News SaaStr runs 21 AI agents in production, revenue swings 66 points YoY
## SaaStr Runs 21 AI Agents, Learns What Breaks in Production
Jason Lemkin's SaaStr is running 21+ AI agents across sales, marketing, and finance operations. Revenue moved from -19% to +47% year-over-year. Episode 007 of *The Agents* podcast covers what broke, what worked, and what that means if you are deploying AI agents in your own go-to-market motion.
### Over-Guardrailing Kills Performance
SaaStr runs a free VC pitch deck grader at saastr.ai. It has processed 4,600+ decks. During their recent SaaStr AI Annual event (10,000 attendees), the agent started failing 88 of 305 submissions and handing out F grades to 53% of completed decks.
The cause: 14 accumulated guardrails. Each rule was added to fix a specific error (pulling projections instead of current ARR, misreading TAM numbers). By the 14th exception, every deck triggered multiple conflicts. Ambiguity defaulted to "no data," which stored zeros, which collapsed all scores to F.
The fix: scrap most rules and rebuild. The lesson: guardrails past a threshold become technical debt that breaks the product.
### Same Spec, Different Platform, Different Agent
SaaStr rebuilt their AI VP of Marketing (called "10K") on Lovable using the same spec that ran on Replit for months. Same data sources, same APIs, same instructions.
The agents behave differently. The Replit version returns three ideas and structures them like a B2C performance marketer (motivation, channel, audience, success criteria). The Lovable version ("10K Prime") returns four ideas and is more aggressive: it recommended paid LinkedIn ads targeting GTM leaders and a flash sale, which the Replit version never suggested.
Two platforms, one spec, two personalities. The platform shapes agent behavior as much as your instructions.
### AI VP of Finance Built Inside AI VP of Marketing
SaaStr planned a standalone AI VP of Finance for collections and cash visibility. Instead, they built it inside their existing AI VP of Marketing (10K).
Why it worked: 10K already had access to Salesforce (closed deals), Stripe (daily ticket sales), YoY data, projections, and historical financials. A clean finance agent would have started blind. The merged agent had full context from day one.
You would not make your VP of Marketing your VP of Finance in real life. In agent land, context matters more than org chart logic.
### Agent Negotiated a Vendor Renewal
One SaaStr agent took over a software renewal negotiation. The vendor did not love it. Lemkin did not elaborate on the outcome, but the fact that they let it run is the story: AI agents are moving from lead gen and SDR automation into account management and renewals.
This is the shift. AI SDRs and AI AEs were the first wave. AI agents for customer success, renewals, and retention automation are the next. If your CS team is not experimenting with this, someone else's is.
### What This Means for ANZ Sales Teams
AI sales agents are moving past the hype phase. SaaStr is not a vendor selling AI SDR software. They are a media company running agents in production and reporting what breaks. That makes this data more useful than most vendor case studies.
If you are deploying AI agents for outbound, lead generation, or account management, the lessons are:
- Guardrails accumulate silently and can strangle performance
- Platform choice affects agent behavior more than you expect
- Context beats clean separation: merge agents if it gets them better data
- AI agents are already handling renewals, not just top-of-funnel
The revenue swing (from -19% to +47% YoY) is not broken out by agent contribution vs. other factors, but the fact that they are running 21+ agents in production while most companies are still piloting one is the real signal.
AI sales automation is no longer speculative. It is in production, breaking things, and getting fixed. If your team is not testing this, you are behind.