12 n8n workflows by industry: real estate, legal, healthcare, HR (with real costs)

Generic examples ("send a Slack message when a form is submitted") are everywhere. The ones in this guide start from a business problem and show what an n8n.io template leaves out: where the human sits, which data must stay out of the prompt, and what it costs to run.

Each card follows the same structure, matching the five steps in the diagram below.

1. Trigger: what starts a run (an email, a form, a webhook, a time). It sets how many executions you are billed for on n8n Cloud.
2. AI step: extraction, classification or drafting, through the Anthropic node, an AI Agent or a Mistral model. The AI produces a structured suggestion.
3. Human approval: with Human-in-the-loop for tools, the agent pauses the run before using a tool and sends an approval request in Slack, Teams, Gmail, Outlook, WhatsApp Business Cloud, Telegram, Discord, Google Chat or n8n Chat. If rejected, the action does not run and the AI is told.
4. Action: the write to the CRM, calendar or inbox.
5. Audit log: one row per run with the execution ID, the decision and, for personal data, the purge date.

Node names are n8n's own. n8n.io templates are free; their update dates are as shown in the gallery on 15 September 2026, and their node lists are summarised.

Sales and lettings agencies repeat the same tasks every week: sorting enquiries, following up, checking applicant files. Industry constraints are covered in our AI for real estate guide. One more useful template outside the three cards: no. 17510 by Oneclick AI Squad, which builds a listing package from room photos with Claude.

In a law firm, the constraint that shapes the architecture is professional secrecy: every document sent to an AI model leaves your premises. Use cases and specialised tools are covered in our AI for legal guide.

Two constraints apply to any healthcare workflow in France: health data entrusted to a provider must be hosted by an HDS-certified host, and Doctolib has neither an n8n integration nor a public API. Specialised tools are compared in our AI for healthcare guide.

Recruitment is the most regulated case of the four industries: AI systems that filter job applications are listed in Annex III of the AI Act. Other HR use cases are covered in our AI for HR guide.

Running cost has two independent lines: the platform that runs n8n, fixed each month, and the tokens the AI model consumes, proportional to volume. Build and maintenance come on top if a provider builds the workflow.

The platform

On n8n Cloud, the annual billing toggle reads "Save 17%". On monthly billing, the dollar version of the pricing page showed $24 for Starter and $60 for Pro in our capture; we could not display the monthly euro amounts, so check the page before subscribing. The Business plan adds self-hosting, SSO and Git version control for 40,000 executions.

On the pages we checked, n8n does not publish an official minimum hardware configuration. The two VPS above are each host's entry offer with 2 cores and 4 GB; queue mode with workers needs more (see below).

Tokens, worked out

Assumption: each run sends 2,000 tokens to the model (instructions and a short document) and gets 300 back (a structured extraction). 1,000 runs therefore use 2 million input tokens and 0.3 million output tokens. Prices are per million tokens (MTok).

Three corrections to apply:

• The tokenizer. Anthropic says its 4.7 and later models produce about 30% more tokens for the same text. For Claude Sonnet 5, the same document becomes about 2,600 input and 390 output tokens: 2.6 × 2 + 0.39 × 10 = 5.20 + 3.90, or $9.10 per 1,000 runs.
• Volume. At 5,000 runs a month, multiply by five: $2.40 with Mistral Small 4, $17.50 with Haiku 4.5, $35 with Sonnet 5. On n8n Cloud, that volume exceeds Starter's 2,500 executions and requires Pro.
• Discounts. Anthropic's Batch API costs 50% less, useful for overnight processing, and cache reads are billed at 0.1 times the input price when instructions repeat.

Full example: a firm extracting 1,000 documents a month with workflow 5 on a Hetzner CX23 and Claude Haiku 4.5 pays €5.49 excl. VAT for hosting and about $3.50 in tokens. Anthropic has announced Haiku 4.5 will be retired no earlier than 15 October 2026: plan to test another model. Other Claude prices and worked examples are in our Claude API pricing guide.

The pricing page defines an execution as one full run of the workflow, however many steps or however much data. The docs add two exceptions:

• runs started manually in the editor do not count, so your tests are free;
• a sub-workflow called by another does not count: only the parent workflow is counted.

The consequence follows from the trigger. A workflow started by a Schedule Trigger every 5 minutes runs 288 times a day, or 8,640 executions over 30 days, even when there is nothing to process: three and a half times the Starter quota. The same processing triggered by an email or a webhook only uses the runs that matter. Steps shared by several workflows (logging, purge, alerts) are good candidates for sub-workflows.

A template imported from the gallery works in a demo. In self-hosted production, five settings prevent most incidents.

1. An error workflow. It must start with an Error Trigger node and is set in Workflow Settings, under Error workflow. One handler can serve all your workflows and send a Slack or email alert. The Stop and Error node forces a failure when a business condition is not met (unreadable document, forbidden file).
2. N8N_ENCRYPTION_KEY. It encrypts the credentials stored in n8n. If you do not set it, n8n generates one on first start and stores it in ~/.n8n: set it yourself, back it up off the server, and give every queue-mode worker the same value.
3. Queue mode, only when needed. It splits a main instance, Redis and workers (EXECUTIONS_MODE=queue). It requires PostgreSQL, since SQLite is not supported, and S3 storage for binary files. With an MCP Server Trigger and several webhook replicas, all /mcp* traffic must go to a single replica.
4. The built-in audit. n8n audit (or a POST to /audit) reports on credentials, database, filesystem, nodes and the instance, including unprotected webhooks and an outdated version. Run it after every update.
5. Locked-down triggers. A Chat Trigger offers three modes: no authentication, Basic Auth or n8n User Auth. For an internal tool, avoid "None" and keep "Make Chat Publicly Available" off.

The Community edition has no SSO (SAML, LDAP), Git version control, environments, external secrets or log streaming. A law firm or clinic that needs those controls looks at the Business or Enterprise plans. Agent-level safeguards (prompt injection, tool permissions) are covered in our AI agent security guide.

An unpatched self-hosted instance is the main weakness of an n8n workflow. Four critical vulnerabilities published between December 2025 and February 2026 are enough to set a minimum version.

• 1.x line: at least 1.123.17 for these four. Advisory GHSA-6xcw-7xm6-48c6 of 2 September 2026 (sandbox escape in the legacy expression engine, score 7.7, default engine not affected) is only fixed in 1.123.76.
• 2.x line: at least 2.5.2 for the four, and 2.37.7 or 2.38.2 for the September advisory. The latest stable release is 2.39.5, published on 14 September 2026.
• The pace: ten security advisories were published on 2 and 3 September 2026 alone. Plan a monthly update, and act within days for an unauthenticated flaw such as CVE-2026-21858.

n8n's Sustainable Use License allows use and modification for your own internal business purposes or for non-commercial or personal use, and files marked .ee. fall under a separate Enterprise licence. Who can build your workflows, and where to host them, is covered in our article on n8n agencies, integrators and freelancers.

A community node is third-party code running inside your instance, with access to its credentials: read it, pin its version and test it on every n8n update. For a CRM, the HTTP Request node against a documented API is often easier to maintain. If you are considering an open, self-hostable CRM, see our Twenty CRM page.

Lumyniq, an AI automation agency based in Paris, designs n8n workflows and AI agents for small businesses in real estate, legal, healthcare and HR, and connects them to the tools already in place. Every project starts with an audit of the process to automate, before any quote. See the n8n offer and custom AI agents.