AI Agents for Lawyers

Use Cases

AI Client Intake: Agents That Qualify, Conflict-Check and Onboard

AI client intake agents qualify leads, run conflict pre-checks, collect KYC documents and draft engagement letters. The flow, the fields and the guardrails.

By Dr. Niklas SchmidtUpdated 13 November 20268 min read
On this page
  1. What an intake agent actually does, and what it must never do
  2. The flow, step by step
  3. The fields to collect
  4. Conflict pre-checks: the step most firms get wrong
  5. Tools: where to build it
  6. Guardrails: consent, confidentiality and unauthorised practice
  7. A system prompt you can adapt
  8. What to do this week

An AI intake agent should do four things and refuse a fifth. It qualifies the enquiry, captures structured facts, runs a conflict pre-check and collects onboarding documents, then hands a clean file to a lawyer. It does not advise, price or accept the engagement. A three-lawyer German firm running this on a EUR 20–30 a month server cut client acknowledgement from one or two days to minutes.

What an intake agent actually does, and what it must never do

Intake is the one workflow in a law firm where the counterparty is not yet a client, the data is not yet privileged and the volume is highest relative to fee value. That makes it the best place to start with agents and the easiest place to get sued if you get the boundary wrong.

The boundary is this. Everything up to "a lawyer will review your enquiry" can be automated. Everything after it needs a human: the decision to act, the fee quote, the scope, the engagement letter signature. The FTC's final order against DoNotPay (February 2025) fined the "world's first robot lawyer" $193,000 and banned claims that a service "operates like a human lawyer" without competent evidence. That order was aimed at a consumer product, but the wording is a ready-made drafting standard for the disclaimer at the top of your bot.

The numbers on what automation returns are unglamorous and specific. The German firm documented in a Lyron case study (spring 2026) runs self-hosted n8n, Nextcloud, local OCR and a local model on a German server: file creation fell from 15–40 minutes to a five-minute check, deadline-relevant items missed went to zero, and the firm reclaimed six to seven hours a week after a six-week rollout. A Zapier-plus-DocuSign stack costed at about $210 a month by Automation Atlas (April 2026) moved intake from five business days to same day and document assembly from 45–60 minutes to 8–12. Neither firm let the software practise law.

The flow, step by step

The skeleton below is tool-agnostic. The playbook on building an n8n intake workflow implements it node by node; the vibe-coded intake app is the front-end alternative.

Enquiry arrives (web form / email / phone transcript)
        |
        v
[1] Identify as AI + disclaimer shown; consent to processing recorded
        |
        v
[2] Capture Tier-1 fields (identity, parties, matter type, urgency)
        |
        v
[3] Conflict PRE-CHECK against PMS / conflict DB  --> hits?  --yes-->  HOLD: human reviews before any facts taken
        |  no
        v
[4] Capture Tier-2 fields (facts summary, documents, deadline, source)
        |
        v
[5] Classify + score (practice group, complexity, fit, urgency)
        |
        v
[6] Route: partner / associate / decline template  -- always a human decision on "act"
        |
        v
[7] On acceptance: KYC/AML document request, engagement-letter DRAFT, calendar slot offered
        |
        v
[8] Human signs engagement letter; matter opened in PMS; audit log closed

Two design choices matter more than any tool. First, the conflict pre-check sits before the facts are taken. If the prospect explains their dispute with your existing client in step 2, you may have received confidential information from an adverse party before you knew they were adverse. Second, "route" is a recommendation, not an action. The agent proposes; a named person disposes. Gartner's rule of thumb, quoted in the agents hub, is that agents belong where decisions are needed; here the decision is precisely the thing you keep.

The fields to collect

Collect in two tiers. Tier 1 is what you need to run a conflict pre-check and route. Tier 2 is what you need to open a file. Do not merge them.

TierFieldWhy it existsFormat the agent should enforce
1Full name, company, roleIdentity; conflict searchFree text, then normalised (legal name, trading name)
1Other parties and known affiliatesConflict searchList; agent prompts for parent companies and group names
1Matter categoryRoutingClosed list matching your practice groups
1JurisdictionRouting; licensingCountry plus state/Land where relevant
1Urgency and hard deadlineTriageDate field; agent asks "is there a court or contractual date"
1Source of enquiryBD reportingClosed list
2Facts summaryFile openingFree text, capped at 300 words; agent asks for gaps
2Documents heldChecklistUpload or list; no upload before pre-check clears
2Preferred meeting slotsBookingCalendar picker
2Consent to AI-assisted processingCompliance recordExplicit yes, timestamped

Note the cap on the facts summary. Longer free text is where prospects paste emails from the other side, and where a chatbot with an over-eager model starts to opine. Keep the field short and the model constrained.

