Working demo, built in n8n
Two revenue digital workers running on real executions: a pre-call brief where every claim cites the signal it came from, and a post-call follow-up that updates the CRM and drafts the email but never sends it. Six minutes, narrated.
Each one answers a different part of the question: where is the agent allowed to act on its own, and where is it not.
The brief can only draw on signal rows pulled from a table. The model never touches the internet.
After the model writes, code checks every cited id against what was actually retrieved and strips anything invented.
An account with no signals returns "not enough on file", not a confident brief.
A rep commitment with no agreed date is written as NO DATE AGREED and flagged. Invalid model output writes nothing.
The follow-up lands in an approval queue. The workflow contains no send step at all.
Every brief, draft and CRM write is logged with its flags, so the team can see what the worker did and why.
Every output in the video is from a logged n8n run. Nothing on screen is mocked.
The four accounts, the signals and the two call transcripts are fictional, written to exercise the edge cases. Your CRM and your call notes replace them.
n8n for orchestration, Gemini for the two reading steps, n8n data tables as the CRM stand-in. Swaps to Salesforce, HubSpot or Pipedrive at the node level.