The main stage at SaaStr AI 2026 in San Mateo ran end to end on agents.
Agents in production, carrying quota, writing to systems of record, and reshaping how companies are built. Below is a session-by-session recap of the day, with every speaker and their title, followed by the themes that showed up again and again no matter who was talking.
The CPO Panel: The Hardest Job in B2B Right Now
Jason Lemkin opened by arguing that chief product officer went from the cushiest job in B2B to the most stressful one in about 18 months. For a decade the job was bringing your mug to the office and deciding what to push to next quarter. Now every CPO is under the gun to ship agentic features that customers will actually pay for.
The panel:
- Emrecan Dogan, Head of Product, Glean
- Anneka Gupta, Chief Product Officer, Rubrik
- Rachel Wolan, Chief Product Officer, Webflow
- Anique Drumright, Chief Product Officer, Harvey
Rubrik’s Anneka Gupta walked through Ruby, Rubrik’s agent for operating the platform. It started as a simple RAG application on top of documentation and is now agentic, taking on tasks like forward-looking capacity planning that used to take a customer a full day. The constraint in cyber recovery is that nothing can take the core service down, so Rubrik uses LLMs to generate the plans but keeps the recovery plans themselves deterministic and explainable.
Webflow’s Rachel Wolan described Webflow as an agentic web marketing platform that started as a website builder. The new launch is an answer engine optimization agent that detects when you need to upgrade technical AEO, looks at what competitors are doing, drafts content, and runs autonomously with a human approving changes. Webflow has roughly 18% of the top 2,000 websites and saw customers get a 75% bump in organic traffic from technical AEO automation. Her framing: answer engines now drive about 50% of visitors, up from 10% a year ago, so a brand that does not show up in answer engines is invisible to buyers.
Glean’s Emrecan Dogan made the case that retrieving information was the original job, but the world is moving from getting informed to getting AI to perform. Glean now operates two ways: as the AI assistant front door, and as a single MCP server that feeds superior company context into tools like Claude Code, Cursor, and Codex. The interesting wrinkle he flagged is the spillover effect: the more usage happens in external harnesses, the more it pulls users back into Glean’s own surfaces.
Harvey’s Anique Drumright described an entire platform that is agentic, not a single agent feature. Litigators are now asking how to leverage agentic capabilities, how to coach associates on using AI safely, and how to verify an agent’s plan, not just its output. Harvey deploys legal engineers, most of whom were practicing lawyers for eight to ten years, alongside forward deployed engineers. Jason’s read: selling AI into law firms is harder than selling security, because partners often pay out of their own pocket, which makes Harvey’s scale even more impressive.
The recurring point across the demos was responsibility. As agents take more core actions, the vendor stays accountable even when customers do things with the agents nobody anticipated. Jason and Rachel both landed on the same idea: the new race is agent-led growth and agent experience, where the default choice of an agent matters more than ranking in a chatbot answer.
Top 3 takeaways
- Use the LLM to generate the plan, but keep the action deterministic. Rubrik’s split, probabilistic planning with deterministic and explainable execution, is the pattern for any agent touching production systems where a wrong action loses the customer.
- The buying decision is moving from humans to agents. Webflow optimizing for answer engines and the panel’s agent-led growth point are the same shift: if your product is not the default an agent recommends, you are invisible regardless of your AEO or your sales team.
- You are accountable for what customers do with your agent, well beyond whether your core product works. Build the guardrails before you ship the autonomy, because you will not predict every workflow customers run.
Jason Lemkin’s Keynote: Tired vs Wired
Jason’s solo keynote was a state of the union. Anthropic was reportedly projecting a $50 billion revenue run rate, while public software leaders were down as much as 70% on the year. He framed the gap as tired versus wired.
The “you’ll vibe code your own CRM” fear from last year is dead. You can vibe code a lot, including most of the apps running the event, but you cannot vibe code something good enough to replace a real platform. The goal is not a cheap clone of Salesforce from 2006. The goal is a CRM that puts deals on your calendar, which is worth $50,000 to $100,000 even to a small company.
The bigger frame: old SaaS may be dying, but B2B and AI is in a Cambrian explosion. Harvey raised at $9 billion. Replit, founded in 2016, is approaching a billion in revenue. Lovable did not exist two years ago. The catch is competition. Most categories now have two to a hundred times more competitors.
