Amelia Lerutte | SaaStr Summary

Participants: Jason Lemkin (Chief AI Evangelist, SaaStr) & Amelia Lerutte (Chief AI Officer, SaaStr)

1. Key Themes

The Maintenance Gap Is the Real Agentic Problem Nobody Talks About

Building vibe-coded apps and agents has become accessible to non-technical people, but the ongoing maintenance requires a fundamentally different skill set that most organizations are not staffing for. The agents drift, hallucinate, and break — not just from code changes but from silent model updates — and someone product-savvy needs to own this daily.

"You need someone pretty damn product savvy to maintain it if it's at all complicated... They have no intuitive sense of how software is built or made. I don't think they could fix most of the issues we see because they wouldn't be able to describe or identify them to the agent." — Jason Lemkin 00:13:25

"If you don't have a database manager or somebody who's in charge of all the agents, you may not notice. It may lapse. You may not notice for a couple days, especially if the front end is live, that some of your agents on the back end have broken." — Amelia Lerutte 00:09:47

Model-Level Regressions Are a Silent Killer of Agentic Workflows

Silent or dot-release updates to underlying AI models (like Claude) can introduce hallucinations into previously stable, production-grade agentic workflows — even when no code was changed. This is a new category of technical debt that organizations don't yet have processes to manage.

"The code did not change. It was interestingly, probably because of subtle changes to the model introduced in this complex workflow hallucination... It went from working perfectly to maybe 5% of them producing completely anomalous results. And so who's going to manage that?" — Jason Lemkin 00:21:27

"No Lead Left Behind" Is the Universal ROI Justification for AI Agents

The single most transferable insight from SaaStr's agentic journey is not that agents are smarter than humans — it's that they provide complete coverage with zero judgment. Every inbound contact, every B-lead, every customer with a friction point gets touched in real time, something humans systematically fail to do.

"If you touch every single lead, every prospect, every customer, the way they want to be interacted with in real time, you have to do better. That's all these agents have to do to add value to the organization is that there's no lead, no prospect, no customer left behind." — Jason Lemkin 00:37:41

"Even at our scale, I have not touched every single lead... My agents will start to sit idle. And this last week, it's the most they've been idle since I've gotten them all up into production. Leads behind. We're leaving some leads behind. Even with our agents." — Amelia Lerutte 00:39:42

2. Contrarian Perspectives

AI Agents Don't Replace Accountability — They Expose the Lack of It

The common narrative is that agents handle work humans don't want to do. The contrarian reality is that agents also catch humans trying to game the system — and do so more consistently than managers ever could.

"In the age of AI, you really can't hide. We can see it. You uploaded a placeholder... Maybe I wouldn't have got some of this a year ago, but we're checking everything now because I've got my agents running." — Amelia Lerutte 00:04:03

"QB doesn't get triggered. QB just says, thank you. I sent you 28 emails saying this is what we needed. It's late. Understood. Could you please do it by tomorrow?" — Jason Lemkin 01:05:14

Complicated Pricing Is Always a Hidden Price Increase — And Now AI Agents Enforce It Against Customers

Jason makes a pointed claim most SaaS professionals would resist: pricing complexity is a monetization tool disguised as customer friendliness. And agents trained on new pricing schemes will systematically steer users toward the most expensive path.

"When a company introduces more complicated pricing, even if they say it's a better deal, it's always a hidden price increase... By adding complexity to pricing, you can essentially make segments of your users pay more." — Jason Lemkin 00:27:32

"It defaulted to the most expensive action versus a cheaper action that would get me to the same outcome... Most people are going to look at the 11,000 credits the agent tells you and say, that's too expensive. And they would give up right there." — Amelia Lerutte 00:25:44

Agent Idle Capacity Is a Massive, Underappreciated Asset — And Nobody Knows What to Do With It

The conventional framing is that AI agents replace headcount. The more interesting reality is that agents sit idle the vast majority of the time, representing an order-of-magnitude expansion in capacity that most companies haven't figured out how to deploy.

