B2B logistics · Growth Engineer Lead · 2025

AI workflows

IN PRODUCTION

Context

Reps already talk in the tools they live in. The point was to make that the interface, and let the workflow do the typing.

Trigger

Qualification fields were missing often enough that everything downstream ran on guesses.

The build

01
Intake where reps already are
A chat surface takes the note — typed, voice or a photo. Slack in this case; Telegram or WhatsApp would work the same way.
02
Transcribe and read
n8n transcribes voice and reads images, so the format of the input stops mattering.
03
Check against official data
Company details are verified against official government sources — INSEE in France, the FMCSA API in the US — rather than trusted as typed.
04
Write to the CRM
A structured record is written to HubSpot, with the original note kept attached for anyone who wants to check.

Stack

n8n
Slack
OpenAI
HubSpot
INSEE
FMCSA

Result

~10 MIN

of manual CRM admin removed per qualified lead

Honestly

What I'd do differently

I first had one AI step do everything: read the note, decide what the company was, and write to the CRM. When a record was wrong, there was no way to tell whether the model misread the note or matched the wrong company. Splitting it into extraction (what did the rep say) and verification (does this company exist in INSEE/FMCSA) made every error traceable to one step. I should have built it that way from day one.