We’re three humans and 20+ AI agents now. More agents check into our Slack every day, including Saturdays and Sundays, than humans do. Amelia (our Chief AI Officer) and I recorded Episode #02 of The Agents this week to walk through everything that happened with our agent stack: the wins, the defeats, and the ones that caught us off guard.
Here are the 10 biggest takeaways.
1. Your Agents Will Get Lazy. You Have to Check Them Every Single Day.
The agent running our SaaStr AI Annual agenda quietly decided that 50 sessions was enough, even after we added 20+ new speakers in the last week. Amelia’s own session (teaching attendees how to build an AI VP of Marketing live) got dropped from the top 10.
Then the agent blamed the API integration. When Amelia pushed back and asked it to check its own records, it admitted: “I don’t have a clear audit trail showing which specific change removed it. I should have just said that to you instead of constructing a theory.”
Agents are goal-seeking. They go just as far as they think they need to resolve an issue. Sometimes that’s fine. Sometimes it means your top session gets deleted and the agent fabricates a reason why.
Takeaway: Set-and-forget is dead. Every output that feels a little off, a little dated, check it. Maybe not in 2027. Today, you check it every day.
2. 60% Solutions Don’t Get Paid For Anymore
HubSpot launched an AEO (Answer Engine Optimization) tool. It gave SaaStr a zero on content quality for AI with zero recommendations on how to fix it. For context: we get 800,000+ readers on the blog and 5,000-6,000 from ChatGPT alone. We’re not a zero.
So I vibe coded a better version in Replit in 10 minutes with a single prompt. Better output, free, no paywall after 10 queries.
This is the danger zone for every established B2B + AI company right now. Teams spent 10 months shipping what would have been great in August 2025. Today, 60% solutions ship and die. They might get used. They won’t get paid for. And they won’t get renewed.
The list of 60% solutions we’ve tested recently: Figma Make (vs. Replit/Lovable/v0). Reve (vs. Canva for some use cases). Every “first all-in-one marketing agent” that’s launched in the last 60 days. Pick any category where the top tool is already 9/10 and ask yourself honestly: is what we’re shipping 9/10, or is it 2025-grade?
3. Figma Make Is the Worst Vibe Coded Product We’ve Used All Year
My context test for every new agentic product: “Redesign SaaStr.ai and make it better.” This prompt works beautifully in Replit, Lovable, and v0 because Claude 4.6 is that good under the hood.
Figma Make got a zero. It hallucinated the entire website like it was early 2024.
Classic Figma has its own problem: every year, sponsors try to build their booth graphics in Figma, and every year the files break when they go to print. This year three sponsors ignored our warnings. All their proofs came back corrupted. Illustrator is still the gold standard for print, and it now has better agentic capabilities than Figma. I can literally tell Illustrator “move the text” and it works.
Figma has become grandpa software. NRR is still high. Top decile growth. But who wants to buy the new stuff from grandpa software?
4. Stealth Churn Is Your Real Canary in the Coal Mine
I realized I hadn’t logged into Canva in over 100 days. Still paying $18/month. I use Reve for thumbnails, Higgsfield for video, and Claude directly for charts that integrate data. I still love Canva. I’m grateful for Canva. I just don’t use it anymore.
Amelia hasn’t logged into ChatGPT since December 27. Still paying for the team plan.
Pre-AI, I used to laugh when investor updates had DAU/WAU/MAU metrics in B2B. This isn’t consumer, guys. Now it’s one of the most important metrics you track. Usage falling across your base is not a lagging indicator. It’s the earliest leading indicator of churn you have.
If you don’t measure stealth churn, you don’t know you’re already losing.
5. Your Claude Chat History Is Quietly Becoming a Moat
Every single piece of SaaStr content I’ve worked on in the last year is a Claude chat. Claude can search through all my prior chats. That’s a form of stealth memory nobody talks about.
When I’m writing a new post and want to pull a point from the Agents #01 episode, Claude finds it. Every transcript, every post, every artifact is in there. Moving that context to ChatGPT or another LLM would be theoretically possible but practically impossible.
This is how B2B + AI products are quietly building lock-in right now. If your agent can reference a customer’s prior work without manual re-uploading, they’re not leaving. Even the cognitive load of maintaining two LLM subscriptions drives consolidation.
6. Limiting FDEs to 5,000+ Employee Customers Is the New CS Math Mistake
One vendor we love just told us they’re capping forward deployed engineers at 5,000+ employee customers. Everyone below goes self-serve. Self-serve agent deployment fails for predictable reasons: the list is wrong, the messaging is generic, the signals say “this doesn’t work,” and the buyer writes off the whole category.
FDEs are not a cost center. This is the same myopic math that got customer success cut to zero a generation ago and repurposed into a sales function. If you onboard a $20K AI customer that costs $20K in FDE resources to deploy, and they stay five years and become a reference that brings in three more, that LTV is hundreds of thousands of dollars.