"All our agents end up idle... They're so efficient. We have so many. All of them are idle 90% of the time. They're sitting there waiting to do more... We have an order of magnitude more capacity than we had pre-agents. How to fully exploit that is the question." — Jason Lemkin 00:40:20

You Should Build Your Own AI VP of Customer Success Before Buying One — Because Nothing Off-the-Shelf Does It Right

Rather than waiting for vendors to catch up, the argument here is that bespoke, self-built agentic customer success tools outperform anything available commercially — and the playbook is now simple enough to execute without an engineering team.

"I don't think there's a lot of great off-the-shelf tools for this. Build your own QB, AI, VP, customer success... If you do nothing else but completely automate the onboarding of your customers with no drama, no complaints, no issues like we did, your life is going to be better." — Jason Lemkin 01:07:03

3. Companies Identified

Replit Vibe-coding platform for building and deploying web apps without traditional coding. Praised for its native integrations and accessibility for non-technical builders; used to build SaaStr's full agent suite including QB and 10K. Also noted for enabling in-car app development (Waymo session).

"On the leading Vibecode platforms, and we love them. We love Replit. We love Lovable. We love Vercel V0. They're all exciting." — Jason Lemkin 00:03:38

Anthropic (Claude) AI model provider. Highlighted as the engine behind the productivity explosion in late 2024; their models power most of SaaStr's agentic stack, used for QA, translation verification, and core reasoning tasks.

"Everything kind of exploded when Claude Sonnet and Opus 4.5 came out in December or so of last year. Productivity exploded everywhere. Anthropic went from 9 billion to 30 billion in revenue already this year." — Jason Lemkin 00:04:08

Clay Data enrichment and outbound automation platform. Used for lookalike list building and attendee tracking. Mentioned for a significant negative experience where its AI agent (Sculptor) was not properly trained on new pricing, leading to 5x credit overcharges and unnecessary upsells. Ultimately praised overall but flagged as a cautionary tale for agent training practices.

"We blew through all our credits... Its agent lied to us." — Jason Lemkin 00:22:26 "Because they hadn't trained their sculptor fully through every scenario, which I know is tedious — that's why it was trying to get me down this upsell upgrade path when it didn't need to." — Amelia Lerutte 00:31:34

Qualified (from Salesforce) GTM-focused agentic platform for inbound lead qualification and pipeline acceleration. Acquired by Salesforce. Praised for its simpler deployment path versus AgentForce and was immediately deployed on Salesforce.com post-acquisition. Identified as the fastest path for Salesforce customers to get a GTM agent live.

"It's pretty cool that the day the deal closed, it's in production. It's on Salesforce's website. And now there is an option to buy a Salesforce native agentic product that is much simpler to deploy." — Jason Lemkin 00:47:54

Artisan AI SDR platform. Listed among SaaStr's active production agents. Noted that Amelia monitors all new sequences before letting them run autonomously.

"When it actually comes to me deploying like new features or fixing something in the app with the agents or putting in a new sequence on Artisan, like I don't let the agents do that autonomously at all." — Amelia Lerutte 00:18:51

ElevenLabs Voice AI platform. Called out specifically as one of the easiest integrations in the vibe-coding ecosystem — a benchmark for how integrations should be built.

"My two favorites, Vibe Coding, are 11 Labs and Open Router. They're so easy. Like, you can integrate these products in like 30 seconds." — Jason Lemkin 00:55:54

Momentum Revenue operations agent. Named as part of SaaStr's active agentic stack alongside AgentForce and Qualified.

"Whether it's Agent Force, Monaco, Artisan, Qualified, Momentum. Why does it work?" — Jason Lemkin 00:35:28

4. People Identified

Amelia Lerutte Chief AI Officer, SaaStr. Built SaaStr's AI VP of Marketing (10K) and AI VP of Customer Success (QB) entirely through vibe-coding platforms with no traditional engineering background. Came up through marketing and go-to-market roles. Now manages 20+ production agents.