Conflict pre-checks: the step most firms get wrong

A pre-check is not a conflict check. The formal check runs against your full database with a person reading the hits; the pre-check is a coarse filter that protects that person from receiving confidential facts too early. Build it as a deterministic search first (exact and fuzzy matches on normalised names against your practice-management system) and let the model do only the normalisation and the summary of hits. Models are good at recognising that "Müller Holding GmbH" and "Mueller Holdings" might be the same entity; they are bad at deciding whether a hit is a real conflict. That judgment stays with the person who would be sanctioned for getting it wrong.

If your practice-management system exposes an API or an MCP server (Clio does; iManage and NetDocuments shipped MCP servers in May 2026), the agent can query it under a scoped credential that can read the conflict index and nothing else.

Tools: where to build it

OptionBest forCost anchor (as of September 2026)Watch out for
Practice-management AI (Clio Manage AI tiers)Firms already on the platform$49–$149 per user per month for Clio Manage AI tiersIntake logic is theirs, not yours; check where data is processed
n8n (self-hosted or cloud)Control over data path; EU hostingSelf-hosted from EUR 20–30 a month all-in per the Lyron studyYou maintain it; see the n8n for law firms page
Zapier or Make with a model stepFastest to ship; 9,000+ app connectors on ZapierAbout $210 a month for the Zapier-DocuSign stack cited aboveEach step routes data through a US SaaS; DPA required
Custom Claude or GPT app (vibe-coded)A branded front endNear zero to prototype; hosting extraNever test with real client data; use synthetic records
Vendor intake agents (Ironclad Intake, LegalOn Intake, Lawmatics)In-house teams and high-volume consumer firmsEnterprise pricing, mostly unpublishedBuilt for contract requests or consumer leads, not bespoke matters

Zapier's own legal team runs its intake and triage ticketing on Zapier, which is a reasonable proof that the low-code route holds up in a legal department. For a solo or a five-lawyer firm, the small-firm guide argues for starting with the form-plus-automation route rather than a chatbot: it is cheaper, it is auditable and prospects fill in a well-designed form faster than they type to a bot.

Identify as AI. Since 2 August 2026, Article 50 of the EU AI Act requires chatbots to tell people they are interacting with an AI system; the Digital Omnibus deferred the high-risk rules but left Article 50 in place. The AI Act page for law firms has the detail. A one-line notice at the top of the chat satisfies it; hiding the bot behind a human first name does not.

Confidentiality of the data path. Prospects assume that anything they type to a law firm is confidential. Judge Rakoff held in United States v. Heppner (S.D.N.Y., February 2026) that a defendant's consumer-Claude memos were not privileged because the vendor's terms allowed logging and disclosure, while signalling that privilege might attach where counsel directs the AI use; the Harvard Law Review note has the reasoning. The lesson for intake is not that AI breaks privilege but that the data path decides. Use business or API tiers that do not train on inputs, sign a data-processing agreement, host in the EU where your bar expects it (ÖRAK's first rule is that client data stays on EU servers; German firms need a § 43e BRAO agreement alongside the Article 28 GDPR contract), and write the intake record into your own system rather than leaving it in a vendor's chat history. The privilege page covers the US and European positions side by side.

Unauthorised practice. Docketwise's August 2026 survey found 40% of immigration lawyers cite unauthorised-practice worries about AI, and they are right to. The model must be told, in its system prompt, that it collects and classifies and never advises. It should also be told what to do when asked for advice, because prospects will ask.

Disclosure and trust. Clio's 2025 Legal Trends Report found 78% of clients want to be told when AI is used and 36% say they would trust an AI-using lawyer less. An honest, short notice at intake costs nothing and pre-empts the awkward conversation later; the client disclosure page has wording.

A system prompt you can adapt

Module 4 of the AI Lab for Lawyers course builds an intake agent like this live in n8n and Make; the prompt below is the core of it.

You are the intake assistant for [FIRM]. You are an AI system and you say so
if asked or at the start of any conversation.

