The GTM stage at SaaStr AI 2026 was the one where the abstract finally got concrete. Everyone talks about how AI changes go-to-market. These six sessions actually showed the data, the org charts, the pricing models, and the failures behind it.
What struck me listening to all of them back to back was how much they agreed without coordinating. Stripe, Google, Canva, Cloudflare, Owner, and Higgsfield are wildly different companies, and yet the same patterns kept showing up. Build and sell in parallel. Centralize your AI intelligence. Treat agents as a real customer. Charge for outcomes, not seats.
The lineup:
- Maia Josebachvili, Business Lead and GM of Enterprise Product, Stripe
- Paige Bailey, AI Developer Relations Lead, Google DeepMind, with Scott Barneson, Managing Director of North America Startups, Google Cloud
- Anwar Haneef, GM and Head of Ecosystem, Canva
- Stephanie Cohen, Chief Strategy Officer, Cloudflare (interviewed on stage by Kyle Norton, Chief Revenue Officer, Owner)
- Kyle Norton, Chief Revenue Officer, Owner (separate solo session)
- Alex Mashrabov, CEO and Co-Founder, Higgsfield AI, in conversation with me, Jason Lemkin, Founder and CEO, SaaStr
Here’s a detailed look at each session, with the takeaways that matter most for founders building in B2B + AI right now.
Maia Josebachvili, Business Lead and GM of Enterprise Product, Stripe: The Four Patterns Behind the Fastest-Growing AI Companies
Maia Josebachvili runs work with the fastest-growing AI companies at Stripe as Business Lead and GM of Enterprise Product, which gives her a front-row seat to what’s actually working. Before Stripe she was founder and CEO of Urban Escapes, an adventure company she later sold to LivingSocial, where the original checkout flow told customers to mail a check to her apartment. People did it. Her whole talk was a contrast between how she built then and how she’d build today.
The headline data is staggering. The top AI companies grew 120% in 2025 and 175% in 2026, nearly tripling in a single year. Lovable hit $100M in revenue in eight months and $400M eight months after that. Cursor reached a $1B run rate in under two years, then $2B three months later. Anthropic went from $1B to a $30B run rate in two years. On the consumer side, Stripe’s Link data shows AI adoption doubled from under 6M to 14M buyers, and top buyers now spend $371 on AI, more than the average American spends on internet, streaming, and phone combined.
She organized everything around four patterns.
Speed. As late as 2024, iOS app releases were declining. Once agentic coding tools went mainstream, releases jumped 24% month over month, and new Delaware incorporations followed the same curve. The surprising part: as startup creation grew, the share of technical founders went up seven points in a year. Builders embedding Stripe now get to their first paying customer in under six weeks.
Global by default. A few years ago the fastest-growing companies reached 25 countries in year one and 50 by year three. AI companies in 2025 hit 42 countries in year one and 120 by year three. Gamma reported $100M in its first year with the majority of revenue from outside the US. Across top AI companies, 48% of revenue comes from outside the home market, up from 33% three years ago. Localized pricing drives 18% higher cross-border revenue, and adding even one local payment method lifts conversion 7%.
Pricing. Value is now elastic. The engineer running agents overnight and the parent who replaced Safari with ChatGPT use the same tool and get completely different value. Replit layered credits on top of flat subscriptions and hockey-sticked toward a $1B run rate. Two in three Forbes AI50 companies now have some form of usage-based pricing, up from under 50% last summer, and most run a hybrid model: subscription to anchor the relationship, credits that scale with value.
Adapting GTM motions fast. Building enterprise sales used to be a decade-long journey. Now it happens in year one. Cursor launched self-serve in 2023 and layered in sales-led motion to land enterprise contracts that would have taken others decades. Channel sales have become core, with buyers purchasing through cloud providers. And the new buyer is the agent: agent traffic to Stripe docs 10xed in a single year, and by the end of this year agents will read more Stripe docs than humans.
Key takeaways from Maia Josebachvili’s session:
- Stop building linearly. The old model of build, then sell, then expand internationally is dead. The best teams build, sell, and iterate in parallel, and they go from PLG to enterprise before raising a Series A.
- Treat global revenue as the baseline, not the bonus. Nearly half of every dollar at top AI companies now comes from outside the home market. Localize prices, automate tax collection, add local payment methods, and track conversion country by country.