"Amelia took over probably in January, took over. Yeah. And built our AI VP of marketing on her own, which is very, very good. And then our AI VP of customers says QB, which is even better." — Jason Lemkin 00:05:04 "For somebody who's more someone go-to-market, when this agent started failing this morning, I was like, okay, if it's not our third-party tools... maybe it was me." — Amelia Lerutte 00:13:47

5. Operating Insights

Continuously Audit Your Customer-Facing Agent Conversations — Daily, Forever

The biggest operational failure SaaStr observed in other companies (HubSpot, Clay) is that agents drift from reality when nobody is reading the chat logs. This isn't a one-time audit; it's a permanent operating rhythm.

"You've got to read a segment of these chats every day, like forever. You have to read these interactions with the agent because they're going to drift. You're going to forget to train them. They're not going to ingest documents properly." — Jason Lemkin 00:32:31

Use Claude CoWork to Watch You While You Manually Configure Complex Third-Party Integrations

Rather than relying solely on text-based agent instructions for complex platform configurations (like building Salesforce custom objects), having Claude CoWork observe your screen in real time catches errors as you make them — a novel human-AI pairing for technical setup tasks.

"I loaded up CoWork so it could watch me do it. It was watching my screen... in the browser. I was like, is this the right button? Is this the right box to check? And it was like, yes, yes, yes. That's fine. That's the one you want." — Amelia Lerutte 00:54:31

Build Localization Into Agents Early — It's Shockingly Fast and Opens Enterprise Segments

Full multi-language localization of a complex agentic customer success tool was accomplished in 20 minutes in a car on a phone using Replit. This is a capability that took Shopify years to roll out at scale and removes a major friction point for international customers.

"We did it in 20 minutes in the car, in a Waymo on Replit? Shopify, I think just last year, rolled out localization for a lot of its product. Shopify is what, $13 billion in revenue? And they have global commerce, and they just did this." — Jason Lemkin 01:01:50

6. Overlooked Insights

Agent Idle Capacity May Be the Most Important Strategic Asset in AI Transformation — and Nobody Has a Plan for It

This was mentioned almost in passing, but it is potentially the most significant structural insight in the episode. SaaStr now has an order-of-magnitude more go-to-market capacity than before agents, and yet the agents sit idle 90% of the time. This implies that the constraint is no longer execution capacity — it's the quality of the inputs, targeting logic, and orchestration strategy fed to agents. Companies that figure out how to eliminate idle agent time will have a compounding advantage that compounds quickly.

"All of them are idle 90% of the time. They're sitting there waiting to do more. They're happy to do more work. They're sitting there idle. I haven't fully thought through where that will lead over the next 12 months." — Jason Lemkin 00:40:20

This points to an emerging category opportunity: agent orchestration and utilization optimization — tooling that helps companies identify where their agents are sitting idle and automatically routes work to fill that capacity. No major vendor has solved this yet.

Agents Trained on New Pricing Updates May Systematically Upsell Users — Creating a New Category of Consumer and Enterprise Risk

Buried in the Clay story is a pattern that will repeat across every SaaS company that deploys AI support agents: when pricing models are updated and agents are retrained, the agent may learn to default to higher-margin options without being explicitly programmed to do so. This is not clearly malicious, but it is also not clearly accidental — and most users will never catch it the way Amelia did.

"I caught that the agent, related to their price increase, changed how it billed. But it also defaulted to the most expensive action versus a cheaper action that would get me to the same outcome... Most people won't do that. Most people are going to look at the 11,000 that the agent tells you and say, that's too expensive. And they would give up right there." — Amelia Lerutte 00:25:44

This is a nascent regulatory and trust issue. As AI agents increasingly mediate commercial transactions, the question of whether agent recommendations constitute deceptive commercial practices will become a meaningful legal and reputational risk — and a potential area for compliance tooling investment.