AI Agent for Legal Decision Support
In-house legal work arrives as a question attached to a pile of documents. This agent does the reading — analysing what the documents say, identifying the clauses that bear on the question, comparing them with your policies, and citing every point — then hands a lawyer a prepared matter.
What is an AI legal assistant?
An AI legal assistant is a governed software worker that prepares legal matters. It reads all documents on a matter together, identifies the provisions bearing on a question by their effect rather than their label, compares them with internal policy, surfaces the organisation’s prior advice, cites every point to its source, and routes specific tasks to specialist legal agents.
What it does
What it is not
The question is easy, the reading is four hours
A large share of in-house legal work is not difficult, it is long. Establishing what the documents actually say, which clauses bear on the question and whether the organisation has taken a position before takes hours, and the judgement that follows takes twenty minutes. The reading is the constraint, and it falls on the most expensive person available.
Matters arrive as a document pile
A question comes with fourteen attachments and no indication which three of them actually matter.
Clauses are found by memory
Whether an obligation exists somewhere across the agreement set depends on someone remembering it does.
Prior positions are lost
Legal already answered a comparable question eighteen months ago, and nothing surfaces that work into the new matter.
Tools cover one document type
A contract tool handles contracts, and a matter that spans a policy, a contract and correspondence needs all three read together.
The matter read, the sources cited
Matter view
Read The Whole Pile, Not One File
Documents read against each other.
Every document on a matter is read together — contracts, policies, correspondence, prior advice — so an obligation created in one and qualified in another is identified as the combination rather than as two separate readings nobody joined up.
- All matter documents read as one set
- Interactions between documents identified
- Irrelevant documents ruled out explicitly
- Prior advice on the point surfaced
All documents together
Clauses
Find Every Provision That Bears On It
Including the ones nobody remembered.
Provisions relevant to the question are located across the whole document set by what they do rather than by how they are labelled, so an indemnity buried inside a schedule and a liability cap in an order form are both surfaced.
Not by heading
Routing
It Calls The Specialist
Playbook, research, policy or NDA.
When a matter becomes a specific job the agent hands it on — playbook grading to contract review, a jurisdiction question to legal research, an internal conflict to policy review, standard confidentiality paper to NDA triage — and returns the result here.
Returned in one place
How the AI Legal Assistant runs a task
- STEP 01
Take in the matter
Every document attached to the matter is parsed with its type identified, and those with no bearing on the question are ruled out explicitly so that the exclusion is a recorded judgement rather than an oversight.
Document intakeRelevance triage - STEP 02
Read them against each other
Provisions are assessed in combination rather than document by document, because the operative position on a question is routinely created in one instrument and qualified by a schedule, an amendment or a side letter in another.
Cross-document readingDefinition resolution - STEP 03
Search by effect
Relevant provisions are located by what they do — limit liability, create an obligation, impose a condition — rather than by their heading, since the clause that matters is frequently the one titled something else entirely.
Effect-based searchProvision extraction - STEP 04
Check the internal position
What the documents say is compared with your own policies and with any prior advice the team has given on the same point, so a new matter starts from the organisation’s existing position rather than from nothing.
Policy comparisonPrecedent retrieval - STEP 05
Route or hand over
Where the matter resolves into a specific job it is passed to the specialist agent that owns it, and otherwise the cited preparation goes to a lawyer who reaches whatever conclusion the matter requires.
Specialist routingCounsel handover
Systems the AI Legal Assistant connects to
Document intake
Analysis
Inputs, outputs and runtime
- Ingests
- Matter documentsThe legal questionInternal policiesPrior advice archiveGoverning jurisdiction
- Produces
- Cited matter summaryRelevant provisions with passagesPolicy conflict listPrior advice referencesSpecialist agent output
- Triggered by
- New matter openedLegal question raisedDocument bundle received
- Human oversight
- A qualified lawyer reaches every conclusion
- Models
- Open-weight LLMs you host — Llama, Qwen or Mistral class
- Typical latency
- Under an hour for a document bundle
- Deployment
- On-premise or sovereign cloud with egress control
- Data residency
- Privileged material never leaves your network
Where the Legal Assistant pays back
Matter Preparation
Read every document on a matter and produce a cited summary of what they say and how they interact.
Clause Discovery
Find every provision across an agreement set that bears on a question, by effect rather than by heading.
Policy Comparison
Compare an external document against your internal policy and report where the two are incompatible.
Prior Position Retrieval
Surface the advice the team gave on a comparable question and whether the basis for it still holds.
Document Summarisation
Reduce a long agreement or bundle to its operative effect with each point cited to its passage.
