AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. The €25-a-month law firm
  2. What n8n is, and why "self-hosted" is the point
  3. Workflow 1: intake, from web form to opened file
  4. Workflow 2: deadline capture with a human check
  5. Workflow 3: the Friday regulatory digest
  6. Security and professional secrecy
  7. When Make or Zapier is the better answer
  8. A build order for your first month

A three-lawyer employment and tenancy practice in Germany runs its post-room, client onboarding and deadline capture on a self-hosted n8n server that costs €20–30 a month, and reclaimed six to seven hours a week within six months. n8n is a visual workflow builder you can host yourself, with AI nodes that call any model. For a law firm it is the automation tool that keeps client data where you put it.

The €25-a-month law firm

The case worth studying is small and specific. The German firm combined self-hosted n8n with Nextcloud for files, a local OCR engine and a local language model, all on a server in Germany with no licence fees. The workflows sort incoming post and email, open files, send acknowledgements and capture deadlines with a lawyer confirming each one. Rollout took about six weeks (Lyron case study, spring 2026).

TaskBeforeAfter six months
Sorting the post and inbox2 hours a day40 minutes a day
Opening a new file15–40 minutesA 5-minute check of a pre-filled file
Client acknowledgement1–2 daysMinutes
Deadline-relevant items missedNot reported0
Time reclaimed6–7 hours a week
Running cost€20–30 a month

Two details in that study matter more than the numbers. The deadline step is human-in-the-loop: the workflow proposes, the lawyer confirms. And the firm treated the setup as a regulated outsourcing, putting a contract under § 43e BRAO and an Article 28 GDPR processing agreement in place for the hosting. That is the shape of a defensible automation under the ÖRAK and BRAK guidance: data on EU servers, no client data used for training, the lawyer keeping final responsibility.

What n8n is, and why "self-hosted" is the point

n8n is a node-based workflow builder from a Berlin company. You drag a trigger (a form submission, an email, a schedule) onto a canvas, connect it to actions (create a record, send a message, call a model, wait for approval) and the workflow runs on a server. There are native nodes for Gmail, Outlook, Slack, Teams, Google Drive, OneDrive and SharePoint, generic HTTP and database nodes for anything else including Clio's API and most document management systems, and AI nodes that call Anthropic, OpenAI, Google, Mistral or a local model.

The distinction that Anthropic drew in December 2024 still governs how to use it: workflows are "systems where LLMs and tools are orchestrated through predefined code paths", agents "dynamically direct their own processes and tool usage" (Anthropic). Almost everything a law firm should automate first is a workflow with a model doing one bounded job inside it. n8n also ships an AI Agent node that lets the model choose tools, and Gartner's rule of thumb applies: "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval" (Gartner, June 2025). The workflow agent glossary entry has the vocabulary.

Hosting is the decision that separates n8n from its competitors.

OptionWhere data sitsWho maintains itCostFit
n8n CloudVendor-hosted; check the current data-region pagen8nSubscription by usageFirms without anyone to run a server
Self-hosted on a rented server (Docker)The data centre you choose, e.g. Germany or AustriaYou or your IT providerServer only; the German firm pays €20–30 a monthMost European small and mid-size firms
Self-hosted on-premisesYour own hardwareYour ITHardware and timeFirms with existing infrastructure and strict clients

Whichever you choose, model calls are the part that leaves the building. The GDPR guide sets out the Article 28 questions; the short version is that a provider under business terms with a processing agreement, an EU-domiciled provider, or a local model on the same server are all defensible, and a consumer API key on a personal account is not.

Workflow 1: intake, from web form to opened file

This is the German firm's onboarding flow generalised, and the n8n intake playbook builds it in full. Node by node:

  1. Form Trigger. An n8n-hosted form on your site: name, email, counterparty, matter description, urgency, consent checkbox.
  2. IF: required fields present. Missing consent or counterparty ends the run with an email asking for the missing item.
  3. AI node: classify and extract. The model receives the description only and returns JSON: practice area, matter type, urgency (1–3), parties named, one-line summary. Prompt below.
  4. HTTP Request: conflicts search. Query your practice-management or DMS API for each party name; return hits.
  5. IF: conflict hit. Any hit routes to step 6; none to step 7.
  6. Send and Wait: partner decision. A Teams or email message with the hit, two buttons (proceed, decline). The workflow pauses until someone clicks.
  7. HTTP Request: create matter. Create the file in Clio or your system with the extracted fields.
  8. Outlook: acknowledgement. A templated email to the client with the file reference and what happens next.
  9. Database or Sheets: log row. Timestamp, fields, conflicts result, who approved.
You are a classification step in a law firm intake system.
Input: the matter description below. Output: JSON only, keys
practice_area (employment | tenancy | commercial | other),
matter_type (one short phrase), urgency (1 = court deadline or
dismissal within 14 days, 2 = within 30 days, 3 = none stated),
parties (array of names exactly as written), summary (max 25 words).
Do not give legal advice. Do not infer facts not in the text.
If the description is empty or unreadable, return
{"error": "unreadable"}.

The rule that makes this safe is in the prompt: the model never sees anything but the description, never advises, and returns fields a paralegal checks in the pre-filled file. The client intake use case covers the conflicts and engagement-letter steps that follow.

Workflow 2: deadline capture with a human check

Deadlines are where automation earns trust or loses it forever, which is why the workflow proposes and a lawyer disposes.

  1. Outlook Trigger. New email in a "Courts" folder, or a scanned-post folder in Nextcloud or SharePoint.
  2. Extract attachments. PDFs only; anything else is logged and skipped.
  3. OCR. A local Tesseract call through the Execute Command node, or a cloud OCR node under business terms. Feed the text, not the image, to the model.
  4. AI node: extract. Document type, court, case reference, date of service, every sentence containing a period or deadline, verbatim. No arithmetic.
  5. Code node: compute candidates. Your own rules (for example, service date plus the statutory period, adjusted for weekends) produce candidate dates. Deterministic code does the counting, not the model.
  6. Send and Wait: lawyer approval. An email or Teams card to the responsible lawyer: document, extracted sentences, candidate date, three buttons (confirm, edit, reject).
  7. On confirm: calendar and file. Create the deadline in the practice-management system and Outlook calendar entries at 14, 7, 3 and 1 days out, the reminder ladder a small-firm Zapier stack documented in April 2026 also uses (Automation Atlas).
  8. On edit or reject: task. A task for a paralegal with the original document attached.
  9. Log. Every run, every decision, who confirmed.

The design principle is that the model reads and quotes; the code counts; the lawyer decides. The German firm reports zero missed deadline-relevant items in six months on exactly this pattern, and the human-in-the-loop glossary entry explains why the approval node is the control rather than an inconvenience.

Workflow 3: the Friday regulatory digest

A weekly digest is the safest place to let a model summarise, because every item carries its source link and nothing is acted on automatically. Anthropic's regulatory-legal plugin does a richer version (feed watcher, policy diff, comment deadlines); this is the twenty-node equivalent.

  1. Schedule Trigger. Friday, 07:00.
  2. RSS Read, one node per source. Regulator and legislature feeds relevant to your practice.
  3. Merge and Filter. Items from the last seven days; optional keyword filter.
  4. HTTP Request. Fetch each item's page and extract the text.
  5. AI node: summarise. Prompt below; input is the fetched text and the URL, nothing else.
  6. Markdown to HTML. Group by source.
  7. Outlook or Teams. Send to the practice group.
  8. Log. Items sent, items that failed to fetch.
For each item below, write exactly two sentences: what changed and
who it affects. Use only the fetched text provided. Do not add
background, dates or citations that are not in the text. End each
item with its URL on its own line. If the fetched text is empty or
looks like an error page, output the title and URL with the note
"(could not fetch, read at source)". Order: items with the word
"deadline", "consultation" or "in force" first.

The regulatory monitoring use case discusses what to do once a digest has proved reliable for a quarter, including adding a diff step against your own policy documents.