On the public markets, Jason laid out the SaaSpocalypse: nearly zero B2B IPOs from late 2021 through 2025, most 2025 IPOs were duds, private equity stopped buying, and public software traded at a discount to the S&P 500 for the first time ever. His counter: the three biggest IPOs in tech history are coming, naming SpaceX, Anthropic, and OpenAI, with Databricks added in. Cursor is being acquired for around $60 billion, and he predicted a $100 billion startup acquisition within 12 months.
He used SaaStr itself as the proof point: 20-plus humans in 2024 doing less than three humans and roughly 20 agents do now. He walked through the agent stack, including 10K (the AI VP of Marketing) and QB (the AI VP of Customer Success), and made the point that the cost is about $257 a month against roughly $500,000 of replaced labor, while productivity went up. He showed QB proactively servicing 120-plus sponsors, the kind of work human CS reps at SaaStr had skipped entirely.
His action items for founders: build the number one agent in your space, be open and agent-friendly (Salesforce gets praised, Workday gets flagged as struggling), and make your API agent-friendly even if you do not have an agent yet. He pointed to the SaaStr.ai API Report Card, where Stripe earned the only A+, as the way to grade yourself. He closed on Gartner data showing overall software spend reaccelerating from about 12.8% to 15% growth, with Palantir going from 27% growth to 85%, and reps from Twilio, Atlassian, Datadog, Cloudflare, and Shopify cited as companies tapping AI budget.
His four-category framework, borrowed from a segmentation Rory did with Harry Stebbings on 20VC:
- New AI native, no legacy: everything is up.
- Pre-AI but AI is driving deals and new customers up, like Twilio, WorkOS, and RevenueCat: the best place to be.
- AI is expansion only, like Atlassian: better than most.
- Not getting AI budget: the recovery is not coming.
Top 3 takeaways
- Stop selling incremental software, start selling instant ROI. A $50,000 AI SDR is an easy yes when it books revenue tomorrow; the same $50,000 for a workflow tool is a hard sell. Price and pitch on the outcome, which is why agents can command two to ten times software pricing.
- Be honest about which of the four categories you are in. The dangerous spot is category four, slowed-but-not-zero growth running last year’s playbook while waiting for a reacceleration that will not arrive on its own. Performative AI in board decks is not the same as products customers will buy.
- Agent-friendliness is a near-term, low-effort moat. Making your API agent-ready can be a couple days of work and decides whether agents and the builders behind them pick you. Grade yourself against the API Report Card and fix anything below a B+.
Salesforce and PayPal: Turning SMB Into a Fast-Growing Segment
Amelia Lerutte hosted a session on how Salesforce uses AI to make SMB one of its fastest-growing segments.
The speakers:
- Adam Alfano, EVP of Global SMB and Emerging Products, Salesforce
- Eitan Saban, Sales Leader, PayPal
Salesforce’s Adam Alfano said Agentforce has closed thousands of deals across the SMB and startup segment, with examples spanning a heavy equipment company (Equiptor) running an outbound agent delivering emails about twice as effectively, a healthcare company (Mindt Health) doing patient onboarding and surgery prep, and AI companies like Windsurf and 11 Labs using Momentum for revenue orchestration. He works almost entirely in Slackbot rather than logging into Salesforce, managing a team of thousands of SMB account executives in a headless environment.
PayPal’s Eitan Saban runs an organization of about 200 reps. PayPal onboards north of 100,000 merchants monthly and faced roughly 8,000 leads a month with no human power to handle them. After deploying an SDR agent fully in production 14 weeks ago, meeting conversion went up about 50%, with the agent running ten nudges religiously and humans showing up only for the meeting that matters. He emphasized that in a regulated environment you cannot launch an agent without marketing, legal, and compliance bought in, and that agent maturity is a real curve: the agent at week one against 200 leads is a different agent than the one running against 80,000 leads a month.
On messy data, the consensus was do not wait to solve world hunger on data before deploying. Adam’s point: you can set up a web agent that understands your site, FAQs, and product data and start accumulating better intel while you fix the rest. Eitan added that messy data is permanent and also relevant to humans, who jump between apps to find answers, so an agent that figures it out yourself produces a more consistent outcome.
