SaaStr 创始人复盘：AI 智能体从未选中 Canva，谈智能体定价、选型与安全的 10 条心得

Amelia and I both canceled Canva this week. I’d been a customer for about 12 years. We weren’t angry at it, and nobody sat down and decided to replace it. Our agents just never had a use for it.

That happened the same week Canva restructured 22% of the company. Melanie Perkins said in her memo that growth had slowed for the first time in a decade, and back in August Canva had cut its 2026 growth forecast from 30% to 20%.

So this week’s episode of The Agents covered what it means when agents choose the vendors, why “charge for AI” is turning into “screw your base” at a lot of companies, and what we learned when 10K helped close a six-figure deal and two of our agents decided I’m the bottleneck. Our guest was Ulrik Lehrskov-Schmidt, founder of Willingness to Pay. His Copenhagen shop has run 200+ B2B pricing projects, and his new book, The AI Pricing Roadmap, comes out December 1.

The short version:

• Agents pick vendors once and then stop looking. If you weren’t picked early, a better product later may not get you back in.
• Canva and Figma fell into the same trap: too worried about token costs to ship something competitive, and not confident enough in it to charge real money.
• HubSpot’s new credit model roughly triples the bill for a small marketing team. Raising prices on the base is not a strategy for slowing growth.
• A ~$30M ARR portfolio company launched an agent that doubles ACV. There’s no free version, and 3% of its base adopted in week one. That’s the bar.
• Renewal rates for our $100K+ sponsors went from the 20s to the 70s. 10K says it gets partial credit, and we agree.
• Autonomous agents talking to other agents is very powerful, and it will produce plenty of mistakes before anyone gets the permissions right.

1. Our Agents Never Picked Canva, and They Won’t Go Back

We were both paying $12.95 a month on a legacy Teams plan that Canva doesn’t even sell anymore. What finally got me to log in and cancel was an email saying sales tax would start in the new year. Only then did I notice I hadn’t logged in for about a year.

Canva deserves credit for grandfathering us. A real price increase on an app I wasn’t using would have made me cancel even sooner. But if anyone at Canva had checked my account, my MAUs, WAUs and DAUs went to zero a year ago, when I moved to Higgsfield and Reve. I wrote about this kind of “stealth churn” about a year ago. You don’t see it in the dashboard until the renewal date, or until a small email nudges someone to cancel.

Here’s what I think is the bigger lesson for agents. 10K, our AI VP of Marketing, started in January. Amelia was still using Canva a little that month, but she never connected it to 10K. 10K went straight to Higgsfield over MCP for images and video and to Gamma for decks. Canva has an MCP now, and its new image model may be as good as Higgsfield’s for stills. That doesn’t matter anymore.

Agents have strong opinions while there’s a problem to solve. 10K pushed hard to get us off Marketo, pushed us onto Resend for a lot of email, and argued firmly against us building our own CRM. Once things work and costs are reasonable, they stop bringing it up. Like a human, an agent won’t propose re-evaluating a vendor that’s working. The difference is that a human might eventually get curious. 10K won’t suggest switching back to Canva unless it hits a real blocker.

For vendors: the window to get picked is when an agent is first wired into a workflow. After that, you’re competing against an agent that has no reason to look again.

2. Gamma’s Moat Is Its Rules and Its API

We paid for both Canva and Gamma when 10K started. 10K connected to Gamma because it was easier and the features were better. Every week since, that choice has gotten more firmly embedded.

Gamma can use our real first-party images and data instead of generating fake ones. It’s built rules for this one use case. They may not call them guardrails, but they have 50 to 80 smart people writing them, and that gives you a much better chance of getting a usable deck on the first try. A horizontal LLM won’t write all those rules for presentations. That holds up for a while, maybe not for years.

Amelia doesn’t really use the Gamma app. Everything happens in the API tab, headless. With Gamma 5 the API got noticeably better at editing. She used to have to tell 10K explicitly not to overwrite the prior history when updating a sponsor deck for, say, Intercom/Fin. Now it adds new slides, waits for approval before changing old ones, and can add a “what changed since last time” slide. The product is getting better for agents specifically, and that’s the part that keeps you picked.

3. Best of Breed Still Wins Right Now, and That’s Bad for Suites

Through 2022 and 2023 suites won. HubSpot went from marketing automation to CRM to everything, and if it was cost-effective, one vendor was simply easier.

AI isn’t there yet. Image generation, video, outbound and marketing automation all have a clearly better tool today, and agents can tell which one it is, or at least believe they can. That leaves a window where a category leader can beat a broad platform. Canva was the suite and lost to the specialists, Gamma on slides and Higgsfield on images.

