AI for Legal Intake, Triage, and Routing
On this page
TL;DR: Ask an in-house team where the day goes and a surprising share is intake overhead: deciphering a three-paragraph email to find the actual ask, chasing missing details, re-explaining that NDAs go through the form, and telling stakeholders “it’s with legal” in fourteen threads. None of that is legal work, and all of it is tractable for AI, extraction, classification, routing, and status communication over messy text. It is also the lowest-risk workflow in this cluster, provided one boundary holds: the intake layer routes legal questions and never answers them. This guide covers the intake pipeline, the triage rules, the advice-creep problem, and what to measure. As everywhere in this cluster: nothing here is legal advice, and a qualified lawyer owns every legal outcome the intake system routes toward.
This guide is part of the AI for Legal hub. Intake is the front door to the other workflows: contract requests flow into review or template drafting pipelines downstream.
What intake AI actually does
Four functions, all process-layer:
- Structuring. Turning a forwarded email chain or chat message into a normalized request: requester, business unit, request type, counterparty, ask, stated deadline, attachments. A large language model is genuinely strong here, extracting structure from messy human text is its home turf, and the output is instantly checkable against the source message.
- Triage. Scoring urgency and complexity against written rules: a litigation-hold trigger is not a routine NDA, a regulator letter is not a marketing-claims check.
- Routing. To the right destination: the commercial lawyer who owns that business unit, the template drafting pipeline for standard paper, the self-service policy page for questions legal already answered publicly, or straight to the GC for the categories you define.
- Status and expectations. Acknowledgments, SLA-based updates, and “here’s what legal needs from you to proceed” follow-ups, the communication overhead that erodes lawyer time without requiring a lawyer.
What it must not do is the fifth thing everyone is tempted by: answering the question. More on that below, because it is the failure mode of the entire category.
The intake pipeline
- One front door. A form, a dedicated address, or a chat entry point, pick one primary channel and route the strays into it. AI can normalize requests from anywhere, but a single queue is what makes SLAs and metrics possible.
- AI structuring pass. Every request becomes a ticket. A working prompt pattern:
You are an intake assistant for an in-house legal team. From the request below, produce a structured ticket: requester, business unit, request type (one of: [your taxonomy]), counterparty, one-sentence ask, deadline as stated by the requester (quote their words; if none stated, write NONE STATED, do not infer one), attachments referenced, and missing information legal will need. Rules: do not answer any legal question in the request; do not advise the requester; if the request mentions litigation, a regulator, a government inquiry, a data incident, or employee discipline, set flag ESCALATE-HUMAN regardless of other classification. If you cannot classify the request confidently, set type to UNCLASSIFIED for human review. Request: [text]
- Rules-based triage on top. Urgency comes from written rules the team authored, request types, counterparty classes, deadline proximity, escalation flags, applied to the structured ticket. Keeping the rules explicit and human-authored (rather than letting the model improvise urgency) makes triage auditable and tunable. This is prompt engineering in the plainest sense: the quality of the taxonomy and rules you write is the quality of the triage you get.
- Route, with defaults that fail safe. Confident-and-routine goes to the assigned queue or the self-service path. UNCLASSIFIED, ESCALATE-HUMAN, and anything matching your high-stakes categories goes to a person, same day. Requesters always have a visible urgent-override that bypasses the machine entirely.
- Close the loop. Acknowledgment with the ticket summary (“reply if we’ve misread the ask”, a free accuracy check by the requester), SLA-based status updates, and a missing-information chase if the ticket flagged gaps.
Wired end-to-end, structuring, routing, status, this is a light AI agent pattern. The agent moves work between humans; it does not resolve legal matters.
The advice-creep problem
Every legal intake deployment drifts toward the same cliff. The requester doesn’t want a ticket; they want an answer. “Can I sign this?” “Is this clause OK?” “Can marketing say this?” A helpful assistant will try to answer, fluently, plausibly, and without the facts, the current law, or the judgment. That answer is unreviewed legal output delivered under the legal department’s banner, subject to the same hallucination failure that produced fabricated citations in Mata v. Avianca, except here nobody ever reviews it, because it went straight to the requester.
Defenses, in order of importance:
- Scope the system prompt hard. “You do not answer legal questions, evaluate legality, or advise. You structure, route, and point to approved resources.” Include refusal examples.
- Give it somewhere to send people. “I can’t advise on that, I’ve routed it to [lawyer] with priority X” plus links to approved, lawyer-published self-service content. Pointing to a published policy your team wrote is routing; generating an answer is advice. Keep that line bright.