Triage Of Incoming Work
Assess what an incoming request actually needs and route it to the specialist agent that handles it.
AI Legal Assistant vs chatbots and SaaS copilots
Legal tooling has specialised by document type, which works until a matter consists of a contract, an internal policy and a chain of correspondence that only make sense read together.
| Generic chatbot | SaaS copilot | VDF AI | |
|---|---|---|---|
| Unit of work | One document | One document type | The whole matter |
| Finding provisions | By heading | By keyword | By what the clause does |
| Cross-document effects | Missed | Missed | Read in combination |
| Prior advice | Unavailable | Not indexed | Searched before external |
| Citations | Rare | Document level | Passage level, verified |
| Gives advice | Freely | Freely | Never — counsel advises |
| Where the matter sits | Vendor service | Vendor tenancy | Inside your own perimeter |
Governance and controls
Privilege attaches to legal work and survives only if the material stays within the circle that is entitled to it, which makes matter-level access control a legal question rather than an administrative one.
Privileged material contained
Matter documents never leave the network
Matter-level access
Visibility limited to the matter team
Output marked preparation
Nothing is presented as advice
Passage-level citation
Every point names its source text
Exclusions recorded
Documents ruled out are logged as such
Named lawyer concludes
A person owns every legal position
Evidence it leaves behind
What changes after rollout
Who runs the AI Legal Assistant
General counsel
Gets a consistent front door for incoming legal work, so a request arrives already read and routed rather than joining a queue where the first hour of every matter is spent working out what it is.
In-house solicitor
Opens a matter with the relevant provisions already located across fourteen documents and the team’s prior position on the point attached, and spends the time on the judgement instead.
Legal operations manager
Sees which kinds of matter arrive most often and which specialist agent each is routed to, which turns intake from an anecdote into something that can actually be resourced.
Questions about the AI Legal Assistant
What is an AI legal assistant?
It is the entry point to the legal agent cluster: reading every document on a matter together, identifying the provisions that bear on the question by effect, comparing against internal policy, surfacing prior advice, and routing specific work to the specialist legal agents.
How is an AI legal assistant different from a generic chatbot?
A chatbot reads one document you paste. This agent reads the whole matter, finds where documents qualify each other, and cites the passage behind every point it makes.
Can an AI legal assistant run on-premise on legal matter data?
Yes, and some of the material is privileged. Matter documents, prior advice and the question itself stay inside your own perimeter rather than passing through a third-party model.
What does an AI legal assistant produce, and in what format?
A cited matter summary, the provisions bearing on the question with their passages, policy conflicts identified, relevant prior advice, and specialist agent output consolidated.
Where does an AI legal assistant fit in a governed AI programme?
It prepares; counsel decides. Legal conclusions, advice and risk acceptance are reserved to a qualified lawyer, and the four specialist legal agents own their specific tasks.
Does it replace the contract review, research, policy and NDA agents?
No, it is how you get to them. Each specialist does something narrow and does it thoroughly: contract review grades against your playbook ladder, legal research scopes to jurisdiction and verifies citations, policy review compares the internal estate against itself, NDA review triages confidentiality paper at volume. This agent holds the matter, works out which of those the situation calls for, and consolidates what comes back.
Is anything it produces legal advice?
No. It reports what documents say, which provisions bear on a question, where they conflict with policy and what the organisation has concluded before, all cited. Whether a position is defensible, what risk is acceptable and what the organisation should do are legal judgements reserved to a qualified lawyer. The output is labelled as preparation throughout and the file records who reached the conclusion.
What does finding provisions by effect mean?
Searching for what a clause does rather than what it is called. An effective limitation of liability may appear inside a definitions section, an order form or a schedule, under a heading that says none of that. Keyword and heading search miss those reliably, which is why the provision that matters in a dispute is so often one nobody knew was there until it was quoted back at them.
How does it handle privileged material?
It processes everything inside your own infrastructure and respects matter-level access, so the agent working a matter cannot read material from a matter the user is not on. That containment is what makes it usable at all: privilege can be waived by disclosure, and sending privileged documents to a third-party service for analysis is a disclosure whatever the contract says about it.
Can it work on litigation or only commercial matters?
It reads whatever documents a matter contains, including pleadings, correspondence and disclosure material, and produces the same cited preparation. What it does not do is form a view on the merits, draft advocacy, or make any assessment of prospects — those are advice. Its contribution in contentious work is the same as elsewhere: doing the reading fast and citing where every point came from.
Arrive at the matter already read
See the AI Legal Assistant prepare a matter and route the specialist work.