AI Lead Qualification with n8n and HubSpot: What It Costs and How It Works

The brief shows up most weeks. Leads arrive from a form, a chat widget and a spreadsheet. Someone reads each one, decides whether it is worth a call, and types it into the CRM. They are slow on Fridays and they miss the good ones in a busy week.

Automating it is a well-understood build. Here is what it costs and where it actually goes wrong.

What it costs#

Running cost is usually $20 to $80 a month: the LLM calls, an n8n plan, and whatever your CRM already costs.

The architecture that works#

Boring on purpose, because the interesting version breaks at 2am:

1. Capture. A webhook per source. Form, chat, inbox, CSV. Each one normalises into the same shape before anything else touches it.
2. Enrich. Company size, industry, domain. Only if you will actually use it. Most teams buy enrichment and then score on fields they never look at.
3. Score. One LLM call with your criteria, returning structured output, not prose. Ask for a band and a reason, not a number to two decimal places.
4. Route. Hot leads to Slack immediately. Everything else into a queue. The notification is the point; a score nobody sees changes nothing.
5. Write. Create or update the CRM contact, attach the score and the reason, set the pipeline stage.
6. Log. Every decision, with the input. You will need this when someone asks why a lead was marked cold.

Step 3 is maybe a tenth of the work. Steps 1 and 6 are where the time goes, which surprises people every time.

Ask for structured output, not a number in a sentence#

The single most common production break is parsing a score out of prose. The model returns "I'd rate this a strong 8" one day and "This looks like a good fit (7-8)" the next, and the regex that worked in testing silently writes null.

Use the model's structured output or tool-calling interface so the shape is enforced at the API, then validate before writing:

{
  "band": "hot | warm | cold",
  "reason": "one sentence, specific to this lead",
  "missing": ["budget", "timeline"],
  "confidence": "high | low"
}

Bands beat numeric scores. Nobody can tell you the difference between a 71 and a 74, and arguing about it wastes more time than the automation saves.

The three ways this fails#

Nobody defined what qualified means. The team says "score our leads" and never agrees what a good lead is. The model invents a definition, everybody disagrees with the output, and the project is declared a failure of AI.

Fix this before writing anything. Pull your last twenty closed deals and your last twenty wasted calls, and write down what separated them. That document is the prompt. It is a sales conversation, not a technical one, and no tool can have it for you.

The score changes nothing. A number appears in HubSpot. Nobody's day is different. Decide in advance what happens at each band: hot gets a Slack ping and a call within an hour, warm goes into a nurture sequence, cold gets an automated reply. If no behaviour changes, do not build it.

There is no feedback loop. Scores drift as your market shifts and nothing tells you. Once a month, pull the leads marked hot that went nowhere and the ones marked cold that closed. Ten minutes, and it is the only thing that keeps the thing honest.

Handling the parts that will break#

Everything downstream of your webhook belongs to someone else, which means it has bad days:

• Retry with backoff on CRM writes. HubSpot rate limits, and a workflow that gives up on the first 429 loses leads quietly.
• Prevent duplicates. Webhooks get delivered twice. Deduplicate on email before creating a contact, or your sales team will call the same person twice in an hour.
• Alert on failure. An error branch that posts to Slack. A lead pipeline that stops silently is worse than one that never existed, because you believe it is running.
• Cap the spend. One LLM call per lead sounds cheap until a bot submits your form four thousand times overnight.

When to skip the AI entirely#

A surprising number of these do not need a model.

If your criteria are "company size above 50 and country in this list and budget field not empty", that is an if statement. It runs instantly, costs nothing and never hallucinates. Use rules for anything you can express as rules, and reserve the model for the genuinely fuzzy part: reading a free-text message and judging intent.

The best version of this is usually rules doing the obvious filtering and one LLM call doing the judgement, not a model deciding everything.

Hiring for it#

Search for "n8n developer", "CRM automation", or "HubSpot integration developer". Someone who has only done CRM work will struggle with the AI step; someone who has only done AI work will underestimate the webhooks and rate limits.

The screening question: how do you stop the same lead being written to the CRM twice? Anyone who has shipped this answers immediately with idempotency keys or a dedupe lookup. Anyone who has not will say "we check if it exists first", which is the answer that produces duplicates under concurrency.

Get in writing: the workflows exported as JSON into your own repository, credentials in your own accounts, your CRM, your API keys, and a short document describing what each workflow touches.

Related#

This is AI agents and automation work. Our own AI Client Qualifier runs exactly this pipeline, and AI sales and outbound automation shows the larger version.

If you are costing the automation layer generally, what an n8n automation costs covers the pricing and the error handling in more depth. If the chat agent needs to answer from your own documents, what a RAG chatbot really costs prices that separately.