviewing, building the timeline, tracking evidence, comparing sources, drafting follow-up questions and summarizing. Each agent proposes a draft with its source attached. A person on the firm’s team approves, edits or rejects it before the record changes or anything reaches the client.
Software that works toward a goal, one step at a time.
OpenAI describes agents as “systems that independently accomplish tasks on your behalf.” A chatbot answers the question in front of it. An agent works through a sequence: it reads what it has, decides what is missing, takes a step with a tool, looks at the result and decides again.
Researchers call this interleaving of reasoning and action the ReAct pattern. It is what lets an agent notice that a date is vague and ask about it, instead of filing the vague date and moving on.
Every agent is built from the same three parts. MatterGraph keeps the tools narrow and the instructions strict, because in legal work the safest agent is the one with the fewest ways to act on its own.
- A model that reads and reasons about language
- Tools it can use, limited here to proposing records inside the matter
- Instructions and guardrails that define what it may never do
One agent step
- 01
Read
The new answer, document or reply
- 02
Decide
What is missing, what conflicts, what comes next
- 03
Act
Propose an event, a flag, a task or a question
- 04
Check
Attach the source and the review status
- 05
Hand off
Queue the draft for a person on your team
The loop repeats as new material arrives. It never ends in an action outside the firm without a person.
Intake is the kind of work rules engines struggle with.
OpenAI’s guide names three signs a workflow needs an agent rather than fixed rules. Client intake shows all three.
Stories, not fields
Clients describe events out of order, in their own words, across calls, messages and scanned documents. Turning that into dated events means reading language, not matching patterns.
Every matter is different
Whether a missing date matters, or which record should exist, depends on the rest of the story. Fixed checklists either miss the exception or ask everything of everyone.
Templates that never finish
An intake form that tries to anticipate every follow-up grows until nobody maintains it. An agent follows the firm’s template and asks only the next question that matters.
One job each, and a clear handoff.
Narrow agents are easier to test, easier to trust and easier to switch off than one agent that does everything.
- Step 01
Interview agent: gets the story out
Runs the guided interview after the client signs. It asks one question at a time, chooses the next one from what is still missing, accepts approximate dates, and saves progress so the client can stop and return. It says it is an AI assistant working for the firm, and it routes legal questions to the attorney instead of answering them.
Guided intakeGuided questionsApproximate datesSave and resume - Step 02
Timeline agent: builds the chronology
Turns answers and documents into proposed events, each with a date type (exact, approximate, range or unknown), the people and place involved, and the exact answer it came from. Every event starts as client stated. Only a person can mark it document supported or attorney confirmed.
Case timelineEvent proposalsDate typesSource links - Step 03
Evidence agent: works out what should exist
Maps each important event to the records that would normally support it, tracks what has arrived and what has not, records why a missing item cannot be obtained, and opens a task for each gap. It never contacts a third party itself.
Evidence trackingSupport mappingGap trackingLinked tasks - Step 04
Consistency agent: finds what does not line up
Compares sources inside the matter: two dates for one event, a person present in one answer and absent in another, a document that disagrees with the account. Both sources are quoted side by side in neutral language, for a person to resolve.
Contradiction detectionSource comparisonSide by sideNeutral wording - Step 05
Follow-up agent: asks the right question once
Turns each gap or conflict into one targeted question instead of the whole intake form again, links it to the event it resolves, and writes the reply back with its source. Questions about conflicts or sensitive events wait for paralegal approval.
Task boardTargeted questionsParalegal approvalReply linking - Step 06
Summary agent: briefs the attorney
Drafts the matter summary and the open-questions list from reviewed events, with a source on every line. Anything it cannot source becomes an open question, not a sentence. Wording, relevance and final use stay with the attorney.
Cited summaryOpen questionsAttorney review
Review gates, set by risk.
Agent guidance from OpenAI and Anthropic converges on one rule: sensitive or irreversible steps wait for a person. These are the gates the product is designed around.
Next interview question
Automatic within templateChosen by the interview agent, inside the firm’s own template.
Proposed event or evidence link
Paralegal reviewShown as a suggestion on the record until a paralegal approves or edits it.
Possible source conflict
Internal onlyVisible to the legal team only. A person decides what it means and whether to ask about it.
Question to the client
Approval requiredDrafted by the follow-up agent and sent only after a paralegal approves it.
Matter summary or outline
Attorney reviewA draft for the attorney. It cannot be exported until an attorney marks it reviewed.
Anything leaving the firm
Never automatedRecord requests, filings and messages to third parties are always a named person’s action.
Designed around what is known about AI, and about memory.
Three findings that shaped how the agents are allowed to behave.
19% to 79%
Legal professionals using AI, 2023 to 2024
Adoption is already here. The open question for firms is which work AI should touch, and under what review.
Clio Legal Trends Report 2024, via LawSites17% to 33%
Hallucination rate measured in leading legal research AI tools
Even purpose-built legal tools invent material. MatterGraph agents do no legal research, and every proposed fact must point to an answer or document in the matter.
Magesh et al., Stanford RegLab, 2024Common
Discrepancies when people retell stressful events
Research with refugees found that repeated accounts often differ in peripheral details. Agents raise differences for review and never treat them as findings about credibility.
Herlihy, Scragg and Turner, BMJ, 2002About the agents.
- Do the agents give legal advice?
- No. MatterGraph organizes facts. The client-facing assistant identifies itself as AI, does not answer legal questions, and passes them to the attorney. Legal analysis, strategy and final wording stay with the firm.
- Can an agent contact a client or a third party on its own?
- The intended design is no. Questions to clients wait for paralegal approval, and requests to third parties are always sent by a person. Agents can only propose records inside the matter.
- What happens when an agent is wrong?
- Its output is a suggestion with the source attached, so a reviewer can check it against the original answer or document, then edit or reject it. Rejected suggestions do not change the record.
- Are the agents running today?
- The agents on this site are a preview of the intended workflow, shown on a synthetic matter. A pilot can be evaluated on synthetic or fully redacted matters before any decision about live client data.
- Can our firm turn an agent off?
- That is the intent of giving each agent one job and one handoff. A firm should be able to keep the manual workflow for any step it prefers to run by hand, and a pilot is the place to decide which steps those are.
Test the workflow on a representative matter.
Start with a synthetic or fully redacted matter. An initial evaluation does not require live client data.