Amelia demonstrated headless Salesforce: Jason does not have a Salesforce seat, but uses it constantly through the API, with 10K pulling sponsorship pipeline, media pipeline, the full deal list, and historical sponsorship data, then executing campaigns directly. The closing exchange, with Eitan’s line that the real question is not whether AI replaces people but which people AI makes irreplaceable, set up the action-bias theme that carried through the day.
Top 3 takeaways
- Point agents at the leads no human was ever going to work. PayPal’s 8,000 monthly leads and SaaStr’s thousand-person follow-up backlog were not stolen from reps; they were revenue nobody had the capacity to touch. That is the cleanest place to start because there is no human displacement debate.
- Do not wait for clean data to deploy. A web agent can run on your site, FAQs, and product data while you fix the rest, and the act of using agents tells you which data actually needs cleaning. Action bias beats a perfect data project.
- Treat agent onboarding like hiring. PayPal’s agent at week one against 200 leads was a different agent than the one running 80,000, and no seller closes on day one either. Budget for a maturity curve, with humans curating and tuning, instead of expecting set-and-forget.
Databricks Fireside: What Enterprises Are Really Doing
Jason interviewed Databricks co-founder and SVP of Field Engineering Arsalan Tavakoli. Databricks is over $5 billion in revenue, reaccelerating at 50 to 60% growth, with deep penetration across enterprises, and Arsalan argued no one has better visibility into what enterprises actually do versus what they say on Twitter.
The reality check: on X everyone looks like they have figured out AI. In the rest of the world adoption is nascent. Every CEO is telling employees that if they are not using AI they are behind, which has created a token-maxing problem where spend goes up but ROI is unclear. He laid out five phases of AI adoption, with almost everyone sitting between phase one and three, very few reaching real process redesign or net-new capabilities.
The product story centered on context. Databricks built Genie Enterprise Context, which looks at the questions people ask and the data they pull, then builds and evolves company context on the fly rather than letting a static document go stale. He used a 7-Eleven example for store-level inventory decisions and a car manufacturer that added 70,000 users who can now self-serve questions instead of waiting a week for a data analyst. His blunt line on dashboards: he calls them dashboard graveyards, and the interface is moving to text and voice while visuals remain the right way for the brain to interpret data.
On the SaaS apocalypse, Arsalan argued any business with a monopoly today will not have one in 12 to 24 months, because the cost of building software has collapsed and the cost of migration has dropped. Databricks runs LLM-driven migrations (analyze the environment, convert and modernize the code, migrate the data, reconcile outputs) that used to be multi-year and can now happen in roughly 30 days. His moat philosophy: lock customers in with value, not proprietary APIs. Jason’s counter, which Arsalan agreed with, is that there is now a soft moat in organizational context. When all of 7-Eleven understands a structure, leaving is hard even when it is technically easy. And the final twist: what the agent wants increasingly decides the vendor.
Top 3 takeaways
- Token spend going up is not the same as ROI going up. The token-maxing trap is real: employees told to use AI or be fired default to the most expensive model for everything. The work is in routing, caching, and matching the model to the task, not in maximizing usage.
- Cheap migration kills incumbency. When LLMs cut a multi-year migration to roughly 30 days, the willingness to switch vendors jumps and monopolies erode from the bottom. If you are the incumbent, your pricing power is on a clock; if you are the challenger, the door is open.
- Context, not raw data, is what makes agents useful, and it goes stale fast. A static glossary written two years ago is already wrong. The durable version learns from the questions people ask and evolves itself, which is also where organizational lock-in quietly forms.
Lovable: No More Feature Moats
Amelia Lerutte interviewed Elena Verna, Head of Growth at Lovable, on Lovable’s one-year anniversary and a launch day. Lovable shipped native SEO and AI discoverability for every app it builds, from server-side rendering and pre-rendering to a Semrush integration for keyword research.
Elena’s background is traditional tech (Dropbox, Miro, Amplitude, SurveyMonkey), where her Dropbox team was a couple hundred people. At Lovable she went back to being an IC and called it the next flex of careers: not climbing to VP titles, but becoming a high-powered IC who can do what used to take dozens of people.
