Our inbound AI agent has had 17,000 conversations with prospects in the last 12 months and booked about 600 meetings for SaaStr AI Annual. Together with the newer self-serve agent we added on top of it, inbound drove a 60% increase in new business. Three humans run all of this.
Amelia walked through the full build on the latest episode of The Agents. This is the step-by-step version: what we replaced, how the first agent is set up, and how we added the newest agents on top of it.
What we replaced: a long form, a round-robin, and a one-day lag
Thirteen months ago, a prospect on the sponsor page hit a long contact form. It came to Amelia. She round-robined it to herself or David. Someone replied within about a day.
The reply was nearly always: “Hey [company], you look like a great fit for SaaStr, let’s book a time.” Amelia calls it the worst email on planet Earth. The prospect reached out to you, and the first thing they get back is a form letter.
That’s the bar. Almost anything beats it, which is why the first agent paid off so fast.
Part 1: The inbound agent on the site
Step 1. Put the agent on your highest-intent page first
Ours lives on the SaaStr AI Annual sponsor page. That’s where people who are already evaluating a ~$90K purchase land. We run it on Qualified, as an avatar called Amelia AI.
Start with the one page where a fast answer is worth the most money. Don’t start with your homepage.
Step 2. Give it a real qualification job
The agent answers questions in real time, but its actual job is to qualify. Ours works through:
- Why they want to sponsor
- Budget
- What they’re buying for: lead gen, brand awareness, or speaking
- Which competitors they’re watching or asking about
Every one of those answers makes the first human call better.
Step 3. Let it book the meeting on the spot
No handoff, no waiting for a rep to write an email. The agent books directly. This is the single biggest change from the old flow, because the old flow lost people in the gap between “submitted a form” and “got a decent reply.”
Step 4. Open the call with what the agent captured
When a prospect tells the agent what they want, we start the meeting there. A real example from the pod: “You said you were interested in coffee and newsletters. Coffee sold out, but let me walk you through newsletters and the other things we have.”
The prospect doesn’t repeat themselves, and the first ten minutes of the call aren’t discovery.
Step 5. Measure the full funnel, not the chat count
Our last 12 months:
Some of those 17,000 conversations were goofball conversations. That’s fine. Meetings and closed deals are the numbers that matter, and we had real logos come through this path, including OpenRouter.
One caveat: we drive the top of that funnel with content and community. The agent converts intentional traffic. It doesn’t create it.
Step 6. Know who this works for
This works because our buyers are tech-centric and increasingly AI-native. They do discovery on their own, talk to the agent, book their own meeting, and close. For those buyers, seeing an agent in the sales process is part of the evaluation. If you sell yourself as a top AI event and the buying experience is a PDF and a two-day wait, they notice.
A non-tech buyer who’s happy scheduling a call two weeks out may not care. Know your buyer before you assume this transfers.
Step 7. Don’t make it the only door
A real share of buyers will not talk to an avatar. Some find it intimidating. Most just want the packages and the pricing without a conversation yet.
We debated killing the self-serve download entirely. We kept it, and that decision is what led to Part 2.
Part 2: How we added the newest agents
For a year, the self-serve path was a download link to a Google Slides prospectus. It converted worse than the agent. The fix came from the agent itself: after we built our renewal agent, it told Amelia she already had most of the pieces and should do the same for inbound.
Inbound is harder than renewals because you know much less about the person. So everything below is about getting the most out of the little you do know.
Step 1. Replace the PDF with a tokenized page on your own site
Shorter form. On submit, the prospect gets their own version of the prospectus on our site, with a link unique to their company. Ours is built on Replit.
This matters because a static PDF tells you nothing after it’s downloaded. A page you host tells you everything.
Step 2. Add heat mapping
We use Microsoft Clarity. 10K, our AI VP of Revenue, picked it as the vendor itself.
On the Base44 lead the pod walked through, the heat map showed time on the package overview, a little on Super Gold, most on Gold, then a long stretch on the contact form. He’d also told us in the form that he was interested in Gold and Super Gold, so the heat map confirmed what he’d said.
