The takes I started on the pod and didn’t get to finish.
We had more news this week than almost any week since Harry, Rory and I started doing this. That meant a lot of threads got cut off, by Harry moving us on, by Rory pushing back, or by me running out of air. These are the ones I’d have kept going on.
1. Anthropic’s 6,000 Employees Matter More Than Its $1.4 Trillion Valuation
At $100B in revenue, Google had 81,000 employees. Meta had 71,000. Anthropic will be there with under 6,000. I said on the pod that for a moment this is the most important thing in the whole IPO, and Rory pushed back: old-school companies won’t end up with $5 to $7 of inference for every $1 of labor, so it won’t be as grim for jobs as I think.
He’s right that a consulting firm won’t run like Anthropic. My view is it doesn’t have to. It just has to compete against someone who runs closer to it. The average AI-native go-to-market team is about 30% smaller. Anthropic is 92% smaller. The other side of that is pay. When the average employee is making $41 million on the IPO and a manager earns $400K, a huge pie is being split across very few people. Every AI company will try to copy that.
My learning → When you build your 2027 plan, start from the smallest team that could hit the number, and justify every hire above that. Don’t start from the org chart your pre-AI competitors built.
2. 2027 Could Lose Its Two Biggest Buyers at Once
I started on this and Harry told me not to be negative. I’m not. But it’s worth finishing. For two years, a lot of AI deals had an unspoken backstop. Nvidia or one of its friends would buy the team, fund the next round or bail out the infrastructure play. Reflection, Poolside, Hugging Face. Harry has spent 12 months at IC telling us we can’t lose on these deals because of our liquidation preference. I was joking about that on the pod, but not really.
With Jensen gone, I don’t think a CFO promoted to CEO two days ago is buying all these kids for $15B or $17B. Now add Salesforce, which did a lot of deals this year and may spend 2027 digesting them. If Nvidia isn’t doing deals and Salesforce isn’t doing deals, 2027 looks very different from the “you can’t lose with your liq pref” year of 2026. Poolside said plainly it sold because it couldn’t raise the $10B to $20B gap. What was plan B? And will Thinking Machines survive? I don’t know.
My learning → A liquidation preference is only worth what someone will pay for the company. If your exit math assumed Nvidia or Salesforce as the buyer, redo it without them.
3. I’m Not Sure Intercom Proves Founder Returns Work
Every VC points to Eoghan and Intercom as proof that bringing the founder back works. I was a tiny investor from the early days, and I said on the pod I’m not sure it proves that. It definitely worked. He fixed the endless price increases in his first week and went all-in on agents before almost anyone. But even with all of that, it got to about $100M of agentic revenue, and then it became part of Salesforce. We never got to see the next act.
That matters for HubSpot, Workday and UiPath. Founder mode will help all of them. Founders can make decisions others can’t. But Rory’s point holds: no B2B company from the pre-AI era has reaccelerated above 30%. The question for Brian and Dharmesh is whether it’s enough, not whether it works.
My learning → Judge a founder return by one number: net new customers a year later. For HubSpot, that means getting back above 5,000 a quarter, and ideally toward the old 10,000.
4. Salesforce Stopped Disclosing Its Customer Count Four Years Ago
I mentioned this in passing and it deserved more. HubSpot still reports customers, which is how we know net new adds went from 10,000 a quarter to guiding below 5,000. Salesforce, on its way to roughly $50B in revenue, stopped disclosing its customer count about four years ago. Whenever a company stops disclosing a metric, it’s telling you something one way or the other.
Salesforce is a master at selling you five to ten expensive products. It’s not as good at bringing in new logos. That model works until the expansion runs out. You can mask 20% growth for years with decent NRR and steady price increases, and people who don’t look closely won’t know.
My learning → When a public company drops a metric from its reporting, go find the last three quarters it did report. And never drop net new logos from your own board deck, especially the quarter it turns down.
5. SMBs Are Skipping HubSpot and Monday and Buying AI Receptionists Instead
This is the one I didn’t anticipate. We grew up in a world where you couldn’t run a business without CRM and work management software. A new VP of Sales would roll out Salesforce in week one. It turns out SMBs can survive without it, and right now they’re choosing to.
It isn’t substitution. I looked at Ramp and others, and Attio picking up another $10M doesn’t explain what happened to HubSpot and Monday. SMBs are just putting it off. A little data analysis in ChatGPT. Google Sheets for another six months. They may buy HubSpot or Monday in a year or two. That may also end up being never.
What they are buying is agents that replace a person. Three of my portfolio companies sell front-office and back-office agents to SMBs at $500 to $1,000 a month, and if they get a prospect on the phone, they close that day. The SMB’s P0 isn’t a better Kanban board. It’s that the receptionist quits every Wednesday. An agent that answers 24/7, reschedules every appointment and handles billing for $800 a month gets signed that afternoon. The budget didn’t go away. It moved from software that helps a person to software that does the person’s job.
