Jason Lemkin Says Your AI Agents Should Be 120% Better Than Your Best Rep. Our Data Shows He’s Only Half Right.

I stumbled on this piece from Jason Lemkin at SaaStr about the 15 hard truths of B2B AI today, and it’s got the kind of founder-battle-scars vibe you can’t fake. His core takeaway? Stop building agents that hit 80% of your best rep. Aim for 120%.

He describes the moment an agent drafted smarter outreach than any human on his team, an inbound agent that booked 682 qualified meetings better than any BDR he’s ever worked with, and social support that actually solves problems inside the comment thread. It’s the “better, not cheaper” gospel, and for once, it’s not just marketing fluff.

Our data backs it up. PainSignal tracks 126 problems in Sales & CRM with an average severity of 3.7 out of 5. Twenty-two of those are specifically about AI for sales automation, with an average severity of 3.8. That’s a lot of unsolved pain—bad lead qualification, missed follow-ups, generic outreach that customers ignore. If an agent can genuinely outperform a top closer, there’s a hungry market waiting.

But here’s where Lemkin’s ear is only half to the ground.

His entire thesis, and the bulk of AI builder energy right now, is trained on the top of the funnel. Sales. Inbound. Outbound. Social selling. It’s the sexy stuff, the revenue-generating stuff, the stuff VCs will wire money for based on a 48-hour glance at growth. But we’re tracking something that should make every founder stop and stare: 47 problems around customer churn prediction and onboarding automation, with an average severity of 4.1 out of 5. That’s higher than sales. That’s higher than marketing. That’s the quiet, boring, retention-focused work that nobody builds flashy Replit demos about, but where the pain is bleeding companies dry.

Think about it. Lemkin jokes that two reps earning six figures each made one closed deal between them and quit without saying goodbye. That’s painful, sure. But what about the customer who churns after three months because they never got onboarded properly? The SaaS subscription that dies silently because nobody noticed the usage drop? The renewal that could have happened automatically if an agent had just checked in at the right time with the right message? Those are 120% opportunities hiding in plain sight, and almost nobody is building for them.

Then there’s voice. Lemkin’s article is heavily text-based in its agent examples—email, social comments, even the “contact sales” gatekeeper he wants to kill. But we’re tracking 89 customer support problems with an average severity of 4.0 out of 5, and a huge chunk of them mention long wait times, IVR hell, and infuriatingly bad call resolution. Voice AI is finally crossing the uncanny valley. The tools exist to build an agent that answers calls, understands context, pulls up account history, and resolves issues without ever putting anyone on hold. It’s a 120% play against the typical outsourced call center, and it’s barely mentioned in the B2B AI conversation.

Lemkin’s vertical AI point is spot on. You need your restaurant upsell agent to know about guacamole margins, not the entire history of marketing. PainSignal data reinforces this hard: there are 1,243 industry-specific problems with severity above 3.5. Specialization wins. But if you’re building that specialized agent, why stop at the sale? The same deep domain knowledge that optimizes the menu should also be able to predict when a restaurant owner is about to cancel because their online ordering setup is broken, and proactively reach out to fix it.

This is the danger of chasing the obvious sizzle. Lemkin’s audience is founders and operators thriving on momentum. But the data says the less glamorous problems—the ones you only notice when you’re staring at a churn report or listening to recorded support calls—are actually rated as more severe by the people living with them. If you’re a vibe coder looking for your next build, you could spend months trying to make a sales agent better than a top 1% closer, only to find the market is already crowded with well-funded teams. Or you could spend a week looking at the 4.1-severity churn problems, and ship something that makes a finance leader’s jaw drop within a month.

And the “something jaw-dropping” advice is, frankly, the most practical takeaway in the piece. Skip the deck. Skip the memo. Put the most magical thing you’ve built live in front of a skeptic. For retention-focused AI, that could be an agent that correctly identifies the 15% of customers most likely to churn this month, generates a personalized save offer, and shows you the projected revenue impact in real time. Show a CFO that, and they won’t ask about your gross margins.

Where our data nudges Lemkin’s advice to the side is on hiring. He says find the internal tool nerd instead of hiring a GTM engineer. That’s smart for the current moment, but our data on engineering demand (anecdotal, but directionally consistent with what we’re hearing) suggests the arms race for truly great engineers is only accelerating. The tool nerd is your first 10%, but eventually you’ll need to hire outside, and the bar is insanely high. If you’re building in the retention or voice support space, you need people who understand the unglamorous plumbing: integrations with billing systems, telephony APIs, customer data platforms. Those aren’t tool nerds; those are domain nerds.

The overall piece is a great gut check. Build agents that beat humans, not undercut them. Go vertical. Trust the slope of improvement, not today’s bugs. But the step Lemkin misses—and where our data practically screams—is to look beyond the revenue engine at the engine that keeps the revenue around. If you’re a builder, the highest-severity, least-hyped problems are your easiest path to a true 120% agent. The sales inbox is full. The retention inbox is wide open.