How Our AI Agent Booked 614 Meetings from 442K Chats, And Why B Leads Are Pure Gold If You Add AI. The Top Learnings from Agent’s Day of SaaStr AI Annual 2026

The final day at SaaStr AI Annual was our agent stack deep dive day.

• Amelia and I walked through the 20+ agents we run our business on.
• Adam Guild from Owner.com showed how they crossed $100M ARR by going all-in on AI three years ago.
• Andrew Bialecki from Klaviyo broke down how a public company rebuilds its product and engineering process around agents.
• Then a pitch comp judged live by Gemini, then a 90 minute AMA.

Here are the 10 things that mattered most.

1. Our Amelia AI (Qualified) Agent Booked 614 Meetings From 442K Chats. Itself.

Numbers from this event alone:

• 2.2M website sessions handled
• 442,000 individual chats
• 614 qualified meetings booked
• Average sponsor ASP around $85K

No human team could have done this. We would have needed 3-10+ BDRs who would turn over every 3-6 months. Instead one agent, well-trained, with access to the full Salesforce context and live website crawls.

Anyone running an AI startup with a “Contact Us” form on their website in May 2026 is leaving money on the table. Replace the form. Today. Qualified is one option.

2. Every Agent We Run Started as a Boring Tool

10K, our AI VP of Marketing, started as a dashboard in January 2026. Just a way to stop copy-pasting numbers from Marketo and Salesforce into Notion.

QBee, our AI VP of Customer Success that now manages 150+ sponsors, started as a project management tool.

Annie, our event producer agent, started as a Squarespace replacement.

None of them were architected as agents on day one. They became agents through 600 to 1,000 commits each, 7 to 8 commits a day, over a few months. That is the pattern. Pick something boring and broken in your stack, replace it with something you can vibe code, then keep adding context and tools until it becomes an agent.

3. Headless Salesforce Is the Highest-ROI Move You’re Not Making

If you do nothing else after reading this, do this one thing.

Spin up Replit or Lovable or v0. Connect it to Salesforce via the API. Build a dashboard, an analysis, or a workflow you cannot do natively in Salesforce.

I haven’t logged into Salesforce in two companies. Now I query it in real time, all the time. Ticket sales by hour. VC attendees by region. Look-alike sponsor scoring. None of this is possible in native Salesforce.

The same goes for HubSpot when their agents launch. Whoever owns your CRM also owns the maximum context for your agents. Use it.

4. Spend MORE Time With Agents, Not Less

The X and OpenClod narrative is that autonomous agents work on their own. That is dangerous nonsense.

The number one lesson from a year of running 20+ agents in production is the more time you spend with them, the better they get. Last week 10K started writing better re-engagement emails than any human marketer we could hire. When I asked Amjad at Replit how, he didn’t know. When I asked their head of field, he didn’t know. The answer is just hours of interaction and context building.

Even your best human gets better when you spend a little time with them each week. Agents are exactly the same. The narrative that you can fire-and-forget an autonomous agent and walk away is the opposite of what actually works.

5. The Same Spec Produces Different Agents on Different Platforms

Amelia rebuilt 10K on Lovable on Wednesday with Elena from Lovable. Same spec, same inputs as the Replit version.

The Lovable version came back with four ad-spend ideas. The Replit version was biased toward email marketing. Different brains, different conclusions, both reasonable.

This matters because picking a vibe coding platform is not interchangeable. The model under the hood and the training data of the platform itself bias your agent’s outputs. We are going to run them in parallel and have them argue with each other about the three best ideas each day.

6. Put Your Agents on B Leads, Not A Leads

A leads close themselves. A million dollar sponsor inbound is going to get a response from your laziest rep in 60 seconds. Don’t put AI there.

The gold is in B leads. Real prospect signal, real ICP fit, but a human rep won’t follow up because the per-lead expected value isn’t worth their time.

Artisan (our slightly-warm outbound agent) recovered about $500K of sponsor revenue from B leads this year. Without that, we eat ramen.

Every founder here has B leads sitting in Salesforce right now. Wire up an agent. The ROI is sitting there waiting.

7. 83% of Owner.com’s New Customers Start With a Free AI Product

Adam Guild from Owner.com walked through this in detail.

Their the AI restaurant website generator that hit 2M+ views on X two weeks ago, costs them about $1 in compute per restaurant. They give it away free, or for $1/month for the first three months.

Result: 83% of new customers start there, then expand to the $500/month bundle. Their CAC profile looks like a cheap product. Their LTV looks like a premium one.

Stop asking what tier you should be in. Start asking what magical thing you can deliver in 5 minutes for free, and whether it pulls people into the full stack from there.

8. DAU/MAU Is the Wrong Quality Signal in an Agentic World

Adam said this and it stuck with me.

In old-school B2B, users logging in daily meant your product was sticky. Today, every login is a failure. It means the user is doing work the agent should have done for them.

The best AI products are the ones users barely touch. The agent runs the workflow. The user gets the outcome. Logins are friction, not engagement.

If your product strategy is still optimizing for time-in-app, you are building for the last decade.

9. Treat the LLM Like an Athletic Middle Schooler

This was Andrew Bialecki’s framing at Klaviyo as a mental model I’ve heard for building AI products.

The base LLM is a generally talented athlete. Good at lots of sports. To make it elite at one specific thing (sending marketing campaigns, scoring leads, handling customer support), you need three things: coaches that train it, drills that put it through reps, and proprietary domain data it cannot get from Claude or ChatGPT directly.

Klaviyo trains their Composer agent against real-time feedback from how consumers actually react across all their customers’ campaigns. That feedback loop is the moat. The base model is a commodity. The harness, training data, and feedback loop are not.

If you are building an agent and you cannot describe the proprietary training loop, your product is one Claude release away from irrelevance.

10. Agents Are Power Users From Day One

Andrew also flipped how we think about product feedback at scale.

In normal B2B, you have a power-law distribution of users. A few power users push the limits, most are novices. You build customer advisory boards to find the power users because they tell you what to ship next.

Agents start as power users on day one. They read your docs, they call every API, they push every edge case in their first hour of usage. The right product feedback loop now is to ask the agent what is holding it back.

Klaviyo’s Composer agent told them it needed better AMP for Email APIs because it could see the ROI lift from interactive email. Humans had been too intimidated to ask for the same feature for ten years.

Stop running customer advisory boards as your primary signal. Start asking your agents what’s broken.

Bonus Lesson: Promote Your AI-Fluent People on Monday

This came up in the AMA. Compensation in the agent era is broken. Your best people are 3 to 10 times more productive than they were 18 months ago. You probably cannot pay them 3 to 10 times more cash. Anthropic and OpenAI can. You cannot.

What you can do is promote them, give them outsized equity, and tell them you love them. Don’t wait for review cycle. Don’t worry about how Jane and John will react. The half-life on talent retention for AI-native operators is dropping fast.

If you even think you should promote someone for what they did with AI this quarter, go back to y