The Seat Is Empty

For twenty years, the software subscription business ran on one arithmetic. A company buys Salesforce. Each salesperson gets a seat. The company hires more salespeople, buys more seats, pays more money. Salesforce grows. Everyone is happy. The SaaS industry built $3 trillion in market capitalization on that one equation: headcount goes up, revenue goes up.

An AI agent does not need a seat. It logs in, does the work of three salespeople, and logs out. The company does not buy three seats. It buys one agent subscription, or it pays per task completed, or it just uses the agent that came bundled with the software it already owns. The headcount goes flat. The seat count goes down. The old equation stops working.

This is why Salesforce stock is down 26% this year while its committed future revenue is $72.4 billion and growing 14%. The revenue that already exists is fine. The market is pricing what happens to the next dollar of revenue — the one that used to come from the next new hire buying the next new seat.

---

The Brief

• The seat-based pricing model is the specific thing under pressure — and the incumbents know it. Salesforce introduced "Agentforce Work Units" — charging per AI task completed rather than per user login. Workday launched "Flex Credits." ServiceNow moved to consumption and value-based pricing for its agent products. HubSpot converted to a credits model early. The companies that moved first are doing better: HubSpot grew 20% year-over-year through the worst stretch of the software selloff, while more traditional-priced competitors saw multiple compression. The transition from "pay for access" to "pay for work done" is the real restructuring happening under the surface of these stock prices. ✓ Fortune · Feb 10, 2026 / Bain Technology Report 2025

• This week, Anthropic launched Claude Design — a direct entry into territory software companies have owned for 25 years. The tool generates complete interactive prototypes, website drafts, and pitch decks from a text prompt. No designer required. Anthropic's CPO Mike Krieger had resigned from Figma's board three days earlier. Figma — the professional design platform used by 95% of Fortune 500 companies — fell 7% on launch day. Figma crossed $1 billion in revenue in FY2025, growing 41%, with net dollar retention of 136%. The stock is down 85% from its August peak. A company with extraordinary fundamentals, being repriced because the pool of people who need to buy a design seat just got smaller. ✓ VentureBeat · Apr 17 / Figma FY2025 earnings

• But the $2 trillion selloff is too blunt. Gartner estimates 35% of single-function SaaS tools will be replaced by AI agents by 2030. That implies 65% are in a different position. Healthcare record systems like Epic, pharmaceutical compliance platforms like Veeva, financial data terminals like Bloomberg — these hold legally required data with compliance logic embedded in the data model. A hospital cannot replace Epic with a chatbot because Epic is the legally mandated audit trail. AI agents use Epic. They do not replace it. The market has sold everything equally. The businesses are not equally threatened. ✓ Menlo Ventures · Apr 2026 / SaaS Capital · Mar 2026

• The SaaS companies with the deepest data moats are actually becoming more valuable, not less. When AI agents need to act on business data — book a meeting, update a customer record, process a claim — they need to connect to the system that holds that data. That system is usually a SaaS platform. The more deeply a SaaS product is embedded in a company's data infrastructure, the more it becomes infrastructure for the agents. Salesforce and ServiceNow are racing to become the operating system that agents run on. Bain's framing: the winners own the "agent operating system," not just the application. ✓ Bain Technology Report 2025 / Deloitte Tech 2026

• The real victims are single-function tools that do one thing a general-purpose AI can now do for free. Survey software. Document formatting tools. Basic dashboard generators. Meeting schedulers. These cost $20–50 per month per user and do one specific thing. When Claude or ChatGPT can do the same thing from a prompt — and the company already pays for Claude anyway — the $30/month subscription loses its justification. IT departments are already quietly killing these. The enterprise average of 170+ SaaS applications per company is heading down. That is where the 35% replacement estimate lives. ✓ Gartner via Register · Dec 2025 / BetterCloud · Jan 2026

---

The Arithmetic That Broke

The best way to understand what is happening is through a company that runs 100 Salesforce seats.

Each seat costs $150 per month. That is $180,000 a year. The company has 100 salespeople. The math is simple. When the company hires 10 more, it buys 10 more seats and pays $18,000 more per year. Salesforce loves this. The revenue is recurring, predictable, and tied directly to the customer's own growth. As the economy grows, companies hire, and Salesforce's revenue grows automatically.

Now introduce an AI agent. The agent handles initial customer outreach, qualifies leads, schedules meetings, and drafts follow-up emails. Three salespeople's routine work, done by one agent. The company does not fire those three people — they move to higher-value work. But the company also does not hire three new people for those tasks. Headcount growth slows. The number of seats the company will ever need has a lower ceiling than before.

Multiply this across Salesforce's 150,000-plus enterprise customers and the math on future seat growth starts to look different. The existing $72.4 billion in committed revenue is locked in. The question is what comes after that. The market has decided the answer is: less than it used to be. Hence the stock price.

Salesforce's answer is Agentforce — its own AI agent platform. Marc Benioff's argument: customers who use agents get 10x the value, so Salesforce can charge 3x to 4x more per customer even if they have fewer seats. The Agentforce Work Unit replaces seat counting with task counting. That is a coherent strategy. The open question is whether customers will actually pay 3x more for software they were already buying — or whether they negotiate the seat count down and resist the premium price on the new model simultaneously. HubSpot's early data suggests the transition can work. It has not been tested at Salesforce's scale.