Security and professional secrecy

n8n stores credentials encrypted on the server, but a workflow is only as safe as its least careful node. Five settings to fix before client data flows:

  • Own the server. A rented machine in a German or Austrian data centre under a written contract; the German firm's § 43e BRAO and Article 28 paperwork is the template.
  • Choose the model deliberately. Business API terms with no training and a processing agreement; an EU-domiciled provider; or a local model for classification, which is what the case-study firm did. Cheap models suit bounded steps: Claude Haiku 4.5 costs $1 per million input tokens and $5 output as of September 2026.
  • Never export credentials. Workflow JSON exports can carry keys; a June 2026 report counted 24,008 secrets in MCP configuration files on GitHub, 2,117 of them valid (Practical DevSecOps).
  • Vet community nodes and MCP servers. The same report found 82% of 2,614 public MCP servers vulnerable to path traversal. Use official nodes where they exist; n8n's AI Agent node can call MCP servers, which is powerful and exactly where to be careful. The MCP glossary entry explains what you are connecting.
  • Register it. Microsoft Agent 365, generally available since 1 May 2026, treats automations as identities to govern, and n8n was a launch partner (Microsoft Security). If your firm runs it, register the workflows rather than hiding them.

The agent security guide covers prompt injection, which matters here in one specific way: workflow 2 reads documents sent by opposing parties. Keep the model's job to extraction and quotation, and keep the send-and-wait node in front of every calendar entry.

When Make or Zapier is the better answer

QuestionChoose n8n ifChoose Make or Zapier if
Who maintains it?You have an IT provider or a partner who enjoys thisNobody will patch a server
Where must data sit?Bar rules, clients or insurers require your own infrastructureVendor hosting under a processing agreement is acceptable
How many runs a month?Thousands; self-hosting has no per-task feeDozens to hundreds
How complex is the logic?Branching, loops, code steps, local OCR or modelsLinear trigger-to-action
How fast do you need it?Weeks, including setupThis afternoon

Vendor risk cuts both ways. Relay.app, a well-regarded no-code agent tool, closed to free users on 15 August 2026 and to paying customers on 14 September 2026, telling them to export workflows as JSON (Relay.app); OpenAI announced in June that its Agent Builder shuts on 30 November 2026. A self-hosted n8n instance cannot be switched off by a vendor, which is a real advantage, and it also means nobody else will keep it running. The Zapier and Make page makes the case for the hosted tools in detail.

A build order for your first month

Week one, stand up a self-hosted instance with your IT provider on a server in your jurisdiction and put the hosting contract and processing agreement in the vendor file. Week two, build the Friday digest, because it touches no client data and teaches you the canvas. Week three, build intake against a test form and synthetic data. Week four, build deadline capture and run it in shadow mode for a month alongside your existing process before anyone relies on it. Module 4 of the AI Lab for Lawyers course, taught by Dr. Niklas Schmidt of Wolf Theiss, builds an n8n workflow live in the browser with no installs, which is the fastest way to see the canvas before you commit a server to it. The German firm's six-week rollout is a realistic target; six to seven hours a week is a realistic reward.

Frequently asked questions

Is n8n free for a law firm to use?

The self-hosted Community Edition costs nothing to run beyond the server, which is why the German firm in the case study pays only €20–30 a month for hosting. n8n Cloud, the vendor-hosted version, is a paid subscription priced by usage. Model calls from the AI nodes are billed separately by the model provider, unless you run a local model on the same server.

Does n8n keep our client data in the EU?

If you self-host on a server in Germany or Austria, all workflow data, credentials and execution logs stay on that server. The pieces that can leave are the model calls: use a provider under business terms with an Article 28 processing agreement, an EU-domiciled provider, or a local model. Check n8n Cloud's current data-region page before relying on the hosted version for client data.

Can n8n build an AI agent, or only fixed workflows?

Both. Most legal automations are workflows in Anthropic's sense, predefined paths with a model doing one bounded step such as classification. n8n's AI Agent node also lets a model choose tools and steps itself. Start with workflows; use the agent node only where the task is open-ended, and always with an approval step before anything is sent, filed or calendared.

What is the difference between n8n, Make and Zapier for a law firm?

Zapier is the fastest to learn and has the most integrations, Make gives more control in a visual canvas, and n8n is the only one you can host entirely on your own infrastructure. For a firm whose bar rules or clients require data to stay in the EU, or that expects thousands of runs a month, n8n is usually the right answer; for a first simple automation, Zapier or Make is.

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