My learning → If you sell to SMBs, rebuild your pitch around a specific job a person does today, priced below what that person costs. If your product can’t be described that way, plan for your net new customers to keep falling.
6. You’ll Replace Marketing Automation Without Noticing
Picture how this plays out in B2B. You get the HubSpot quote. It’s $30,000. Then you realize Muse can do more of it than you thought, so you push the purchase out a month. Ask Muse to go into your CRM, find the top 200 founders, and invite them to a VIP dinner tomorrow night. It does it. We’ve already done exactly that.
That’s what companies pay $100,000 a year in marketing automation software to do, except with humans picking the list and getting it wrong. Nobody is going to announce they replaced their marketing automation vendor. They’ll just stop renewing, without thinking of it as a replacement at all.
My learning → Before your next big software renewal, ask Muse or another horizontal agent to do the core job for two weeks. If it gets you 70% of the way there, renegotiate or cut the contract.
7. Muse Is Overengineered for What People Think It Does
The press covers Muse as a consumer app for making reservations. It’s massively overengineered for that, in a good way. The team built what I think is the ultimate expression of OpenClaw. It’s the best software I’ve used since Claude 3.x got good. The memory, the browser agent, and the way the model is tuned to finish tasks instead of arguing with you or being developer-focused all add up.
Rory joked that I’m not the typical user, and the business model will survive me. I said maybe I’m just the future in the present. He said that’s the caption. I meant it as something to learn from, not hubris. I’d still worry about Instinct, which Harry likes, on one thing: runaway infrastructure costs. Meta can give away 24/7 agents because it has a cash cow behind it. Rory’s answer that great founders will figure it out is not a cost model.
My learning → Find your five heaviest users and look at how many workflows they run. If it’s 20 to 30 on one agent, that’s where your average user will be in two or three years, and your pricing and infrastructure costs need to work at that level.
8. OpenAI Buying Adessia Means Vertical Agents Get Absorbed Next Year
I got cut off on this one. I think we’re past the era of agentic apps and AI CX. Next year is the year of horizontal, autonomous agents, and that’s where most people at a lay level will feel AI. Claude Code may be the greatest app of all time, but most humans aren’t Claude coding all day.
That’s why I think OpenAI buying Adessia, a travel agent, matters, and why we’ll see 20 more deals like it. The vertical assistants become features inside the horizontal agents. Mark was right to push Muse now. In January, with everyone launching, it would be much harder to get 20 million people running workflows on it.
My learning → If you’re building a vertical agent, check whether Muse or ChatGPT can already do your core use case today. If it can, your most likely exit is an Adessia-style acquisition, and you should start those conversations in 2027, not 2028.
9. We’ll Stop Trying to Get Humans to Work Like Agents
Harry said his biggest challenge is workforce mindset: getting people who’ve worked the same way for 10 years to ask whether an agent could do it first. I think that’s a transition phase, and it inverts. Right now we’re trying to get humans to work with agents. Soon we’ll just ask the agents first and plug humans into the gaps.
When you ask Muse first, you just won’t make the next hire. Our agents do all our collections now, and the contractor who billed $3,000 to $4,000 a month now bills about an hour. You don’t even fire people. They think they found a better job, they leave, and you don’t backfill.
My learning → Before your next backfill, give the job to an agent for two weeks. Hire only for what it can’t do.
10. The One-Person Billion-Dollar Startup and S-Tier Security Can’t Both Be True
I said on the pod that I find this confusing, and I still do. On one hand we want one-person billion-dollar startups with, frankly, no security. Have you built a real production app yourself? On the other hand, I don’t think you can trust anything right now except companies with S-tier security teams. Optimizely is a moldy oldie of B2B, and it got compromised across CNN, the BBC and Zoom. Fireworks has a best-in-class team, and it still got named by Notion, Cursor and Harvey.
And it’s not just credentials. I think the Microsoft outage is DDoS for the AI era. It was a load bug that sat there for four months until load hit a new peak. Anyone can now spin up 10,000 agents and see what they can take down. If you’ve ever run a B2B service through cascading server failures, you know there’s nothing more stressful. Ten servers go down, you shift the load, ten more go down, and it gets worse even as you find the fix.
My learning → For now I’d be very careful adding any new third-party plugin, analytics tool or payments startup beyond Stripe or Adyen. Every vendor with access to your site or your keys is now part of your attack surface.
The Common Thread: Nobody Announces the Switch
Most of these come back to the same pattern. Software built for humans is getting replaced without anyone saying so: SMBs who put off buying HubSpot and buy an AI receptionist instead, agents that do the marketing automation job once and keep doing it, teams that stop backfilling. None of it shows up as a customer switching to a competitor. It shows up as a missing new logo, a renewal that doesn’t happen, or a role nobody posts.
Jason’s Takes is the SaaStr AI companion to our weekly 20VC x SaaStr recap with Harry Stebbings and Rory O’Driscoll.