- Audit transcripts. Sample weekly for boundary violations. Drift happens through edge cases, a “quick clarification” here, a summarized policy there.
- Label the layer. Requesters should know they’re talking to an intake assistant that cannot advise, so nobody later claims they “cleared it with legal” on the strength of a bot acknowledgment.
Confidentiality and privilege at the front door
Intake messages carry the facts of disputes, incidents, and sensitive plans, often exactly the material your team wants privilege over. Three rules:
- Sanctioned infrastructure only. The intake assistant runs on an enterprise deployment or legal-ops platform with contractual confidentiality and training on inputs disabled in writing, per the legal hub channel rules and your AI acceptable use policy. Never a consumer chatbot fronting the legal department.
- Counsel sets retention and access. Intake records are discoverable business records unless handled otherwise; who can see tickets, how long transcripts persist, and how privileged matters are segregated are questions for your own counsel in your jurisdiction, decided before launch, not after the first litigation hold.
- Sensitive categories skip the funnel. Whistleblower reports, harassment complaints, and suspected privileged-dispute material should have a direct human channel, advertised alongside the assistant. Some conversations should never have a bot in the loop.
Measuring whether it works
Intake is the most measurable workflow in this cluster. Baseline for two weeks before the pilot, then track:
| Metric | What it tells you |
|---|---|
| Time from request to routed ticket | The core speed win, hours or days down to minutes |
| Misroute rate (tickets re-routed by a human) | Triage quality; tune taxonomy and rules against it |
| Escalation-flag precision | Whether fail-safe categories fire when they should, review every miss |
| % requests resolved by self-service pointers | Deflection of already-answered questions |
| Requester corrections to ticket summaries | Structuring accuracy, measured for free |
| Lawyer hours on intake admin per week | The number that justifies the project |
A realistic 90-day outcome is a structured queue with measured routing accuracy, a visible drop in intake admin time, and clean escalation behavior, at which point the same discipline extends into the contract review and drafting pipelines the intake layer feeds. For the org-level rollout sequence, see the AI adoption roadmap.
FAQ
Can an AI intake assistant answer employees’ legal questions? No. It structures, routes, and points to lawyer-approved resources. Answering the question is unreviewed legal advice under legal’s banner, scope it out, and audit transcripts for drift.
What if the AI triages something wrong, marks an urgent matter routine? Design fail-safes: high-stakes categories escalate to humans by default, requesters get an urgent override, and a person reviews the queue daily during the pilot while you measure and tune misroutes.
Do intake requests create privilege issues in an AI tool? They can, intake carries dispute facts and sensitive plans. Sanctioned infrastructure only, and have counsel set retention, access, and segregation rules before launch.
Do we need a legal-specific intake tool? Not to start: a form plus an enterprise assistant routing into your existing ticketing system proves the workflow. Buy the dedicated front door when volume justifies it.
Next in this cluster: the requests this pipeline routes most often land in AI contract review and template-based drafting, or return to the AI for Legal hub.
Not sure which legal workflow to start with? Take the free AI readiness assessment, ten minutes, and you’ll get a prioritized starting point for your team.
Frequently asked questions
Can an AI intake assistant answer employees' legal questions?
No, and this boundary is the whole design. An intake assistant structures the request, points to approved self-service resources (published policies, the NDA request form), and routes to a lawyer. The moment it starts answering 'can we do X?' it is generating unreviewed legal advice under your legal department's banner. Scope it in the system prompt, test the boundary, and audit transcripts for drift.
What if the AI triages something wrong, marks an urgent matter routine?
Design for it rather than hoping against it: uncertain or high-stakes categories escalate to a human by default, requesters get a visible 'this is urgent' override that skips triage entirely, and a person reviews the queue daily during the pilot. Measure misroutes and tune. AI triage should beat the status quo, an unstructured inbox, not a perfect router that doesn't exist.
Do intake requests create privilege issues in an AI tool?
They can. Intake messages often contain the facts of a dispute or a sensitive plan, material your team may want privilege over. Run intake on sanctioned infrastructure with contractual confidentiality and training disabled in writing, keep it out of consumer tools, and have counsel set the retention and access rules for intake records in your jurisdiction.
Do we need a legal-specific intake tool?
Not to start. A form plus an enterprise AI assistant that summarizes, classifies, and routes into your existing ticketing or matter system proves the workflow. Dedicated legal front-door tools add matter-management integration, reporting, and self-service libraries, worth it once volume and metrics justify them.