Seth Levanen

AI agents in revenue operations

The agents I build draft the work and a person approves it. This page shows what exists, how each agent is controlled, and what I learned running them. It uses no employer or client data.

Sales Agent Intel Suite

Built as Claude skills for HubSpot. Not released yet. Walkthrough on request.

Fifteen agents that cover the sales workflow, from first research on an account to the hand-off when it changes owners. One setup interview configures all of them, so no agent asks the same question twice.

Setup
A single interview reads the website and the CRM schema first, then asks about target customers, buyer personas, competitors, field mapping, email voice and guardrails. Every agent reads that shared configuration.
Control
Draft and approve everywhere. Nothing is sent and no CRM record changes until a person approves it.
Evidence
Each claim an agent makes carries a quote, a link or a record it came from.
Hand-offs
Agents pass work to each other, for example research into outreach, and a call recap into the right CRM record.
Testing
A QA agent checks the other fourteen against their own specifications.

The fifteen

  • Prospecting: outreach batches matched to a segment and a persona, waiting in an approval queue.
  • Deal progression: every open deal with its momentum, what blocks it and one concrete next step.
  • Deal loss: patterns in closed-lost deals: objections, competitors and failure modes.
  • Customer health: green, yellow or red ratings with cited evidence and next actions.
  • Customer handoff: a complete transition report when an account changes hands.
  • Customer response: answers to inbound questions from an approved knowledge base, with human hand-off rules.
  • Company research: news, website review and CRM summary before outreach.
  • Call recap: a transcript turned into a CRM note and follow-up email.
  • Data agent: plain-language questions answered from the CRM, with changes queued for approval.
  • Commercial performance: metrics, trends and recommended actions on the deal book.
  • Clip search: the exact quote, speaker, account and date from past conversations.
  • Content research: topic ideas, drafts and content gaps.
  • AEO visibility: how a brand shows up in AI answer engines.
  • Suite QA: the regression check described above.
  • Setup: the interview that configures the rest.

Solo go-to-market build: agents in daily use

Built in my most recent role. Described in general terms only.

Scale
More than 30 agents and a dozen recurring automations on Claude and the HubSpot API, built with no engineering support.
Jobs they did
Call prep, competitive monitoring, pipeline forecasting, CRM hygiene, speed-to-lead and executive reporting.
Control
Every output was a draft that I reviewed before it went to anyone else.

My rule of thumb is 80/20. The agent does about 80 percent of the work. I do the other 20 percent: checking it for made-up facts and weak writing before it leaves.

How a run flows

  1. A trigger starts the run: a schedule, a new deal, a call transcript or a question.
  2. The agent reads the shared configuration and pulls the records it needs.
  3. It drafts the result with the evidence behind each claim.
  4. The draft waits in an approval queue.
  5. A person approves, edits or rejects it. Only then does anything reach the CRM, an inbox or a customer.
A flow diagram. A trigger starts an agent that reads a shared configuration and pulls the records it needs. The agent drafts a result with evidence for each claim. The draft waits in an approval queue where a person approves, edits or rejects it. Only an approved draft reaches the CRM or inbox; an edited or rejected draft sends nothing.
One run, start to finish. The person at the approval step is the only way a draft reaches a customer. Swipe sideways to see the whole diagram.