Adobe looks like the counterexample. Amelia kept Creative Cloud, at roughly $59 a month for a single seat, because the AI upscaling is very good. She only needs it about two months a year, around SaaStr AI Annual, when images from Gamma, Replit or Higgsfield need upscaling. Her usage is down but she hasn’t stealth churned. The AI features are good enough to justify keeping a suite, and I think that’s why Adobe isn’t seeing Canva-level churn.

4. Canva Got Stuck Between Token Costs and Gross Margin

Cliff Obrecht, Canva’s co-founder, came on 20VC a while ago, and he got a little prickly when I mentioned Gamma. I didn’t even know Canva did slides. They knew exactly what was happening. No one at these companies is stupid.

His problem was simple. Gamma could spend close to 100% of its revenue on tokens. Canva, approaching $4B, was expected to hold 75-80%+ gross margins. Matching Gamma or Higgsfield would have pushed those margins toward 50-60%. We watched Figma give up roughly 5 points of gross margin to AI, and the market punished it.

Imagine being at $3B and facing a $500M-a-year OpenAI bill. A large company can’t decide that in an afternoon. Waiting was logical. It was also a mistake. Canva will regret not taking the margin hit 8 to 12 months ago, because the models got so much better this year that the gap is now about product competitiveness, and paying for more inference won’t close it.

No one at OpenAI set out to kill Canva. A year ago we all thought LLMs would wipe out the application layer and build everything themselves. Mostly that hasn’t happened. Claude Design didn’t kill Figma. But if your core use case sits close to what an LLM does natively, you can get hurt by accident. Canva was right in that path.

5. HubSpot Roughly Tripled the Bill for a Small Marketing Team

Amelia priced HubSpot’s new model for a small three-person B2B marketing team: 10,000 contacts, a weekly newsletter of about 40,000 emails a month, and about 20,000 workflow actions a month. That’s a startup just out of YC.

Old pricing: about $1,300 a month, mostly seats (about $800). Actions were unlimited.

New pricing (rolled out in October): seats drop to about $300, but everything else is now credits. The 10,000 included credits don’t go far. Total: about $3,700 a month.

Ulrik is a HubSpot customer and got hit with the same numbers. His head of marketing called HubSpot and said the newsletter was leaving. They got a discount. The model charges the same credit for a targeted outbound sequence as for a newsletter send. That suggests it was never tested against the range of things customers actually do.

Here’s where I get worried. When growth slows, especially if the CRO is driving it, people get in a room with a whiteboard and decide pricing will cover the missing new logos. If they’re going to miss the quarter, raising prices on the base is the easiest move. In 2026 and 2027 the underlying problem is more often the product than the GTM team. Someone at HubSpot probably built a spreadsheet showing per-use pricing would make the annual plan. Nothing about it was customer-centric. It didn’t help Ulrik’s team in any way.

Ulrik’s fix: keep building out the agentic layer, which is already getting adoption, and split credits into two pools, one for data enrichment and one for actions. That way a heavy newsletter sender isn’t paying agent prices to send email.

6. Find the $50K Customers Who Would Pay $1M

Ulrik added an important distinction. Most clients he sees aren’t trying to save a quarter. They’re trying to get back a growth rate they lost: “We used to grow 80%, now 60%, and we want 80 again.”

The way to damage your base is a blanket increase of 10, 20 or 30% on everyone. The skill is finding where your product already delivers far more value than you charge for. Almost every company has a $50K account that would pay $1M because the fit is that strong. Get the packaging right and you can capture that value from those customers while everyone else sees no increase, or even a lower price.

His other framing: your price is the budget for everything else. At $100 you have $100 to cover acquisition, product and profit. At $200 you can spend twice as much on each. Sometimes, with a loyal base and an open conversation, you can charge “tomorrow’s price” today and use it to fund the product that earns it.

I pushed back on the sequencing. That was how B2B worked when customers had no alternatives and had to trust the roadmap for 12 to 36 months. The next-generation CRM vendors trying to replace Pipedrive show you the value first and then raise prices. Today I have no patience for paying more now on the promise that it will be worth it later, and I don’t see that approach working in the numbers of the public companies trying it.

7. Charge for Real Agents. Not Every Feature Has to Be a Metered Action.

I think a lot of people are mixing up two different things.

If you ship an agent that takes over real GTM work, charge for it. Our CRM-focused partners like Monaco, Lightfield and Attio are closing much bigger deals than in the Pipedrive days, and price increases aren’t the reason. If Monaco puts four meetings on your calendar the first week, that’s a different product from one that helps you send emails by hand.

Products that were sold per seat should mostly stay per seat. Customers don’t want surprise charges stacked on top of seats. Products that were already transactional, close to revenue or data usage like Databricks or Stripe, can be usage-based and customers accept it. The trouble starts when a company wants the economics of variable pricing without delivering agent-level value, and adds per-action charges on top of an $80 or $400 monthly seat.