Her biggest take-away was on moats. Feature differentiation used to be the moat in the SaaS era because development was expensive. With AI writing 80%-plus of code in AI-native organizations, any feature can be copied in months. The moats that still work: hardware, network effects, data, security and compliance, and brand. On brand, Lovable leans hard into building in public. She shared internal mechanics: a flat structure with no titles, a #shipped Slack channel where work goes to production multiple times a day, and a principle that if you can convince one other person an idea is good, you ship it. Lovable crossed $400 million in ARR with just under 200 people, and engineering (which they call product engineering) is the top hiring priority because, even near half a billion in revenue, they consider themselves still on the product-market-fit treadmill.
On freemium, she argued it has never been more important. Lovable treats freemium cost as marketing budget and announced a LinkedIn partnership giving LinkedIn premium members free access, with double-digit conversion to paid. Her advice for non-AI-native marketing teams: let the engineering team lead the change in marketing structure, because otherwise marketing becomes the blocker to the velocity AI-native development unlocks.
Top 3 takeaways
- Feature moats now last months, not years. When anyone can vibe code a clone, differentiation has to come from data, network effects, security and compliance, or brand. Plan your moat around those, and treat any feature lead as temporary.
- Org structure is a velocity lever. A flat team, a #shipped channel, and a convince-one-person rule are why Lovable ships multiple times a day at under 200 people and $400M ARR. If marketing or approvals cannot match engineering’s pace, you become the bottleneck.
- Freemium is marketing spend in the AI era. Giving away expensive features is how you change buyer habits; the LinkedIn deal converting at double digits is the proof. Stop gating your highest-cost features if your goal is adoption.
Snowflake CMO: Deploying Agents at Scale With Compliance
Amelia Lerutte interviewed Denise Persson, Chief Marketing Officer of Snowflake, on what it takes to deploy agents inside a large, compliance-heavy organization. Snowflake marketing is about 700 people, owns the new business pipeline, and is aiming to be the most AI-assisted marketing team in B2B.
The contrast with Lovable was the point. Where Lovable ships to production multiple times a day with no hierarchy, Snowflake runs a centralized AI and data organization that certifies skills before they are used broadly, because the company has very low risk tolerance. Their GTM agent, Raven, is what the entire sales and marketing organization uses, with skills certified centrally to avoid duplicated effort.
Denise no longer logs into a dashboard in the morning. She interrogates the data directly through Snowflake Intelligence, getting a morning brief on pipeline, the health of the organization, who joined and left marketing, and even travel and expense flags. On results, she cited a 30% reduction in cost per opportunity over six months from optimizing media spend daily instead of at the end of a campaign.
On data, her line was that bad data plus AI means bad decisions faster and at scale, so the data estate has to come first and cannot be siloed in marketing. Snowflake’s chief data officer is now chief data and AI officer. On the organization, attrition fears from a year ago have flipped to people leaning in, supported by weekly AI skills training, a weekly AI challenge where someone shares what they built, quarterly AI days, function-level hackathons, an AI council, and a usage leaderboard she is careful to frame as being about outcomes, not token count. On hiring, certifications matter less and soft skills (adaptability, curiosity, change management, collaboration) matter more, with GTM engineer the function they are actively hiring for, often converting marketing ops and business analysts into the role. Events demand is going through the roof, and product marketers finally have time to spend with customers.
Top 3 takeaways
- Scale changes the rollout, not the thesis. Snowflake runs the same agent playbook as Lovable through a centralized layer that certifies skills before broad use. If you are compliance-heavy, the governance layer is the price of moving fast, not a reason to move slow.
- Adoption is pulled, not pushed. Snowflake’s fluency came from peer learning, hackathons, and an outcomes-based leaderboard, not mandates. The surprise power users came from unexpected teams, so make curiosity easy rather than forcing usage.
- Hire for adaptability over tool certifications. The skills that age are specific platforms; the skills that compound are change management and curiosity. GTM engineer is the role to staff now, and your existing marketing ops and analysts are the most natural conversions.
Monaco: What Everyone Gets Wrong About AI Sales
Sam Blond, co-founder and CEO of Monaco, closed with lessons from seeing 200 AI sales startups as a VC at Founders Fund before building his own. He was previously a go-to-market and sales leader, including as CRO at Brex.
