# When Agents Take Over the System of Record: 40GB Nobody Typed, a Renewal Agent Built in Half a Day, and Why Our Agents Love Clay: The Agents #013 > Source: > Published: 2026-08-28 15:28:16+00:00 Episode #013 of The Agents is here! Same setup as always: three humans, 20+ agents in production, an 8-figure B2B + AI business. This week was about what happens when agents take over the system of record. ServiceTitan just cut off Podium over exactly this question. Our own agents have written about 40GB into Salesforce, 99% of it through the API, and neither of us had logged into Salesforce in a week when it flagged us for overages. And just how safe are “Systems of Records” in the Age of AI? ## #1. Our System of Record Told Us We Had Too Much Data. We Hadn’t Logged In All Week. Salesforce told us we were already running overages just weeks after moving our marketing data over. Amelia’s first reaction was that it couldn’t be right, because we’re barely in there. We’re not the ones writing. The agents are in there constantly, writing as they go: task records, email sends, opens, clicks, call metadata, enrichment. We went from about 5GB to about 40GB in roughly 30 days. Roughly 21 million records. If you’re deploying agents against your CRM, your storage line is going to move dramatically. And the storage costs are often much higher than you might expect. ## #2. ServiceTitan Cut Off Podium After Nine Years and ~1,000 Shared Customers ServiceTitan is the system of record for a huge share of the trades. Podium fed it leads for years, and both sides did well. Then Podium went agentic and crossed nine figures of agent revenue. The agent talked to the customer, scheduled the appointment, tracked the job, and started holding the record. A partner that used to hand off one discrete thing was suddenly doing a lot of what ServiceTitan does. ServiceTitan’s public position was more mature than the fight around it. You can compete with us, you can access records, you can’t use the partnership to displace us. Podium got 30 days’ notice, which is short for a broken integration serving a thousand joint customers. Smaller competitors are still on the platform. Watch for this in CX and support next. If a customer talks to your agent on someone’s website and that agent keeps them, the CRM underneath that interaction may not be needed at all. ## #3. The Same 40GB Is a Lot in Salesforce and a Rounding Error in Postgres We asked Claude what those 21 million records would cost on Postgres. The estimate came back around a thousand times cheaper. That’s an agent’s estimate rather than a quote, but the direction isn’t in dispute. On Neon or Supabase or Databricks, 40GB is nothing. This isn’t Salesforce being a bad actor. They have real costs on a database architected for an era when the customer was a human clicking a UI a few dozen times a day. Nobody priced for agents writing constantly. A year ago people hazed me on X for running the agent stack on Salesforce instead of Postgres. My answer then still holds: we have ten years of hard data in there, it doesn’t drift, it doesn’t hallucinate, and Artisan and Qualified integrate with it natively. If it costs a few nickels more, fine. The question is what happens when it stops being nickels. ## #4. Raising API Prices While Our Data Goes Up 100x Is a Self-Inflicted Wound A lot of vendors are talking about charging more for API access right now, on the logic that customers buy fewer seats and the money has to come from somewhere. I understand why the sales leader in the room wants that. If our data were flat and you raised API pricing 20%, we’d be annoyed and pay it. Our data is up 10x heading to 100x. Raising the per-call price on top of that isn’t a 20% increase, it’s 200% or 2,000%, to reach our own data. What that buys you is our decision to put the next 100GB somewhere else. What customers want is cost tied to outcomes and resolutions, not a meter on reaching their own records. ## #5. Headless Made Salesforce 10x More Useful, and Probably, Also Easier to Leave It took me a while to realize that 10K, our AI VP of Revenue, really is headless Salesforce. It didn’t start that way. Amelia built 10K without a Salesforce integration, we added one to pull a little data, and the agent kept finding more it could do with the connection. Now six things run through it: classic Salesforce CRM for our sales team, Qualified, Momentum for call recordings, Agentforce, Marketing Cloud Next, and a custom quote-to-cash app reaching PandaDoc, Bill.com and QuickBooks. Amelia had never logged into Salesforce in a decade before building 10K. Vendors should sit with the second half of that. The same architecture that makes your system 10x more powerful makes it 10x easier to split. An agent that can sit on top of your API has no trouble sitting on top of three. ## #6. Bizzabo Was a CRM Since 2018 and We Never Treated It Like One 10K pulls from two systems that function as CRMs for us. Salesforce, and Bizzabo. We’ve sold tickets and emailed attendees through Bizzabo since 2018. That is a CRM by any definition. We never thought of it that way because getting data out of it was too much work to bother. Agents made it worth bothering. Go look