Reference architecture · Illustrative scenario, fictional business
Lead intake and qualification with human approval before any send
An n8n workflow that collects inquiries, scores each one against the ideal-client profile with a language model, and waits for a person to approve in Slack before anything is sent or recorded.
Illustrative scenario, fictional business. Ostrand Vale Advisors is invented to show how I would design this workflow; it is not a client.
Solution architect · Professional services · Design, not built for a client
n8n
CRM
LLM scoring
Slack
Design targets
0Contacts sent without approvalEvery send waits for a person in Slack
300 leads / monthSized forA 12-person consultancy
< 15 minTime to first reviewDuring business hours
< $0.05Model cost per leadTarget, checked against the provider price sheet at build
Context and problem
Illustrative scenario, fictional business. Ostrand Vale Advisors is a 12-person consultancy. Inquiries arrive through a web form and a shared inbox, and the partners answer them when they have time. Good leads wait days; poor fits take up the same attention as good ones. The firm wants every inquiry triaged within minutes, but no message leaving the firm without a partner seeing it first.
Constraints
Reputation over speed. A wrong or tone-deaf reply costs more than a slow one, so automation may prepare, never send.
Existing tools. The firm already lives in Slack and a CRM; the workflow has to meet them there.
No developer on staff. The rules for scoring must be readable and editable by a partner.
Architecture
How to read it. Left to right is the happy path. Nothing to the right of the highlighted box happens until a person approves in Slack. A rejection or a timeout ends the run with the lead parked for review, so the workflow can never contact anyone on its own.
Workflow
Intake. n8n receives the form submission or email, normalizes fields and drops duplicates.
Score. A language model compares the inquiry with the written ideal-client profile and returns a fit score with a two-line rationale and a suggested reply.
Approve. n8n posts the lead, score, rationale and draft to Slack using its “Send and Wait for Response” operation, and pauses.
Act. On approval, the CRM record is created and the reply goes out from the partner’s address. On rejection or timeout, the lead is parked for the weekly review.
Decisions and trade-offs
Approval in Slack, not in the CRM. Partners answer Slack quickly; a CRM task list is where approvals go to wait. Rejected: auto-send above a score threshold, because one bad send undoes the time saved.
Model for judgment, rules for facts. Deduplication, required fields and blocklists are plain workflow rules; only the fit judgment and draft use the model. That keeps cost and behavior predictable.
Profile as a document. The ideal-client profile is a short text the partners edit, read by the model on every run. Changing who counts as a good fit does not need a developer.
Governance and cost
Human approval: no contact and no CRM write happens until a named person approves in Slack.
Data handling: inquiries pass through n8n and the model provider. n8n documents that its cloud service acts as controller and processor under a DPA, while self-hosted data stays on your own infrastructure; choose based on what clients send in their inquiries.
Cost per run: one model call per lead. Cost per lead is input tokens (inquiry plus profile, roughly 2,000) and output tokens (rationale plus draft, roughly 400) at the chosen model’s published price; the target is under $0.05 per lead.
Two levers: keep the profile short, and use a smaller model for scoring with the larger one only for drafts of approved leads.
Failure modes
What breaks
How it is detected
Fallback
Model returns malformed output
Schema check in the workflow
Lead posted to Slack unscored
No one approves
Approval timeout
Lead parked; listed in the weekly review
Duplicate inquiry
Match on email and domain
Merged into the existing lead
Score drifts from partner judgment
Monthly sample of approvals vs scores
Edit the profile text
Outcome
Design targets, not results: zero sends without approval, first review within 15 minutes in business hours, under $0.05 of model cost per lead, sized for 300 leads a month.
What I would measure in a pilot
Time to first response: median under one business hour, against the current baseline.
Agreement: partners approve at least 80% of leads the model scores as a strong fit.
Safety: zero messages sent without an approval record.
What comes next
Follow-up reminders for approved leads that have not replied, using the same approval step before each message.