YOUR JOB
1. Collect the Tier-1 fields: full name; company and role; other parties and
   any parent or group companies; matter category (choose from: [LIST]);
   jurisdiction; urgency and any hard deadline; how the person found us.
2. STOP after Tier 1 and return a JSON object. Do not ask for facts of the
   matter until the system tells you the conflict pre-check has cleared.
3. If cleared, collect Tier-2 fields: a facts summary of at most 300 words,
   documents held (names only, no uploads), preferred meeting slots, and
   explicit consent to AI-assisted processing.

YOU MUST NOT
- Give legal advice, assess prospects of success, or say whether a claim is
  strong, weak, in time or out of time.
- Quote, estimate or discuss fees. If asked, say a lawyer will explain fees
  at the first meeting.
- Confirm that the firm will act. Say: "A lawyer will review your enquiry
  and contact you within [X] business hours."
- Ask for or accept documents before the pre-check has cleared.

WHEN ASKED FOR ADVICE
Reply once, in plain language: "I can collect the details so a lawyer can
review them, but I cannot give legal advice." Then continue the intake.

OUTPUT
Return only the JSON schema provided. Fill "confidence" with low/medium/high
for the matter-category and urgency fields. Never invent a field value; use
null and add the question to "open_questions".

Run it under a business or API account, log every exchange to your own store, and have a lawyer read the first fifty transcripts. You will find the questions prospects actually ask, which is the most useful intake research most firms never do.

What to do this week

  1. Write the disclaimer using the DoNotPay wording as the negative: nothing that suggests the bot operates like a lawyer.
  2. Split your current intake form into Tier 1 and Tier 2 and move the conflict pre-check between them.
  3. Pick the build route from the table above and price the data path, not the licence; the confidentiality checklist has the questions to ask the vendor.
  4. Prototype with synthetic data only, following the four-step method Clifford Chance's Jamie Tso describes for vibe-coded tools: one pain point, a written user experience, a platform, then iterate.
  5. Set the routing rule so that "we will act" is never a state the agent can reach on its own; the human-in-the-loop entry explains why that single gate matters more than the model you choose.

Frequently asked questions

Can an AI chatbot do client intake for a law firm?

Yes, for the parts that are administrative: capturing contact details, matter type, urgency, opposing parties, documents and availability, and running a conflict pre-check. It must not give legal advice, quote fees as binding or confirm an engagement. Those steps stay with a lawyer, and the bot should say so in plain words and identify itself as AI.

How does an AI conflict check work?

The agent extracts party names, affiliates and adverse parties from the intake form, normalises them (spelling variants, trading names, group companies) and searches the firm's conflict database or practice-management system. It returns candidate hits with a confidence score. A person then reviews the hits and runs the formal check before any confidential facts are taken.

Is client data entered into an intake chatbot confidential?

Only if the data path is. A consumer chatbot that logs or trains on inputs is not a confidential channel; a US court held in February 2026 that a defendant's consumer-chatbot memos were not privileged. Use a business or API tier with no training, a data-processing agreement, EU hosting where your bar requires it, and store the record in your own system.

What should a law firm intake form ask?

Enough to qualify and conflict-check, no more: name and contact details, role (individual or company), other parties and their affiliates, matter category, jurisdiction, urgency and any deadline, how they found you, documents they hold, and preferred meeting slots. Facts of the matter come only after the conflict pre-check clears.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Playbooks

Build a Client Intake Workflow in n8n: From Web Form to Matter Opening

Build an n8n client intake workflow node by node: form trigger, AI classification, conflict pre-check, engagement letter, lawyer approval, DMS entry and tests.

14 min read

Playbooks

Vibe-Code a Client Intake App as a Lawyer (Lovable, Claude Artifacts, v0)

Vibe-code a client intake app as a lawyer with Lovable, Claude Artifacts or v0: the spec, four exact prompts, a fee-band calculator and the IT handover.

8 min read

Tools

n8n for Law Firms: Building Legal Automations Without Code

n8n is a self-hostable workflow builder with AI nodes. A German firm runs intake and deadlines on it for €20–30 a month; three law-firm workflows, node by node.

9 min read

Guides

AI for Small Law Firms and Solo Practitioners: What to Use and What to Skip

AI for small law firms in 2026 means a business-tier model, one legal tool and simple automation for under €200 a month, plus the policy and privacy settings most solos skip.

9 min read

Regulation & Ethics

Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots

Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.

10 min read

Regulation & Ethics

The EU AI Act for Lawyers and Law Firms: What Applies in 2026

The EU AI Act applies to law firms as deployers. Article 4 literacy and Article 50 transparency are live; high-risk duties are deferred to December 2027.

9 min read