8,000 Leads No One Was Calling: Inside PayPal's AI Sales-Agent Rollout

Every sales team sits on a pile of leads it never gets to. At PayPal the pile had a precise size: roughly 8,000 merchants a month who had started onboarding with PayPal, gone quiet somewhere in the process, and then sat untouched because the North America mid-market team simply didn't have the people to chase them all.

Speaking at SaaStr AI 2026, Eitan Saban, PayPal's Head of Sales for North America Mid Market, walked through what happened when the company handed that pile to a single Salesforce Agentforce agent instead of writing it off. Fourteen weeks after go-live, the agent had scaled from 200 leads in its first week to 8,000 a month, and meeting conversion was running roughly 50% higher than the human team had managed on its own. It's a story worth reading closely, because Saban was specific about how it worked rather than just what it produced, and most of what made it work has very little to do with the model itself.

Capacity was the constraint, not lead quality

It helps to be clear that these weren't cold leads. They were merchants who had already raised their hand by beginning onboarding and then stalled, which by most reasonable definitions makes them warm. The reason nobody called them was arithmetic rather than intent: a rep can only work so many accounts in a week, and a ten-touch follow-up across thousands of merchants a month was never going to happen by hand, so the leads aged out quietly, month after month.

That is why PayPal aimed the agent at this group first rather than at its hottest opportunities. When the realistic alternative is doing nothing at all, the downside of letting an autonomous agent have a go is close to zero, while the upside is pipeline that would otherwise have evaporated. This is also where lead-qualification agents change the underlying economics, because the cost of working one more lead stops being a staffing decision and starts being a rounding error.

The agent books the meeting; the person runs it

What the agent actually does is deliberately narrow, and that restraint is a large part of why it works. It runs a persistent ten-nudge sequence against every lead in the pile, the kind of steady, unglamorous follow-up that no rep sustains across 8,000 accounts, but it stops at the point of booking. Once a meeting is on the calendar a human takes over. As Saban put it, the agent does the heavy lifting to book the meeting so that a human shows up to the merchant as the best version of themselves, with full context already loaded.

That division of labour is what separates a genuinely useful AI sales agent from the automated outreach everyone has already learned to ignore. Send more messages without relevance and you burn sender reputation and train buyers to tune you out; PayPal did the opposite, letting the agent absorb the repetitive qualification work so that scarce human attention went only to conversations that were already warm and briefed. We saw the same pattern in Definity's contact-centre rollout, where the AI carried the routine volume so that agents could concentrate on the moments where judgment genuinely mattered.

The data is what made it convincing

An agent is only ever as good as the context it can reason over, and this is where the deployment becomes instructive. PayPal's agent isn't working from a generic script; it's fed every conversation from Gong along with the transcripts behind them, plus account data held in Seismic, so it can reason about why deals tend to slip and what actually moves a merchant toward a yes.

That grounding is the part most "we deployed an AI SDR" announcements quietly skip, and it's the part that produced the lift. The 50% didn't come from the model on its own; it came from wiring the model into PayPal's real sales intelligence. Teams that treat an agent as a chatbot bolted onto a CRM tend to get chatbot results, whereas teams that treat it as a system grounded in first-party data get outcomes worth putting in front of a board.

The hardest part wasn't technical

Asked for the single biggest lesson of those fourteen weeks, Saban didn't point to anything about the model or the tooling. He said you cannot run it on your own. Getting the agent into production meant bringing marketing, compliance, and the wider organisation along, all of them bought into what the agent was there to achieve, well before anyone tried to scale it.

That maps almost exactly onto what separates the companies that get value from AI from the ones that stall, which is rarely the technology and almost always organisational readiness, ownership, and governance. We looked at that gap in detail in why enterprise AI stalls and what the pacesetters do differently. Seen in that light, PayPal's compliance and marketing sign-off wasn't friction slowing the project down; it was the thing that let the project survive once it was live in front of real customers.

Built for production, not for a demo

The figure worth dwelling on isn't really the 50%, it's the trajectory. The agent worked 200 leads in its first week, reached 8,000 a month by week fourteen, and now runs fully in production across a 200-rep organisation, with a stated target of handling 80,000 leads a week. Most corporate AI never travels that far, because it demos well, earns a slide, and then quietly expires.

What PayPal did differently was scope the first version tightly enough to actually ship, with one agent, one clearly defined pile of dead leads, and one metric everyone could agree on, and then scale from something that already worked rather than from a promise. That is the line between a pilot and production, and it is also the line between AI that shows up in a quarterly review and AI that shows up in the revenue number.

What to take from it

Strip away the logos and PayPal's rollout is a pattern most teams can follow. Start where doing nothing is the current baseline, since the safest place to prove an agent is the valuable work that already gets no human attention. Design for handoff rather than replacement, so that reps experience the agent as something handing them warmer conversations rather than something aimed at their jobs. Ground it in your own conversation and account data before you judge the results, because that context is what turns plausible output into persuasive output. Bring marketing and compliance in ahead of scale, not after it. And ship something narrow first, widening it only once it holds up in production.

None of that requires an AI strategy in the abstract. PayPal took one expensive, well-understood problem, solved it with a system, measured what happened, and grew from there, which is really the whole point of building systems that drive growth rather than experiments that generate headlines. If you're trying to work out where an AI sales agent might earn its place in your own funnel, the pile of leads nobody is calling is usually the best place to start looking.