---

Why Design Software Got Hit Hardest

Figma is worth examining closely because it shows what happens when the seat shrinkage hits a very specific type of user: the professional who has a dedicated tool that a newcomer can now approximate without training.

Figma's real-time collaborative design platform has 136% net dollar retention and is used by 95% of Fortune 500 companies. Those existing customers are not going anywhere. Professional designers who live in Figma every day are not switching to a chat prompt for their production work.

The shrinkage hits the next layer: the product manager who needed a designer and a Figma seat to produce a prototype, but can now get a working prototype from Claude Design in two prompts. The startup founder who used to hire a freelance designer, who used Figma, and now uses an AI tool that costs $20 per month bundled with their Claude subscription. Every one of those paths was, eventually, a new Figma seat. Now it is not.

On April 17, Anthropic launched Claude Design. Figma's stock fell 7% in one day. The same week, Anthropic's CPO — who had been on Figma's board — resigned from that board to launch the competing product. Figma's full-year 2025 revenue was $1.06 billion, up 41%. Its stock is down 85% from its peak. The market is not saying the business today is bad. It is saying the total number of Figma seats that will ever exist is lower than it thought a year ago.

"The expansion of the design user base to non-designers is the real competitive threat, even if the professional designer's workflow remains anchored in Figma for now."

— VentureBeat, analysis of the Claude Design launch · April 17, 2026 ✓ VentureBeat · Apr 17

---

Where the Threat Does Not Reach

The software that is hardest to threaten shares one characteristic: it is the legally mandated place where certain data must live.

Epic Systems runs the electronic health records for most major US hospitals. Every patient interaction, every prescription, every billing code has to be in Epic — because the hospital is legally required to maintain that record in an auditable format for years. A hospital cannot replace Epic with an AI agent. It can deploy an AI agent that works with Epic. Epic becomes more valuable as the agent layer grows, because now there are more systems that need to connect to the source of truth it holds.

Veeva Systems does the same for pharmaceutical companies — managing clinical trial data, regulatory submissions, and drug safety records that must meet FDA standards. FactSet and Bloomberg Terminal hold the financial data that portfolio managers are contractually required to use. Procore manages construction contracts and compliance documentation that owners, contractors, and insurers all need access to simultaneously.

The EU AI Act classifies systems like Workday's HR platform as "High Risk AI" — meaning any startup that wants to replace it now has to meet the same regulatory bar that Workday spent twenty years building toward. That is an accidental moat, created by regulators who are trying to govern AI risk but are simultaneously making the incumbents harder to displace.

SaaS Capital used Claude Opus 4.5 to score 52 SaaS companies on AI disruption risk. The pattern was clean: companies where customers are legally required to use them, or where their data is the source of truth for regulated processes, consistently scored as lowest risk. Companies that do one specific workflow task — and that task can now be done by a general-purpose AI — scored as highest risk. The market selloff did not distinguish between them. The evidence does.

---

The Anthropic Problem — and What It Means for Everyone Building on AI

There is a separate problem inside this story that deserves attention, and it applies to any software company that has built its product on top of an AI model.

Figma integrated Claude deeply into its platform. Its AI design tool, Figma Make, runs on Anthropic's models. Lovable — the AI prototyping startup — also runs on Claude. Both companies built real products using Anthropic as the intelligence layer. Then Anthropic launched a competing product in the same category.

This is a new version of a very old problem in technology. Amazon's AWS has competed with companies that built their products on AWS for twenty years. App developers built businesses on Apple's iOS and then discovered Apple built the same feature into the operating system. But the AI version is more acute, because the foundation model is not passive infrastructure — it understands the domain well enough to enter it as a product. The more capable the underlying AI gets, the more adjacent markets it can enter on its own.

Anthropic's annualized revenue crossed $30 billion in early April. Venture capital is valuing the company at up to $800 billion. A company at that scale and growth rate has strong incentives to expand from model provider to full-stack product company. Every SaaS company that uses an AI model as its core intelligence layer is now one product announcement away from competing directly with its infrastructure provider. That risk was not in the original SaaS pricing thesis. It is now.

---

The Read

The market's $2 trillion judgment on SaaS in 2026 is directionally right and specifically wrong. Directionally right: the per-seat model that underwrote SaaS valuations for twenty years is under pressure. AI agents do work without taking seats. That arithmetic has real consequences for future revenue growth. Specifically wrong: it treated a pharmaceutical compliance platform with the same brush as a $30-per-month meeting scheduler. Those two businesses face completely different futures, and the selloff did not price the difference.

The companies worth watching are the ones that understood the transition earliest and moved first. HubSpot shifted to credits-based pricing before the selloff and grew 20% through it. The companies worth avoiding are the ones clinging to seat counts while quietly deploying agents that make their own product more efficient — that is a strategy that compresses your own revenue. And the companies worth owning long-term are the ones holding the data that agents need to do anything useful — because those become the infrastructure of the agentic economy, whether or not they ever get credit for it from the current market.

The SaaS business model ran on one bet for twenty years: every new employee means a new seat. AI broke that bet. What takes its place is not obvious yet — outcome pricing is the direction, but nobody has proved it works at scale except HubSpot. The companies that figure this out first will be worth more than they are today. The ones that don't will discover that their most profitable customers — the ones paying for seats they don't fully use — were the first to negotiate them away. ~ Framework

---

Market Truths · 財經真言 · Published Tuesday, Thursday, Saturday · markettruthspod.com