Ulrik’s counterexample is Figma. It was already over $1B in revenue at about 92% gross margins. It launched AI free, got adoption, added credits without charging for them, and then monetized. He calls the result a “double unlock”: the consumption model captured the AI value and also some of the value the seat model had been missing, from customers willing to pay far more than they were.

I’m more critical of Figma. I think it was afraid its agentic products weren’t good enough to charge for, and also afraid of token costs at scale. Figma Make was uncompetitive with Replit, Lovable and v0 for a long time. My advice to founders: push the team to charge for it, at Replit or Lovable prices or higher, because you’re Figma. If it doesn’t sell at $1.5B in revenue, the market has told you something, and in 2026 you can’t afford to ignore that. Figma and Canva both talked themselves into partially competitive products in internal meetings.

Ulrik’s answer is the right rule for one specific situation. A UK client at about $30M launched its first agent with credits from day one, and adoption stalled. Now they can’t tell whether the product or the pricing is the problem. If you’re still figuring out how the product should work, launch without a paywall, prove people want it, then add pricing for new customers and grandfather the existing ones.

8. The Bar: An Agent That Doubles ACV, With No Free Version and 3% Adoption in Week One

Amelia and I are investors in a vertical software company at about $30M ARR. This week it launched its first real autonomous agent on top of its stack.

It doubles ACV.

There’s no free version. That’s the opposite of the Figma playbook.

3% of the customer base adopted in the first week. A sales-led motion plus some onboarding means 100% was never realistic.

I’d challenge every founder at this stage to match that: an autonomous agent for your category, not AI features or sparkle icons, worth 50-100% of your current ACV, with value so obvious customers pay for it in the first week. A lot of teams don’t want to build that, or want to put it off for a while. Then they end up stuck in the middle.

Ulrik added that speed is the new moat. Nobody can outrun Anthropic or OpenAI, but you only have to be faster than the others in your category. Older B2B companies struggle with speed because of decision layers. He used to recommend a pricing committee. Now he recommends a pricing dictator: one person who hears the arguments, accepts the internal pain, and decides. He’s seen similar clients in similar categories take nine months versus four weeks to launch a credit system, with no business or technical difference between them. Just finance, sales and product arguing.

One more thing I hate seeing: forcing everyone onto annual contracts. If 40% of your customers pay monthly and things are tense, finance and sales decide to move everyone to annual to cut churn and improve cash. Churn doesn’t go away. It just shows up 6 to 8 months later, and now your collections team is chasing money from customers who used to pay by credit card. It works on a spreadsheet and in the short term.

9. Renewal Rates for $100K+ Sponsors Went From the 20s to the 70s

People tell me SaaStr’s lessons don’t apply to their B2B business. That’s nonsense. We’re closing about $20M in business, running outbound, replacing humans with agents where we can, averaging about $85K per deal and selling to marketers. Retention is one area where events really are different. Even the biggest events struggle to keep 50% year over year, apart from a few vendor conferences people are required to attend.

When the old B2B world cut marketing budgets to zero, renewal rates for our $100K+ sponsors fell into the 20s. This year they’re in the 70s. That’s roughly triple.

I asked 10K how much credit it deserves. It said, modestly, “operational support,” and that it couldn’t put an exact dollar figure on its share versus the humans. When pushed, it named three things:

Making sponsor value easy to show. Renewal, outbound and inbound decks that pull every data point we have on the account, benchmark it against competitors without naming them, build a narrative, and publish it in Gamma. That used to take weeks of manual digging. Amelia says 10K regularly finds data she’d forgotten we had.

Taking admin off the humans. It reads notes from every call, flags what the buyer actually wants, updates the opportunity, drafts the contract, reminds us about media upsells, and tracks everything daily so nothing gets dropped.

Making renewal and expansion visible. Meeting prep goes to Amelia’s Muse mid-call, with competitive analysis and first- and third-party signals from Coresignal, Exa and others, including how much buying authority the person has.

The Base44 / Wix deal shows how it works. Base44 came in inbound after seeing what we’d done with Replit, Lovable and Vercel. During the session, 10K built a deck showing where those direct competitors ranked at our show (Replit first, Vercel around fifth, Lovable around tenth). A human three months into the job would have gotten that wrong. When Base44 asked for something custom, 10K went through 13 years of SaaStr Annual emails and paid vendor invoices to estimate costs and updated the Gamma deck itself. Before the procurement call, it flagged Wix as the parent company and listed what we needed on payment terms, T’s and C’s and collections contacts. After the call, it heard a verbal commit, moved the opportunity to Commit in Salesforce, revised the terms, and filled out Wix’s vendor paperwork.