Step 3. Wait 10 minutes before doing anything
Most self-serve visitors click off after 5 to 7 minutes. We set the wait at 10 so the session is finished and the heat map is complete before the agent starts working.
Step 4. Run first-party signals before anything else
The agent checks, in order:
- Have they been on the site (Qualified, Vector)
- Are they already in one of our outbound sequences
- Have they seen our recent LinkedIn or X ads
- Are they on the newsletter
- Have they attended an event
- Have they spoken at one
- Have we written about their company on SaaStr.com
Then competitors, which are the second-biggest signal. Third-party data comes after that.
The first-party layer is the part no vendor can sell you. On Base44, it surfaced that their CEO had previously come to SaaStr, which Amelia didn’t know.
Step 5. Route it everywhere at once
When the lead lands, the agent:
- Emails Amelia
- Pings her in Slack
- Adds it to a live dashboard queue
- Writes it to Salesforce, with the heat map and custom link attached
That last one used to be a quarterly Google Sheet upload.
Step 6. Have the agent write the pitch, and a human approve it
The agent builds the narrative in 30 to 60 seconds: why this company should be at SaaStr, which of their competitors were at Annual, and recommended packages.
Use only public or anonymized results about other customers. For Base44, the pitch referenced that Replit was top of the leaderboard as the number one sponsor, which is public. Nothing proprietary about any customer goes into another customer’s pitch.
Amelia reviews, gives feedback if needed, and approves. The email goes out from her, not from a generic address.
Step 7. Update the same link in place
This is the piece to copy first. The prospect downloaded something generic. After the pitch is approved, the same URL now opens with “Marlin, here’s why Base44 should be at SaaStr,” with their competitors and packages in it.
Any time they return to the link, they see the personalized version. It becomes a living document you can keep updating for the whole sales cycle.
Step 8. Build your own booker
David used Calendly. Amelia used Read AI. Neither connected to anything. 10K suggested building our own, and it took the agent 20 minutes.
The booking link now shows the prospect’s company name, ties back to the prospectus they looked at, and tracks whether they opened it without booking. When someone bounces, the agent tells Amelia and drafts the next email.
The calendar looks like a minor detail. At 20 minutes of build time it was worth doing, and it closed the last tracking gap in the inbound flow.
Step 9. Route by who owns the most similar accounts
Base44 went to Amelia, not David. David owns Vercel, but Amelia owns Replit and Lovable, and the agent weighted that volume higher. When the agent can identify similar companies in your book, it sends the lead to whoever knows that type of customer best.
The result on one real lead
Base44 inbounded, got the personalized follow-up, and replied that it was a great follow-up and fast. He booked for Friday.
What it runs on
- 10K on Replit as the backend: dashboard, queue, and the layer that writes to Salesforce
- Salesforce, run headless. We rarely log into the UI
- Qualified for the on-site agent
- Microsoft Clarity for heat mapping
- Zoom for meetings
- Our own booker, built by the agent
You don’t need to be an engineer to build this. We aren’t.
Where this goes wrong
Trying to build it all at once. Build it stair-step. The on-site agent came first and ran for a year before the self-serve agent existed.
Letting the backend app get too big. For a few weeks 10K got noticeably worse. It told us it had too much in it: too much data, too many APIs. We cleaned it up and modularized, and the quality came back. Newer models help here too. Fable 5.1, which launched September 1, gives better advice on what to build and improve in a system this size.
Forgetting the rest of the sales team. Amelia works through 10K directly. David doesn’t have the same backend access, so he still logs into Salesforce more. The agent’s next proposal is a version of itself built for reps.
Assuming one inbound path. We run both doors in parallel today. The long-term goal is to merge them so everyone gets the personalized experience whether they talk to the agent or self-serve. We haven’t done that yet.
The build order we’d use again
- The on-site agent on your highest-intent page, booking meetings directly
- Keep the self-serve path, and move it off a PDF onto a tokenized page you host
- Heat mapping and a wait window
- First-party signal checks, then competitors
- Agent-written pitch with human approval, updating the same link in place
- Your own booker, tied into all of it
We ran step one alone for a full year before adding the rest.