- Price in your customer’s language. Developers think in tokens, enterprises think in seats but are warming to consumption. Match how your customer experiences value, show them usage before the bill arrives, and sell credits so they think in value rather than dollars per click.
- Build the unified systems agents need now. One customer object, one product catalog, one data model regardless of how someone arrives. Design so an agent can discover, evaluate, and activate your product with no human in the loop.
- Go to market is the tip of the iceberg. Adding a new motion touches your product, pricing, org, and support model. A Frankenstein revenue stack is manageable when you grow slowly and becomes a liability when you compress a decade of GTM evolution into months.
Paige Bailey, AI Developer Relations Lead, Google DeepMind, and Scott Barneson, Managing Director of North America Startups, Google Cloud: Building the Plane While Flying It
Paige Bailey leads engineering for the developer relations team at Google DeepMind. Scott Barneson is Managing Director of North America Startups at Google Cloud. Their fireside was less about product announcements and more about what it feels like to operate inside a company shipping at this pace.
Paige described the internal reality as building the plane while flying it, with new model capabilities landing daily. She used the example of a DeepMind math model solving a Kirby problem, one of the harder feats in topology, the kind of thing that used to take decades. To stay sane in the noise, her team runs always-on agents, including a social listening agent that monitors Twitter, Reddit, and Discord and summarizes releases across the industry. Her advice to founders drowning in announcements: have an agent help you stay on top of the AI releases, not just your product work.
On reimagining work, she walked through how much is already automated at DeepMind. Documentation, tutorials, and blog posts are generated by agents. Every feature push triggers a CI/CD step that generates an explainer video in multiple languages with a custom avatar. Inbox triage drafts responses to routine email so humans just review and send. Scott added a useful frame: they’re in a messy middle where blessed central agents, organically shared bootstrapped agents, and people building their own all coexist, and that experimentation is healthy.
Both flagged the MCP-to-skills shift. MCP servers were the race for a while, then complexity pushed teams toward skills, which are just markdown files that help models make decisions. The next step is automating skill creation itself, so the system notices you doing the same thing three times and builds the tool for you.
On roadmaps, Paige separated two questions: what improvements belong in the model, and what features belong in the product. For models, the only way to move a research team is to give them an eval they can hill-climb against. For features, low-hanging fruit gets fixed by whoever sees it broken, while larger architectural decisions get adjudicated by customer feedback and field signals.
Key takeaways from the Google session:
- Use an agent to keep up with AI itself. The release pace is impossible to track manually. A social listening agent that aggregates and summarizes across platforms is now table stakes for staying current.
- Automate the byproducts of every release. Documentation, multi-language explainer videos, and blog posts can be generated as part of your CI/CD pipeline, which lets a small team ship and explain far more.
- Separate model needs from product needs. Improvements that belong in the model require evals you can hand to a research team. Features that belong in the product come from customer and field signals. Don’t conflate them.
- The messy middle is fine. A mix of central, shared, and personal agents is a sign of healthy experimentation, not chaos. The hard part is focus versus distraction, not standardization.
- Human connection is the moat AI can’t automate. Paige was direct that real conversations and fast incorporation of customer feedback get more important, not less, as the world gets more AI-inundated.
Anwar Haneef, GM and Head of Ecosystem, Canva: The Agentic Playbook From a Quarter-Billion Users
Anwar Haneef is GM and Head of Ecosystem at Canva, and his talk was explicitly a playbook, not theory, drawn from two and a half years of working across agents and AI assistants. He opened with the questions founders keep asking him: will agents replace my product, how do I build integrations agents can find, and how do I become a tool for other agents. Canva asked a sharper version: how do we make sure agents reach for us when they need design.
He framed three shifts. Behavioral: humans browse UIs, agents call APIs, so value locked behind a graphical interface is invisible to agents. Distribution: agents are a new screen, the way mobile was, and you have to treat them as a first-class surface. Value: a human might pick you for a beautiful UI or a friend’s recommendation, but an agent only cares whether you get the job done reliably and on time, which makes domain expertise the real differentiator.
The scale behind this is real. Canva has more than a quarter-billion monthly users, over 27 billion uses of its AI capabilities, and ranks as the third most popular AI platform per a16z. It hosts over 1,700 apps with more than 4,000 developers. Last month it launched Canva AI 2.0, an agentic conversational platform powered by a foundational design model that understands layout, visual hierarchy, and brand. It launched its MCP server last year, became the number one productivity app on ChatGPT, and is seeing 30 to 40% month-over-month growth in MCP connector usage across ChatGPT, Copilot, and Claude. Canva was the first custom GPT back in 2023 and worked with Anthropic on the launch of Claude Design.
