Automation
Chatbots and CRM automation that feed the loop.
The gap
A chat widget on the site that nobody tracks. Leads scored by whoever picks up the phone. A CRM that is updated on Fridays, if at all, so nothing useful ever goes back to Google.
What we build
Lead classification before upload, where the client agrees it is appropriate. An enquiry can be assessed against a written rubric, with uncertain cases routed to a person. The classification can inform conversion value and routing after a labelled test and ongoing checks against the CRM. In one internal client example, reported cost per qualified lead moved from about €50 to a €20-30 range after the feedback loop was introduced. That is an account result to verify from the underlying records, not a promised causal effect.
Chatbots that capture, book or hand off within an agreed scope. The first interaction identifies the system as AI, a person remains available for handoff, and any CRM write is logged. This supports Article 50 transparency where the EU AI Act applies to a system interacting directly with people.
CRM workflows in n8n. Stage changes trigger uploads. New customers join the exclusion audience. Quotes that go quiet get a follow-up. The CRM stays current because the automation does the boring part.
Workflows can be self-hosted in n8n so the implementation remains portable and there is no n8n per-execution charge. Hosting, model, messaging and other API costs still apply and are documented.
What you get
The scoring workflow with the rubric written in plain English so you can change it. The chatbot, its disclosure copy and a transcript log. The CRM workflows, documented. Monitoring that tells us when something stops running.
Sources checked
Related: CRM feedback loop · Tracking
FAQ
Will an LLM misjudge leads?
Yes, it can. We test a labelled sample, monitor it against later CRM outcomes, retain a human override and adjust or stop the workflow when the error rate is not acceptable.
Do we need HubSpot?
No. Many CRMs can connect through a maintained n8n node, API, webhook or controlled export. We check authentication, fields, rate limits and error handling before saying an integration will work.
What about data protection?
Raw submissions stay in your CRM and your self-hosted n8n instance. Only the minimum fields needed for classification are sent to the model, through API endpoints under the provider's business data terms, and the retention position is documented per implementation.