Skip to main content
All projects

Intelligems / 2026 / Sales / RevOps

Post‑Demo Follow‑Up Agent.

The team's best follow-up lived in one rep's habits. I productized it: an agent drafts the email and proposal after every demo, and the rep decides what gets sent.

30–45 min → ~40 sec

per post-demo follow-up

01The problem

All four experienced sales reps were already drafting post-demo follow-ups the same slow way: copy the call transcript, paste it into an AI tool, wait, edit the output, paste it into email. Roughly 30 to 45 minutes of fiddling per demo. Separately, only one rep was building polished proposal decks afterward, even though the proposal was the highest-converting follow-up artifact. Two valuable behaviors, neither of them on rails.

02The decision

I designed it around the rep, not around the automation. The agent does the assembly work: it reads the call and deal context, drafts a voice-matched email, and builds a proposal deck when the conversation calls for one. The rep keeps the decision. Nothing auto-sends; drafts land in the rep's inbox for review, because the one rep's proposals converted precisely because a person shaped them.

03What I built

  • A cloud AI agent reads the call transcript and deal context, then drafts a voice-matched follow-up email.
  • When the conversation warrants it, the same agent generates an on-brand proposal deck.
  • A lightweight orchestration layer handles the trigger and delivers the result into the rep's email drafts.
  • Wired across several system APIs: CRM, transcription, and email.
  • Solved the binary-artifact retrieval pattern needed to pull a generated deck out of a cloud agent session and attach it to an outgoing draft.

04Watch it run

The system running end to end. Account details are illustrative, not a customer record.

05How it runs

The demo ending is the signal. Nobody starts this, and nobody remembers to.
  1. Demo call endsTrigger
    The recording closing is the event. No button, no checklist
  2. Transcript and deal context assembledTransform
    Pulled from the transcription tool and the CRM record
  3. Agent reads the call and decidesRoute
    Reads the call to decide whether a proposal deck is warranted
    alwaysDraft a voice-matched follow-up email
    when the call warrants itAlso build an on-brand proposal deck
  4. Lands in the rep's email draftsLands in
    Deck attached. Never auto-sends; a human always approves
Built for

Account executives, right after a demo

The follow-up is waiting in drafts before they leave the call. Review and send, instead of 30 to 45 minutes of assembly.

It drafts, it doesn't send. The reason the one rep's proposals converted was that a person shaped them, so the human stays where the judgment matters and leaves where it doesn't.

06Getting it reliable

Seven versions to a production-ready v7, each a hypothesis tested against a real transcript. Along the way: fixed email formatting so it rendered cleanly, cut per-run cost by roughly 60%, and enforced a single clean structured output so downstream steps never broke.

07Results

~$0.42
per run
25–45 sec
for an email-only run, versus 30–45 min manual
~60%
per-run cost reduction across seven versions

08Stack

Claude Managed Agents / HubSpot / Pylon / Attention.ai

Built, demoed, and submitted to the company's internal AI-builder competition.

Let’s connect.

I’m looking for GTM Engineering work with B2B SaaS teams in NYC.

Get in touch