The most useful part was his five hard-won lessons, told as failures.
Key takeaways from Canva’s session:
- Treat your MCP setup as a living system. Canva got its architecture wrong the first time. Early structural decisions on shared tooling and hosting created constraints they had to unwind later. Moving fast for first-mover advantage carried a real tax.
- API design hygiene still matters under MCP. MCP is only a surface. A poorly scoped or ambiguous underlying API doesn’t disappear, it just gets harder to debug when agents start failing. Their investment in a clean REST API paid dividends.
- Build for agents from first principles. Agents use your service in surprising, nonlinear ways. When Canva’s search broke during testing, an agent figured out how to export a presentation to PDF, read it, and summarize it. Retrofitting a human experience does not translate.
- Trust at the protocol level is ongoing work, not a one-time task. OAuth, OKTA configs, and credential management are consistent friction across every platform integration. If an agent fails with your service once, it shuts you down and doesn’t come back, so treat trust as a product feature.
- The agentic landscape does not reward generalists. Canva’s bet is to be the undisputed visual design layer. When EY built an agentic sales solution with Snowflake, they didn’t build their own design engine, they called Canva. Ask what your company does that no one else can do as well, and own it.
Stephanie Cohen, Chief Strategy Officer, Cloudflare: The Business Model of the Internet Is Changing
This session was Stephanie Cohen’s, Chief Strategy Officer at Cloudflare, which powers more than 20% of the internet. Kyle Norton, CRO at Owner, came on stage to interview her, but the content and the company here are Cloudflare’s. This was the most consequential session for anyone whose business assumes humans visit their site.
Stephanie’s core claim: in every prior internet era, PC, mobile, social, the business model stayed the same. This is the first one where it changes. Already on the Cloudflare network, more than 50% of HTML page requests are non-human, and at the current rate that hits two-thirds by the end of the year. Publishers got hit first because people read the AI summary, not the source. Financial services and healthcare are next, with AI agents accessing healthcare sites nearly seven times more often than publisher sites in a 30-day window.
Her framework for taking back agency is transparency, context, control, then the tools to optimize and monetize. Know which bots are on your site, understand the difference between training, inference, indexing, and commerce, and control at the network level rather than with robots.txt. Inside the company, an MCP gateway is zero trust for agents. She made a sharp point about content format: Markdown is 80% less data than HTML, and it makes no sense for a machine to read something optimized for human eyes.
On commerce, she revived the 402 “payment required” protocol with Coinbase and Stripe. On any given day there are a billion 402 responses on the Cloudflare network, which she called a billion voices saying please pay me. Web Bot Auth gives cryptographic verification so you don’t pay what you think is the Salesforce bot but isn’t. Partnerships with Visa, Mastercard, and Amex exist to bring millions of merchants along.
The boldest moment: Cloudflare announced 20% of its organization was leaving, not as a cut but to redesign the factory floor. Her logic was that if you keep doing things the old way, you never do them the new way, and your org structure is effectively your technology. If you’re not fit for change, someone newer and faster beats you.
Key takeaways from Stephanie Cohen’s session:
- Assume most of your traffic is about to be machines. More than half of HTML requests on Cloudflare are already non-human, heading toward two-thirds this year. The question shifts from “is that a customer” to “is that a bot, and is it helpful or harmful.”
- Serve agents the format they want. Markdown is 80% lighter than HTML and easier for machines to parse. Agents like FAQs. You may need an agent-optimized version of your site, not just a human one.
- Every non-human interaction is a potential transaction. What feels like friction to a human just works for an agent. The 402 protocol plus identity verification is the emerging plumbing for agent-to-agent commerce, and the demand signal is already a billion requests a day.
- Protect the customer relationship through the agent layer. Protocols like Web Bot Auth and Visa’s Trusted Agent Protocol let the merchant still capture customer information and maintain loyalty even when an agent does the buying.
- Create scarcity to force real change in your own org. Cloudflare deliberately removed the old way of working to make room for AI-native roles. The point isn’t person versus person, it’s whether a given job function still needs to exist.
