n8n Agency in France: custom automation and AI agents

Concretely, we take a manual, repetitive process — sorting incoming emails, extracting data from PDFs, following up with prospects, generating reports, syncing two applications — and turn it into an automatic workflow that runs 24/7 with no intervention. n8n is the technical building block; AI (language models) is the intelligence layer we plug into it.

The difference with a simple « Zapier freelancer »: we treat n8n as a production system. Error handling, automatic retries, logging, alerting, GDPR-compliant hosting, staging tests. You keep full ownership of the system and can take it back at any time.

Agency, integrator or freelancer: the three labels overlap. Our guide to choosing an n8n agency, integrator or freelancer explains how to tell them apart and which n8n licence each hosting setup requires. For concrete examples, see 12 n8n workflows by industry.

• Data sovereignty. Self-hosted on your infrastructure or a European cloud — nothing passes through a closed US SaaS. Essential in legal, medical and finance.
• Controlled cost. No per-task billing. A modest server absorbs tens of thousands of executions per month at a fixed cost.
• Native AI. Direct connection to Claude, OpenAI, Mistral and vector databases to build agents and RAG pipelines.
• No ceiling. Conditional logic, loops, custom code, more than 400 integrations — where Zapier stops, n8n keeps going.

• Audit & mapping. We identify the highest-ROI processes to automate and cost each piece of work.
• Workflow design. Triggers, API calls, AI steps, error handling, retries — tested in staging.
• Dedicated hosting. Secure self-hosted deployment (Docker, European cloud), backups, updates.
• AI agents & RAG. Connecting Claude / OpenAI / Mistral, vector knowledge bases, autonomous agents.
• Monitoring & maintenance. Alerting, logs, incident fixing, continuous optimization.

Here are automations we regularly deploy for French SMBs and mid-market companies:

• Sorting and auto-replying to incoming emails with an LLM, with human escalation when needed.
• Extracting data from PDFs (invoices, quotes, contracts) into your CRM or ERP.
• Automatic lead scoring and follow-up (see our intelligent real estate CRM).
• Internal chatbot connected to your document base (RAG) to answer your teams' questions.
• Automatic generation and sending of weekly reports from multiple sources.
• Real-time synchronization between applications that do not natively talk to each other.

Two of these deserve their own comparison before you commit to a tool: our guides on document automation software and on AI automation tools lay out where each category actually earns its licence fee.

These building blocks recombine differently by industry. We describe the workflows that actually pay off in real estate agencies, in law firms, in healthcare settings and in HR departments.

The examples below start from the tools named on our four industry pages. Every limit comes from the vendor's or n8n's public documentation, reread on 4 November 2026.

Real estate: Apimo, with no webhook

Apimo does not tell you when a lead comes in. The workflow starts on a Schedule Trigger node, reads the leads or contacts list with the timestamp of the previous pass, then handles whatever changed.

Work out the quota before you build. Reading three lists every 5 minutes uses 864 requests a day out of the default 1,000, which leaves little room for on-demand reads. Spacing out the passes overnight frees up some headroom.

Once the enquiry is qualified, n8n writes back to Apimo: an action (call, viewing) on the contact, or a search with the buyer's criteria. The full flow is described in qualifying enquiries in a real estate agency, and the API operations in our guide to connecting an AI agent to Apimo.

Legal: HubSpot or the firm's practice software

A firm that tracks prospective clients in HubSpot can run all of it from n8n. When a first consultation is booked, the HubSpot node creates or updates the contact, opens a deal and attaches a note to it. The conflict of interest check still happens before the booking.

In the other direction, the HubSpot Trigger node needs a developer account and a public app, and only one trigger can be active. When several workflows depend on HubSpot, they all start from a single workflow that routes the events.

HubSpot retries a notification up to 10 times over 24 hours if n8n does not answer within 5 seconds, and guarantees neither the order nor the absence of duplicates. The workflow therefore checks that it has not already handled the event.

On Secib, the only public connector runs inside Power Automate: it signals a new case file and reads a view, and creates nothing. n8n then handles the calendar and email, and a person opens the file. The detail is on our page about booking consultations in a law firm.

Medical: calendar, reminders and HDS hosting

Doctolib does not publish an open API: its developer documentation requires a login. For a practice on Google Calendar, the Google Calendar Trigger node flags an appointment that is created, updated or cancelled, and n8n schedules the reminder by text message or email.

Health data then passes through n8n itself. With pruning on by default, n8n deletes saved executions after 336 hours, or 14 days. Until then, each execution keeps the data that went through every node. Setting EXECUTIONS_DATA_SAVE_ON_SUCCESS to none stops n8n from keeping successful executions.

Our reading: the instance, its database and its logs then have to sit with an HDS-certified host, the French certification for health data hosting. The component-by-component detail is in our guide on what an n8n workflow must host under HDS, and the reminder flow in appointment reminders for medical practices.

HR: Lucca and PayFit for onboarding

Lucca sends an employee.created webhook when an employee is created. n8n receives it and starts the onboarding: accounts to open, documents to send, first-week meetings.

The limit of 50 requests per minute applies to the company's whole Lucca domain, across every integration. A workflow that rereads the full directory on each new hire eats into the share of the other tools plugged into Lucca: it should read only the record it received.

PayFit states 50 reads and 20 writes per second. Its webhooks are reserved for partners, and its partner page said on 2 October 2025 that new integration requests were paused. Check access before you put a price on the project. The full flow is described in the HR onboarding agent.

Twenty, the open-source CRM

Twenty sends a signed webhook for every record created, updated or deleted. n8n recomputes the HMAC SHA256 signature with the webhook secret and rejects any request that does not match. Twenty cannot yet filter by event type, so the first node in the workflow does the sorting.

Its REST and GraphQL API accepts 100 requests per minute and 60 records per call. Twenty can also be self-hosted, which puts the CRM and the workflows on the same European server. See our page on Twenty CRM integration.

Our approach comes in four phases, designed to deliver fast without sacrificing reliability:

1. Audit (1 week). Process mapping, identification of quick-ROI automations, costed estimate.
2. Build (2–4 weeks). Modular workflows and integrations, tested in staging.
3. Deployment (1 week). Self-hosted production rollout with monitoring, alerting and fallback logic.
4. Optimization (ongoing). Monthly review, integration updates, scaling up.

Want to see how we plug Claude into these workflows too? Read our page on Claude integration for business or our approach to custom AI agents. The detailed platform comparison lives in our resources.

The question comes up in every audit, and the short answer surprises people: a French OPCO does not fund the development of your workflows. Labour law assigns it training and apprenticeship, not engineering work. On a mixed project, only the training share can draw on its funds.

We price the audit knowing which item falls under which scheme, which avoids building a case on a line that will be refused. The full detail, scheme by scheme, with amounts and filing dates, is in our guide on funding an AI project in a French SME. For the obligations that come with training your teams, see the AI literacy obligation in Article 4 of the EU AI Act.