Inspectable workflow example

AI Lead Outreach with n8n: Research, Drafts, Human Approval & Replay Safety

Prepare a relevant first message without starting from a blank page for every company. The workflow gathers website context and saves a short draft for a person to check, edit or reject.

Researching each company and writing a useful first message takes attention. AI can help prepare the work, but someone still needs to check the facts and decide whether the message is appropriate. This case shows the draft queue and a separate test of what should happen when evidence is missing, a message changes after review, or the same company appears twice.

Two separate kinds of evidence:

  • Historical n8n demo: the screenshots show real website research and drafts saved to Airtable. That demo was last verified June 14, 2026.
  • November 2, 2026 local lab: synthetic company data, fixed research, a hand-written draft and simulated review decisions. Its 29 passing checks test review rules; they do not verify a live n8n, Airtable or model integration. No message is sent.

Inspect the input, evidence and output or download the runnable review lab v1.0.0 (ZIP).

Partner disclosure. This guide includes a partner link to n8n Cloud. If you sign up through it, FloxoLab may earn a commission at no extra cost to you.

The bottleneck is not the AI.

It is the human decision: check the research, decide whether the message is relevant, edit if needed, then send or skip.

Lead source Airtable rows with pending websites
Research step Groq Compound Mini visits the site
Draft step DeepSeek writes a short message
Screenshots use real public website research from a small test list. Third-party company names and URLs are redacted because those businesses are not FloxoLab clients and are not shown as endorsements. The visible FloxoLab row is included so the research and draft path can be inspected openly.

One fact, one draft, one review decision

Recorded local result: 29 policy checks pass. A four-event batch produces one approved local handoff, one duplicate, one conflict and one rejection. The technical checks below explain how that result is reproduced.

The kit runs with node run.mjs on Node.js 22 or later, without installation or credentials. It includes a retained website snapshot, research JSON, a candidate draft, the executable policy and its tests, and the recorded run. The .example URLs identify synthetic fixtures; nothing fetches them.

1

Input: Northstar Studio (synthetic), campaign fixture-campaign-01, website https://northstar.example/.

2

Source ledger: fact F1 points to S1, its source URL, observation time, exact quote and retained-snapshot hash.

3

Candidate: a subject, fewer than 80 words, and one fact reference. A supported quote must appear in the message.

4

Decision: a named synthetic reviewer approves or rejects that exact fingerprint. An unexpired approval creates only a local handoff record.

The evidence stays beside the message

Source S1: https://northstar.example/services
Observed: 2026-11-02T03:00:00Z
Fact F1: "Request an estimate using the website form."
Retained source: fixtures.json → sources[0].snapshot
Snapshot hash and fact-to-source mapping: results.json → sourceLedger

Candidate subject: A question about estimate requests

Your services page says: Request an estimate using the website form. I build workflows that route incoming requests to a review queue. Would mapping that handoff be useful to your team?

This message is a hand-written fixture, not a new DeepSeek output. The exact quote can be checked against the retained synthetic snapshot. A matching quote proves traceability; it does not establish that the whole message is true. The reviewer still checks the source, the offer and every other claim.

What the run actually checks

node run.mjs
29 checks passed
approved fixture → approved → 1 local handoff
identical replay → duplicate → no new handoff
changed replay → replay_conflict → no new handoff
rejected fixture → rejected → no handoff

Evidence boundary: these checks execute Node.js reference logic. They do not measure model accuracy, website-research quality, n8n execution, Airtable writes, authenticated approvals or sales results. Reviewer events are fixtures; the kit has no recipient list or send method. Its replay map lasts one run, so a process restart can produce the same local artifact again. Durable atomic deduplication, authenticated review records and integration tests are separate work.

The screenshots below belong to the earlier n8n demo, last verified June 14, 2026. They show its research and Airtable output; the November 2 lab does not re-test those services or their current model availability.

The result in Airtable

The final table is the review queue. Good rows get Status = Done, plus a draft message and research notes. Broken, parked, or unusable websites are routed to Status = Blocked or Status = Needs review instead of forcing a fake personalization.

Airtable table showing lead outreach drafts, notes, done status, blocked status, and one visible FloxoLab test row
Drafts and notes land in Airtable for review. The workflow prepares messages; a human still decides what happens next.

The full workflow map

The canvas starts with pending Airtable leads, loops through each row, researches the website, checks whether the research is usable, generates a draft only when the site is usable, saves the result, and waits three seconds before the next lead.