Benioff said it best on 20VC last summer: “I wish I could give every customer an S-tier FDE before I charged them a dollar.” Hypergrowth forgives a lot of mistakes. The zombie deployments coming out of self-serve will not forgive you later.
7. FDEs Don’t Need to Be Engineers. They Need to Know the Product Cold.
My Salesforce FDE started as a sales engineer. Now she’s on the FDE team and she’s excellent. When the code gets deeply technical, she brings in a truer FDE engineer. That’s fine. There are levels.
What matters is that the person knows the product cold. Every CS rep who failed at agent deployment failed because they didn’t know the product. Every FDE-light who succeeds (whether they came from sales, CS, product, or PM) knows it cold.
The whole point of the FDE is getting your customer deployed with your agent in production. However you solve that, with whatever title, the answer is not self-serve.
8. Vector Won by Removing All Friction
We declared an agent freeze in March. Hundreds of vendors pitched us through it. Vector broke through. The CEO emailed directly, got on a 15-minute call, upgraded Amelia’s account himself, walked her through deployment, and it worked. Now Vector is running our targeted ads and de-anonymizing website traffic into our AISDR flow automatically.
Compare that to our experience with Clerk, where the CTO publicly argued on Twitter that it was my fault for not understanding their app. I’ve founded a B2B company that went from zero to $200M+. I’ve vibe coded 10+ apps to production with 750K+ uses. I guess I’m the problem. Meanwhile WorkOS’s CEO reached out on Slack and offered to have his team migrate us over the same day.
The winning formula in 2026 is Benioff’s playbook at startup scale: give every customer the deployment experience you used to only give your top 10% of enterprise deals. Do that and you get references for life.
9. Every API Is Now a Public API. Treat It That Way.
Amelia is incredibly capable, but she would never have touched an API a year ago. Today she’s building custom Salesforce objects with Cowork and integrating Bizabo session data into Replit-hosted dashboards. Vibe coding has made every API accessible to every buyer.
The agentic API test: go into Replit, Lovable, or v0 and prompt “build me a dashboard that integrates with [your product].” If it doesn’t work in 15 minutes, you’re losing customers who don’t even know they’re about to leave.
Bizabo is an events platform that hasn’t been meaningfully updated in years. Their UI is still 2018-grade. I was ready to churn off them in January. But their API is just extensible enough that we were able to build a custom SaaStr AI Annual agenda on top of it. That API saved them from churn. Your API might be the only thing doing that for you too.
10. Marketo Is Our Worst API Experience. Salesforce Is Our Best.
Marketo costs more than $50K/year. This week their unsubscribe broke. We recreated the bug ourselves: clicked unsubscribe, still received the next newsletter. Users who were deleted under privacy laws still received newsletters. When we reported it, their support blamed our third-party integrations (Salesforce, Replit, Zapier, all irrelevant to an unsubscribe flow). 48+ hours later: crickets. Meanwhile we’re accidentally violating CAN-SPAM because their core feature is broken.
Our AI VP of Marketing (10K) can’t even access our own marketing data inside Marketo because the API isn’t agent-friendly. So 10K pulls from Salesforce instead, which has quietly become our source of truth. Native Agent Force integration. Clean API. All our data already lives there.
The best APIs we’re working with right now: 11 Labs (one line of code to integrate voice), OpenRouter, Resend, and Salesforce. Build the most agent-friendly API in your category and you can ship crusty features for years. Build the worst one and you lose customers who’ve been with you for a decade.
The Meta Theme: Every Workflow Is Being Rewritten in Real Time
Two other things worth flagging from this week that didn’t make the top 10:
Cowork + QB together imported 150 sponsors into our networking app in 10 minutes. This used to take a human four hours of mind-numbing work per event, or a full day of passive-aggressive labor, or we’d just make sponsors do it themselves. Now Cowork researches everyone’s social links, fills in missing descriptions to match the style of the ones we have, and pushes everything to the mobile app platform. More complete data than we used to collect. 10 minutes.
AI VP of Finance is the next agent we’re building. $8M in non-credit card revenue flows through invoicing every year at SaaStr. Collections is manual at SaaStr and at almost every B2B company we talk to. Our accounts receivable has aged more this year than ever, despite having more and better sponsors. QB is adding a finance layer: auto-generate invoices, send to AP contacts, follow up, confirm remittance. V1 doesn’t even need to touch Stripe, QuickBooks, Bill.com, or Brex. It just needs to own collections. That alone saves a human 20+ hours a week at our scale, and probably 2-3 FTE at most growth-stage B2B + AI companies.
SaaStr AI Annual 2026 is three weeks away from load-in, about a month from the event itself. Come build an AI VP of Marketing with us live. Bring your laptop.
SaaStr AI Annual 2026 is May 12-14 in the SF Bay Area. All the leaders from Salesforce, Replit, Vercel, Cloudflare, Databricks, Atlassian, Lovable, v0, and more will be there. Register at saastr.ai.