Kyle Norton, Chief Revenue Officer, Owner: Five GTM AI Decisions Every Company Should Make Now
Kyle Norton, Chief Revenue Officer at Owner, came back for a solo talk, and it was the most operationally dense session of the day. Owner is a vertical AI company serving independent restaurants, think HubSpot plus Shopify for small takeout places. They’re approaching $100M ARR, and Kyle’s been there about four years, starting around $2M. The results he led with: a 20x close-won to OTE ratio, where a $150K rep brings in over $2M in ARR, roughly 4x the AR per rep of direct competitors, and outbound BDRs closing over $100K in ARR each per month. And they’re not selling tokens, they’re selling traditional B2B with a lot of AI inside.
He grounded everything in a sophistication ladder from level zero (reps using ChatGPT as a search engine) to level four (a recursively self-improving system nobody has reached). The gap between companies firmly at level three and everyone else is widening fast, because the gains aren’t 10 to 15%, they’re a doubling of output per person.
Then five decisions.
Centralized versus decentralized. Kyle argued hard for centralized. A decentralized “let a thousand flowers bloom” model builds AI literacy but locks tools away in pockets of early adopters and stalls at level one or two. Worse, it pulls reps into what Stuart Butterfield calls hyperrealistic work-like activities, building a Claude skill that writes mediocre emails instead of being in front of customers. His applied AI team builds things 5 to 10x better than a rep would, and he doesn’t want reps running agents at all. He wants outputs delivered into the surfaces reps already use.
Build versus buy. Buy your infrastructure, build your intelligence. He offered five questions: how critical is uptime, how custom does it need to be, what’s the engineering ROI, is it core intelligence, does it give competitive advantage. A dialer is a clear buy, you don’t rebuild Twilio. Their AI pre-call research was a clear build, and two weeks of one person’s work let 50 BDRs book 85% more.
Where to start. Start with data, both third-party market mapping and well-instrumented first-party customer journey data (they use Momentum to auto-fill Salesforce fields from call transcripts). Then prioritize with a 5P framework: map Possibilities, estimate Payoff, weight by Probability, divide expected value by Perspiration (effort to build and to drive adoption) to get a Priority score. And give your team early wins, because change management is the number one challenge.
The people stack. You need a centralized team of genuine technical experts, owned by someone who deeply cares and can push through budget and behavior change. Jobs are bundles of tasks, so unbundle them: 60% of BDR hours often go to building lists, which is a bad use of a salesperson, so move it to a central data team.
Agentic versus assistive. Beware lossiness. Chain five generative steps with no human in the loop and you get slop. Use deterministic frameworks with humans at decision points, and remember that for GTM, you are the eval. Building the MVP is easy now, the value comes from grinding through iteration until the output is genuinely good. He closed on personal compounding systems, citing Gary Tan’s idea of a personal “Gbrain” and the Karpathy fundamentals videos, and quoted my own line about leading from the front.
Key takeaways from Kyle’s session:
- Centralize the building, decentralize the ideas. Let reps surface what they need, but have a small team of experts bring tools to production grade and deliver outputs into existing surfaces. Reps running their own agents is usually performance theater.
- Buy infrastructure, build intelligence. Run every candidate through uptime, customization, ROI, core-intelligence, and competitive-advantage tests. Boring and undifferentiated gets bought. Proprietary lead scoring and research gets built.
- Start with data or nothing else works. Strong third-party market data plus an instrumented customer journey is the foundation. Slap agents on mediocre data and you get mediocre conversations.
- Unbundle the job before you staff it. Map the tasks inside a role, send what machines do well to machines, and regroup what’s left around what people do well. Roles are already collapsing across onboarding, CS, and account management.
- Manage lossiness and own the eval. Limit how many generative steps you chain before a human checks the output. Building is easy, iterating until it’s actually good is the hard part, and that’s where the gap between winners and everyone else opens up.
Alex Mashrabov, CEO and Co-Founder, Higgsfield AI: The Real Story Behind $300M in 11 Months
I moderated this one as Founder and CEO of SaaStr, and I’m a heavy Higgsfield user, so I peppered Alex Mashrabov, CEO and Co-Founder of Higgsfield AI, with the questions everyone actually wants answered when they hear about these explosive AI numbers. Higgsfield is AI-native video creation for social media, and they got to $300M ARR in about 11 months.