Full n8n canvas for lead outreach drafts with Airtable, Groq research, DeepSeek draft generation, routing, and save steps
The useful parts are visible: input rows, research, parse checks, draft generation, Airtable update, failed-site routing, and pacing between leads.

Research the company first

The Research Company node calls Groq using groq/compound-mini. Its job is narrow: visit the public website and return factual JSON with a research status field. In this workflow, that field is expected to be done, blocked, or unclear.

The research prompt asks for what the company sells, how customers book or inquire, visible calls to action, and a few useful personalization facts. It explicitly tells the model not to invent content for broken or empty websites.

Research Company node showing Groq Compound Mini website research input and structured JSON output
Groq Compound Mini handles the research step and returns structured context for the next model.

Draft from one real detail

The draft step uses DeepSeek to write a short message from the research. The instruction is intentionally restrained: 3-5 short sentences, under 80 words, one specific detail, no fake enthusiasm, no invented statistics, no generic "businesses like yours" line.

Generated Draft node showing DeepSeek output for a short lead outreach draft based on company research
The draft is useful because it has context. It is still only a draft until someone reviews it.

Route bad inputs away from drafting

A practical outreach workflow should not pretend every website is usable. If parsing fails, Airtable gets Status = Needs review. If the website is blocked, broken, parked, or too thin, the row can be updated to Status = Blocked.

n8n status routing section with Parse OK, Site OK, Parse Failed, Prepare Error Info, and Mark Failed nodes
This route is what keeps the workflow honest: bad input becomes review work, not fake personalization.

Historical test costs

These figures belong to the original demo, not the November 2, 2026 policy lab. No new paid model run or current-price comparison was performed for this update.

This pattern is cheap to test because each row is a small research-and-draft job, not a long agent session. In the FloxoLab test account, 16 days of workflow testing showed Current Monthly Usage = $0.26. That included more than this one workflow and a day where we tested around 100 runs across different models. DeepSeek draft testing was under $0.10.

That is not a production guarantee. Real cost depends on website size, model choice, prompt length, retries, and how many leads you run. But for a small outreach test, the bigger cost is usually not model tokens. It is finding a clean lead list and spending the review time well. The B2B prospect list guide shows how to define the market, add exclusions, and validate the first 20 companies before this drafting workflow touches them.

If Apollo is the upstream research tool, use the Apollo free trial checklist to evaluate one qualified segment before this drafting workflow receives it.

1

Airtable: stores the review queue, status, notes, and draft message.

2

Groq Compound Mini: researches each public website and returns compact factual JSON.

3

DeepSeek: writes one short draft from the research instead of a generic template.

4

Human review: protects the workflow from becoming an auto-send spam machine.

Good fits

Niche outreach

A freelancer or small team can review 20-50 researched leads instead of sending the same email to a scraped list.

Shopify wholesale

A store can research potential stockists, draft a note from their product fit, and approve only the contacts that look relevant.

Google Workspace follow-up

Form or Sheet rows can be enriched before a Gmail draft, Viber follow-up, or owner task is created.

Before connecting the review policy to Airtable

Approval status

The local lab checks approval and rejection. A live queue still needs authenticated reviewers, durable version-bound decisions and explicit states for review, expiry and rejection.

Delivery boundary

Keep delivery outside this reference lab. Any later delivery integration needs its own authorization, suppression check, idempotency and delivery record.

Follow-up drafts

Add a 3-5 day follow-up draft for approved leads that did not reply, with a clear stop condition.

Do-not-contact list

Check every lead against skipped domains, existing contacts, unsubscribes, and companies that should never be contacted again.

Channel variants

For some Philippines workflows, the output might be a Gmail draft, LinkedIn note, Viber follow-up, WhatsApp draft, or owner task instead of an email.

Related guide

For the budgeting side of lead work, read How Much Does It Cost to Automate Lead Follow-Up in the Philippines?. It breaks down where the money actually goes once outreach turns into ongoing follow-up.

What you can inspect now

This workflow is useful because it avoids the weakest part of many outreach systems: a generic message sent to every company. It turns public website context into a draft that a person can review.

It is not a promise of sales results. It is a small, inspectable system for doing better first-contact prep: real research in, short draft out, human approval before send.

Want to map one lead workflow?

Send the lead source, the review step, and where the final message should live. We will map the smallest useful version first, including where a human should stay in the loop.

Map my lead workflow