at whatever system in your stack holds years of customer interaction that you’ve never queried because the export was painful. It’s a CRM now. ## #7. With AI, Every Silver Sponsor Now Gets the Deck Only Diamonds Used to Get Amelia built a renewal agent on top of 10K in about half a day. This is what headless bought us. It pulls the Salesforce side first: who the AE is, what they signed, what they’ve signed historically, LTV, email opens, Qualified conversations, Momentum call history. Then everything that was never in Salesforce: the saastr.com WordPress API, our social accounts, our podcast archive, Bizzabo for real event lead counts. Amelia also let it into her Gmail, since we don’t log every email to Salesforce. Then it builds the deck through the Gamma API, which matters because Gamma preserves our real logo and images instead of generating an approximation of our brand. The Replit deck opened with number one sponsor at the show, 5.9 million social impressions, 76% of leads at VP level or above, and a sample of the 214 articles we’ve written about them. Twenty to thirty have gone out. Before agents we’d have built five, for the diamonds, because no human has a week to make 25 custom decks. Every company cuts that corner and it’s indefensible. The silvers, smallest checks and historically the lowest renewal rate, responded at a higher rate than the diamonds did. ## #8. Agree on the Narrative Before the Agent Builds Anything The first decks were methodical and flat. Here’s what you bought, here’s what you got, thank you. Amelia had to teach it to sell: this is a renewal, put the heavy hitters in the first three slides, make the case for doing more, nobody reads the whole deck. Now, before it generates anything, the agent writes the narrative it intends to tell, and we agree on it first. On one silver sponsor its proposed narrative was “you’re a silver, upgrade to gold.” Reasonable, wrong for that account. They’d come out of stealth right before the event and had grown a lot since, so the right pitch was three options including a media tier with content. Amelia corrected the narrative, not the deck, and then it built. Every agentic sales tool we use writes roughly the same pitch every time because it doesn’t have the context. Approving the narrative takes a minute and gets the context in without you building the deck. ## #9. We Told It Four Times, in Caps, Not to Make Up Numbers. It Still Does. Amelia wrote DON’T MAKE UP NUMBERS into the renewal agent four separate times. It still occasionally invents one. The workflow that tested best: let the AI SDR do the first touch and get the open, then send the custom deck yourself as the reply. Amelia A/B tested giving the agent enough context to send the custom results in the opener. It didn’t work. Second step worked much better, and the replies to the first email gave us feedback to fold into the pitch. Our AI SDR emails are better than 99% of humans, which is all that matters most days. They aren’t great. None of them know how many times we mentioned you on the podcast or how many badges scanned at your booth. Never let them send the follow-up unsupervised. ## #10. Our Agent Picked Clay, Then Picked the Browser Over MCP We needed champion tracking: find contacts who left a renewal account and see where they went. 10K did a poor job and kept steering toward a different tool, because the Replit and Clay integration isn’t good. So Amelia handed it to Claude directly and said use the Clay integration or use Cowork in the browser, your call. It used both. First enrichment passes through the connector, then it decided the browser was faster and finished in Cowork. MCP would have taken a day. The browser took about an hour. Then it suggested running ZoomInfo inside Clay, which neither of us had thought of, and that produced the best enrichment list we’ve gotten. Three layers of agents deep. Amelia saved the sequence as a skill so it stops rediscovering it. Everyone is on MCP right now and a lot of these connectors are slow enough that the agent goes around them. Being the answer when someone asks an AI “who’s the best enrichment vendor” is the easy half. The hard half is being the tool the agent reaches for while it’s doing the work, and being fast enough that it doesn’t route around you. Clay won that on our stack. ## What We’re Building Next: Every Inbound Gets a Custom Proposal, Instantly 10K proposed the inverse of the renewal agent on its own. Run the same machinery on inbound. Today, booking a meeting instantly with Amelia AI beats a contact-us form by a thousand times. Next, every inbound that books a meeting or downloads the media kit gets a custom proposal back: who’s similar to them, who’s had success, what they should do with SaaStr and why. No human in the loop. Amelia tested a proof of concept this week and we’ll show how it was built next episode. None of it would exist without the renewal agent. That was the plumbing. *This is a recap of Episode #013 of The Agents, our weekly show with Amelia Lerutte on running AI agents in production, not in demos. What’s working, what broke, and what to do about it. Come see it live at SaaStr AI Annual 2027, May 11-12 in the SF Bay Area.*