PayPal Put Agentforce on 8,000 Leads a Month No Human Was Going to Call. Conversions Jumped 50%.

One of the best sessions at SaaStr AI 2026 wasn’t a grand keynote about the future of AI. It was two operators showing exactly what’s working right now in AI agents for sales. Amelia Lerutte sat down with Adam Alfano, President at Salesforce, and Eitan Saban, Head of Sales for North America Mid Market at PayPal to share the honest lessons learned from rolling out AI sales agents at scale in Agentforce.

PayPal turned on an Agentforce AI SDR agent just 14 weeks before SaaStr AI 2026.

It’s fully in production today across a 200-rep org.

And meeting conversion is running about 50% higher than what the humans were doing alone.

Let’s break down what they actually did.

1. The agents aren’t replacing reps. They’re feeding them.

PayPal didn’t deploy agents to cut headcount. They deployed them to work the leads no human was ever going to touch.

PayPal onboards north of 100,000 merchants a month. A huge chunk of them stop processing somewhere in that journey, and there are roughly 8,000 of those leads to chase every month. No sales org has the human power to religiously work 8,000 nudge cycles a month. So those merchants just got dropped.

The agent runs a 10-nudge cadence against every single one of them. It doesn’t get tired. It doesn’t skip the weekend. It doesn’t pick and choose the leads with the fattest commission. As Eitan put it, the win isn’t the agent closing the deal. The win is the agent doing the heavy lifting to book the meeting, so a human shows up to the merchant as the best version of themselves with full context already loaded.

The conversion lift comes from that handoff. Reps start way further down the funnel, on a qualified lead, with the full story already in front of them.

2. “But my data is messy” is not a reason to wait.

This was the best exchange of the session. Someone in Amelia’s vibe coding class the day before asked whether you can deploy an agent if your CRM data is a mess. Adam’s answer was direct: don’t try to solve every problem with your data before you turn an agent on. Just … start.

His logic holds up. Salesforce runs its SDR agent at 70% deliverability, which is better than its human reps. The reason is the agent has an omniscient understanding of a well-curated data environment. That’s the idyllic end state. But most companies aren’t there yet, and that’s fine.

You can stand up a web agent that understands your website, your FAQs, and your product data while you’re still consolidating everything else. It still has a safe, accurate conversation. It still pulls in rich customer intel. And here’s the part most people miss: using the agent is itself the fastest way to figure out what data and structure you actually need. The work runs in reverse. Action bias beats a six-month data cleanup project that never ends.

Even Amelia’s own story makes the point. When she copied the internal Salesforce use case for following up on stale leads, she went to check how many leads SaaStr had never followed up with. The answer was a spreadsheet with a thousand people on it. She didn’t clean it for a month. She gave herself one day, accepted it would never be perfect, and handed it to the agent. Some of those re-engaged leads were in the room at Annual.

3. The real unlock is headless.

This is where the conversation got to the thesis we’ve been hammering all year. The model is a commodity. The data model is the moat. And the way you actually capture that moat is by running your systems headless.

Adam said it plainly: Salesforce has always been a great place for humans to do work. They now think it’s great tooling for agents. The data is in the CRM. The processes are codified. The agent draws inference from structured data and writes back to it with 100% determinism, which is exactly what you want when you’re in a regulated industry and “probably going to do the right thing” isn’t good enough.

The clearest example came from SaaStr’s own setup. Jason doesn’t have a Salesforce seat. He has an API seat. His AI VP of Marketing, an agent called 10K, talks to Salesforce headless through the API, authenticated through both Replit’s native integration and a Salesforce connected app. 10K pulls the full sponsorship pipeline, the media pipeline, the deal list, and the historical revenue data. Jason can sit in the agent and say “pull up sponsors with competitors bringing a bunch of people to Annual this year,” get the list, then say “great, now email it to that person.” The CEO runs a full campaign against the go-to-market team’s data without ever logging into a CRM.

Adam’s vantage point is Slack. He manages a team of thousands through Slackbot and rarely logs into Salesforce directly. The agents live where the work happens, and the inference in Slackbot is unmetered, so you can take the data for a ride without burning tokens.

The lesson: stop thinking of your CRM as a place humans go to fill in fields. Start thinking of it as tooling your agents execute against.

4. Conversational data is the new training set.

There’s a second data vector most teams aren’t thinking about. It’s not just the structured records. It’s the conversations.

PayPal feeds the Agentforce agent every conversation from Gong and the transcripts behind them, plus account data hosted in Seismic. The agent learns exactly why deals were lost and what drove lower conversion. Adam made the broader point: human-to-human, agent-to-human, and increasingly agent-to-agent conversations all feed back into the dataset. A lot of the orchestration is now coming straight off the transcript and conversational intelligence.

Pull all of that together and the agent has far more context to work with, which means it does a better job on the actual problem. The companies winning here treat every conversation as part of the dataset instead of letting it disappear after the call.

5. Agents have a maturity curve. Treat them like teammates.

The other reframe from the session: agents aren’t magic, they’re employees you have to onboard.

PayPal’s Agentforce agent in week one, working 200 leads, is a completely different agent than the one now working 8,000 leads. In a month it’ll work 80,000 leads a week. That maturity journey takes teaching, troubleshooting, and tuning. Adam called it caring for the agents. Dan Darcy’s customer success team at Salesforce describes it as giving each agent a “virtual mom” to steer it in the right direction. Neglect the onboarding and the agent develops the same flaws a poorly trained human would.

The same example shows up in the tuning itself. SaaStr’s outbound agent got explicit context that some of these leads are stale and might be grouchy about being ghosted, so be more personable than a person typically would. That’s not a prompt you write once. That’s ongoing management of a teammate.

6. Bigger orgs need orchestration. Smaller orgs just need to vibe it.

The deployment playbook splits by company size, and both ends were on stage.

SaaStr AI is, in Amelia’s words, three humans and a dog. So when they need an agent, Jason vibe codes it himself in Replit and hooks it up. Every person who needs to sign off is already in the room.

PayPal is the opposite. In a highly regulated environment you cannot kick off an agent without marketing approval and without compliance having veto power. Eitan’s single biggest learning from 14 weeks: you cannot run it yourself. You have to bring marketing, compliance, and the broader org along, all bought into the agent’s outcome, before you scale it.

The mistake in the middle of that spectrum is having a great agent that works fine in isolation and then stumbling on how to integrate it into the real work process. The integration is the hard part now. Standing up the agent in Agentforce takes a couple of seconds. Orchestrating the team and the agent around it is where the work lives.

And critically, the agent carries a quota. Eitan’s team built an AI institution that brings the use ca