The efficiency is the story. The team is around 160 people, roughly 80 in engineering broadly defined and 70 on the creative side, with about 60 in traditional engineering and product. That’s roughly a 5x ARR-per-engineer ratio against a typical 2x. Alex was clear that vibe coding isn’t yet good for deep infrastructure work like stability, safety, and anti-fraud, which is why engineering still grows as you scale.
The product evolved constantly, which is the part the headline numbers hide. They started with their own open-source-plus model, then made camera controls the first real innovation because creative directors immediately rejected anything without them. Then they leaned into their roots, one-click visual effects (Alex sold a previous face-filter company to Snap), and hit $10M ARR in about five or six weeks. By July they noticed people making commercial projects end to end, reoriented the company around professional creation, and are now reorienting again around agentic marketing workflows with their Supercomputer agent and MCP integrations into Meta and other ad networks.
On the thin-wrapper question, Alex was honest. Higgsfield aggregates every video model (VO, Cling, Seedance, WAN, their own) so customers pick products, not underlying technology. About 40% of usage now goes through higher-level workflows like Cinema Studio and Marketing Studio rather than raw model selection. Average spend is around $1,000 per customer per year, roughly 5x Canva’s $200, and they nearly double ACV every quarter as they move upmarket. The surprise for me: about 70% of that $300M comes from agencies, who found them early because AI made them radically more efficient and gave them something new to sell.
He also gave the most intellectually honest ARR definition I’ve heard. Take monthly and annual subscriptions, divide annual by 12, add the last four weeks of on-demand usage, sum it, multiply by 12. Roughly 40% is annual. His take, and mine, is that the industry overcomplicates this. If Stripe says you’re processing $25M this month, you have a $300M revenue rate, and he expects companies will start sharing Stripe dashboards directly.
Key takeaways from Higgsfield’s session:
- The numbers hide constant reinvention. Higgsfield didn’t go from zero to $300M with one product. Camera controls, then one-click effects, then end-to-end commercial creation, then agentic workflows. Most users today are on a radically different product than the early one.
- Aggregating your “competitors” can be the product. Higgsfield runs VO, Cling, and others in parallel so customers choose the best output per use case. The value is the unified workflow, asset management, and collaboration, not any single model.
- Move up the value stack to expand ACV. Marking up models is the low-margin floor. Workflows that replace an agency or contractor command 5x Canva’s ACV, and customers buy more credits the more value they get.
- Your early adopters may not be who you think. Agencies, not just AI nerds, drove 70% of revenue, because creative agencies were struggling and used AI as a new way to sell to clients while getting radically more efficient.
- Be intellectually honest about ARR. Divide annual contracts by 12, count real on-demand usage, and don’t multiply marketing-discounted or free credits as revenue. If your Stripe dashboard backs the number, that’s good enough.
The GTM Stack Truly is Being Rebuilt for AI
Six companies, one consistent message: the entire go-to-market stack is being rebuilt at once, and the winners are the ones moving on all of it in parallel.
A few threads ran through every session:
#1. The first is that agents are now a real buyer and a real surface, whether it’s Stripe seeing agents about to out-read humans in its docs, Canva treating agents as first-class API consumers, or Cloudflare watching non-human traffic cross 50%. If your value is locked behind a human-optimized interface, you’re becoming invisible to the fastest-growing channel.
#2. The second is that intelligence is the moat, infrastructure is a commodity. Kyle’s build-your-intelligence-buy-your-infrastructure line is the same logic behind Canva’s domain-expertise bet and Higgsfield’s workflow layer on top of commodity models. Nobody wins by being a generalist or a thin pass-through.
#3. The third is that pricing and the business model itself are up for grabs. Usage-based and outcome-based pricing, elastic value, agent-to-agent commerce, and a genuinely honest definition of ARR all point to the same thing: the seat-based subscription assumptions of the last decade no longer describe how value gets created or captured.
#4. And the fourth, said most directly by Paige and echoed by Kyle, is that none of this removes the need to build relationships and lead from the front yourself. The companies pulling away aren’t the ones with the most agents. They’re the ones whose leaders are deep enough in the tools to know what to centralize, what to cut, and what to charge for, while keeping the human connection that AI can’t replicate.
The opportunity is real and the data is clear. The only open question, as Maia Josebachvili put it, is whether you’re driving the change or reacting to it.