His diagnosis of the startups he saw: most were either a feature pretending to be a company (taking call content and updating a CRM field should be a feature of the call recorder or CRM, not a standalone business) or a solution in search of a problem (building something nobody asked for and waving it off with the faster-horse line). His three company-building lessons: build in a space you understand better than anyone, have deep prior experience in the category (citing Brex’s and Ramp’s founders), and respect market timing.
On applying AI in go-to-market, he split the work in two. AI is now better than humans at building and scoring your TAM, enriching and overlaying signals (funding events, new hires, website visits), executing multi-channel outbound, and pipeline maintenance like opportunity creation and field updates. If you are still doing those by hand, you are behind. AI is worse than humans at meeting with customers and building relationships, and at creative or operationally complex campaigns. Buyers of complex enterprise software want to talk to a person.
His takeaway: do not outsource your whole go-to-market to AI and assume a technical founder can just focus on code. Use the leverage AI provides to spend more time on the highest-ROI work, which is customer calls and standout brand campaigns. He walked through Monaco’s cadence of a major initiative every month, mixing organic moments (product launch, Series A in February, Series B announcement, the SaaStr sponsorship) with manufactured ones (the Monaco Invitational poker tournament with a $100,000 prize, San Francisco billboards, a plane trailing a banner over the event). The process recommendation: pick five creative people, send a prompt 48 hours ahead, get in a room, whiteboard five ideas, pick two or three, and the result is the most crowded booth at the event instead of one more iPad giveaway.
Top 3 takeaways
- Test every idea against feature-versus-company and solution-versus-problem. If your product should be a feature of the CRM or the call recorder, it is not a company. If nobody asked for it, the faster-horse defense usually means you skipped the customer research.
- AI does not let founders skip go-to-market, it relocates their time. Hand AI the TAM building, signal enrichment, outbound, and pipeline hygiene, then reinvest that time in customer calls and creative campaigns, the two things AI is still worse at and that actually move the needle.
- Manufacture brand moments on a monthly cadence. You will not always have an organic milestone, so create one, and let them compound. A $15,000 plane against a $100,000 sponsorship is reasonable ROI when each campaign makes the others work harder.
The Common Threads
Different companies, different scales, different risk tolerances. The same patterns kept surfacing.
- Agents carry quota now, not just tasks. PayPal’s SDR agent lifted meeting conversion 50% in 14 weeks. Salesforce’s Agentforce closed thousands of deals. QB serviced 120-plus SaaStr sponsors. Across the day, agents were treated as teammates with outcomes attached, not features in a product tour.
- Data is the foundation, and bad data compounds. Snowflake, Databricks, Salesforce, and Lovable all said it independently. You do not wait for perfect data, but the data estate has to come first and it cannot live in a silo, because AI amplifies whatever you feed it, including the mistakes.
- Headless and agent-friendly wins the vendor selection. SaaStr runs Salesforce headless. Glean serves context as an MCP server. The Databricks close and the SaaStr API Report Card pointed the same direction: when the agent picks the tool, the most agent-friendly product gets chosen. Make your API agent-ready even before you have an agent.
- The dashboard is fading into answers. Denise stopped logging into dashboards. Arsalan calls them graveyards. The interface is becoming conversation, and the value sits in the structured data and context underneath, not in over-applying the most expensive model to every query.
- Context and brand are the durable moats. Feature differentiation lasts months when anyone can vibe code a clone. What holds is data, network effects, security and compliance, organizational context, and brand. Lovable builds in public, Monaco manufactures monthly brand moments, and Snowflake invests in authenticity as the human contribution.
- Human time moves up the value chain. AI takes the TAM building, the prospecting, the field updates, the pipeline hygiene. People move to customer relationships and creative, operationally complex campaigns. Eitan from PayPal’s framing summed up the whole day: the question is not whether AI replaces people, but which people AI makes irreplaceable.
The market is bifurcating, and it is going to feel more bifurcated every month. The companies tapping AI budget are not just growing faster, they are exploding. The rest are waiting for a recovery that is not coming. The good news from every stage: it is still early, and in almost no category is it too late.