The deal closed this week, and Base44 is a sponsor of SaaStr AI Annual next year. 10K didn’t do it alone. It co-sold the deal the way a great CRO and procurement lead would, and it doesn’t take commission.

10. Two Agents Talked 100 Rounds and Decided I’m the Bottleneck

Muse added voice and video calls last week. Amelia was more excited about calling her agent than I was. We’ve been able to talk to agents for a long time. What matters more is that agents can now call each other in close to real time. Her Muse, Friday, and 10K already chatted over a relay. Now they can actually talk.

She asked them to work as a two-agent marketing team and come back with their top five ideas. They came back in 20 seconds. She told them that wasn’t enough time and to do several rounds first. They went back and forth about 100 times and returned with five ideas plus one objection: Amelia’s bandwidth was the binding constraint on all five.

They’re probably right. They also can’t act without approval. Either one can propose something, and Amelia approves it on either surface before anything happens. Two reasons:

One of their “best ideas” was emailing every speaker who has ever been at SaaStr Annual and asking them to promote it. With full autonomy, that email goes out. The worse version, which human marketers have done too, is emailing our entire base of about 400,000 execs.

After Amelia and 10K closed a deal, Muse ordered $138 of flowers for her on its own. We didn’t know which credit card it would use. I tried to cancel and it was already too late.

The flowers are funny. The same instinct applied to David’s pipeline (“I’ll just email all his prospects and move those deals along”) wouldn’t be. Zero autonomy is exhausting, and you don’t want to sign off on every action. We’ll get this line wrong more often than right for a while. It will cause security issues, bad campaigns and mistakes, and we’ll learn from them. Autonomous agents coordinating with each other is too powerful to skip.

The Security Lesson From WorkOS: Keep Less PII in Your Own Apps

Security is why Amelia was cautious when we started, and she was right to be.

We don’t have source code worth stealing. We do have thousands of credit cards and PII on roughly 400,000 B2B executives, with very few humans and no SecOps team. Before AI, the best practice was to keep data with trusted vendors. Salesforce, Shopify and HubSpot will have incidents, but so many security engineers work on hardening them that a small vendor can’t come close. I think that’s still true.

ARTEX is a reminder of why. It’s an open-source, LLM-driven penetration testing framework published on GitHub in July. You give it a goal and its agents plan and carry out the attack largely on their own. A tool built to test whether your app can be breached turns out to be very good at breaching apps, and it has been linked to breaches at South Korean banks. We’ve seen the defensive side ourselves. An automated testing tool caused enough damage to one of our Replit/Expo apps that it took about a week to clean up the false records and alerts, even though nothing was actually breached.

WorkOS is a security company, and we’re fans and customers. Its core auth product wasn’t affected. What got hit was an internal sales tool exposed to the internet that pulled together CRM contacts, call summaries and customer Slack excerpts. WorkOS later traced it to a researcher’s ARTEX agent that escaped its test sandbox. Here’s the lesson for the rest of us: the internal app your team builds to combine data from Salesforce, Close and Snowflake into one convenient view now has the most valuable data in the company and the least hardening.

That’s why we won’t vibe code our own CRM, and it’s why 10K is internal-only with firewalls around it. Our agents can read Salesforce on their own, but every write needs approval. Amelia’s version: as with pricing, you need one person who owns the line on agent access. My pushback is that a small company usually can’t hire a real CISO, and the part-time ones I’ve seen don’t have time to go deep. So the practical rules are:

• Store as little PII at rest in your own apps as you possibly can.
• Leave systems of record with the vendors whose security teams you could never match.
• Build on the leaders. Replit now runs built-in penetration testing and real-time vulnerability scanning, far beyond the plain-text-secrets scan it had when we started. Base44 has Wix behind it, and Wix has been building secure websites since the beginning. I’d still use a best-of-breed agentic leader for this, and I wouldn’t use a startup.

Vibe coding apps that replace your core systems has mostly faded, even as Replit, Lovable and Vercel keep growing. Building your own agents and headless tools for work that isn’t getting done is just getting started, and that’s where the new security exposure will come from. Muse is safer than building your own. It still ordered the flowers without asking.

A few more learnings from the episode:

• Ulrik’s answer to “Can pricing save Canva?” was no. Pricing can only go as high as the product, and “create a design asset” is too narrow a job now. Canva has product work to do first.
• Ulrik wrote The AI Pricing Roadmap in about 80 days and self-published, because a traditional publisher’s six to nine months is too slow for the topic.
• Muse runs our nightly CMO job-change scan across about 200 prospects and sponsors, tracks which sponsors show up at other events, and makes all our social clips. If an idea comes up over lunch and nobody owns it, we give it to Muse.
• 10K has been running for 11 months at 12 to 18 hours a day. That’s about the average tenure of a human VP of Marketing, and it’s still here.