Here’s what’s on our plate today:

• 🧪 AI agents, SaaS pricing, and the quiet rewrite of knowledge work.

• ⚡ AI’s GDP dud, Sora shuts down, and Arm moves into AI chips.

• 🧰 Three Things Worth Trying: Gumloop, Zapier Agents, and Dust.

• 📊 Thursday poll on what changes first as AI agents spread.

Let’s dive in. No floaties needed…

The Laboratory

AI agents, SaaS pricing, & the quiet rewrite of knowledge work

TL;DR

• The SaaSpocalypse was a pricing crisis, not a tech crisis: In February 2026, $285B in SaaS market value evaporated when investors realized that AI agents could bypass per-seat licensing, threatening the revenue model that has powered enterprise software for two decades.

• NVIDIA’s data says the shift is real: The State of AI 2026 report found 86% of enterprises plan to increase AI budgets, 64% already have AI in live operations, and agentic AI is the fastest-growing category, meaning systems that execute work without needing a licensed human seat.

• Incumbents are automating against themselves: Salesforce’s Agentforce already has 8k+ customers, and Microsoft 365 Copilot has hit 15M paid seats, but both are layering AI on top of subscription products rather than replacing them, because true automation weakens the pricing models those products rely on.

• The quiet cost is the career pipeline: What is being automated is not just routine busywork, but the contract reviews, report processing, and junior tasks through which people historically built expertise. Atlassian cut 10% of its workforce to fund AI investment, and if that pattern scales, the industry risks hollowing out its future senior talent.

For the past two decades, enterprise software has been sold using a simple formula built around per-seat subscriptions. Companies would estimate how many employees needed access to a tool, multiply that number by a fixed monthly or annual fee, and sign multi-year contracts with vendors. The model became standard during the rise of cloud software in the 2000s and 2010s, when providers such as Salesforce, Microsoft, and Adobe shifted from one-time licenses to recurring subscriptions.

Driving adoption for this model was the revenue predictability it offered software companies and the ease of budgeting for customers. But since the model tied pricing directly to headcount rather than actual usage, it was ready for the shifts brought in by artificial intelligence. The model worked as long as human employees were the primary users; however, with automated systems taking over, the economics no longer work for software companies or their clients.

Signs of the model’s inability and the shift were evident on 3 February 2026, when about $285B in market value of SaaS providers was wiped out in a single trading session. Atlassian fell 35%, while Salesforce dropped 28%. Interest rates, earnings, or any broader economic shock did not trigger the decline. Instead, investors were reacting to a growing realization that the traditional pricing model for enterprise software may be breaking down, because AI agents do not need user seats the way human employees do.

The selloff that started the question

The “SaaSpocalypse,” as the selloff came to be called, marked the start of a broader shift that became even clearer a month later at NVIDIA’s GTC developer conference in San Jose in March 2026. During the event, NVIDIA released its State of AI 2026 report, based on a survey of 3,2k enterprise professionals across five industries, and the results painted a picture that was encouraging for AI but unsettling for traditional software vendors.

According to the findings, 86% of enterprises plan to increase AI budgets in 2026, while 88% report AI-driven revenue increases. 64% now have AI running in live operations, not just experiments, and crucially, the type of AI spreading fastest in enterprise environments is agentic: systems that plan, reason, and execute multi-step tasks autonomously, without a human approving each step.

The findings are not just a snapshot of the shift in the software industry, but also explain why the SaaSpocalypse played out the way it did.

One employee, four seats,  one agent, none

Consider a typical contract review workflow. A legal associate logs into the contract management system, checks clauses against a compliance database, writes a summary in a document editor, and sends it for approval through a project management tool. Each step often requires a separate product and license, resulting in several paid seats for a single employee.

Now imagine the same workflow handled by an AI agent. The agent connects to the same systems but does not log in as a human user; rather, it runs as a software process. In many current enterprise licensing agreements, this creates a gray area. Agents may be billed for API usage or compute time, or not billed at all, because the contracts were written before this kind of automation existed. What they are rarely billed as is a user seat.

Analysts at Deloitte say this shift will push enterprise software away from pure subscription pricing toward hybrid models that combine usage-based and outcome-based fees. Most observers agree that this transition is coming. However, the real debate is over how quickly it will happen and which companies will capture the value when it does.

The transition is real, but slower than expected

On how fast this shift will happen, the evidence points in two directions. Some forecasts suggest rapid adoption, but real-world data show the transition may be slower and messier. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, largely because systems that work in controlled pilots struggle once they meet the complexity of real enterprise environments.

Composio’s analysis of failed deployments describes the same pattern: early demos run on clean data and narrow tasks, while production systems must deal with messy inputs, edge cases, and difficult integrations that quickly eat into the promised efficiency gains.

Deloitte’s State of AI 2026 report adds another layer, finding that only a quarter of organizations have moved a large share of their AI experiments into production, and even then, most companies have only a handful of agents running in live workflows. Being technically in production is not the same as being operationally indispensable, and the gap between the two is where many projects stall. In other words, reaching production is common, but becoming truly essential is still rare.

Incumbents defending a model that AI weakens

Markets, however, are not waiting for that distinction to play out before repricing the sector, and neither are the incumbents. Salesforce says its Agentforce platform already has more than 8k customers and roughly $100M in annual recurring revenue. At the same time, Microsoft reported over 15M paid Microsoft 365 Copilot seats in its fiscal Q2 2026 earnings call, a 160% year-over-year increase, even though paid Copilot users still account for only around 3.3% of its roughly 450M commercial seats. Adoption is real, but it is not yet replacing the existing model; instead, it is being layered on top of it.

That is where the value-capture problem becomes more complicated than the disruption narrative suggests. The fastest way for incumbents to defend themselves against AI-native competitors is to build agents directly into their own products, yet doing so also weakens the very pricing structure that made those products valuable.

When Salesforce automates parts of CRM workflows, fewer human operators need CRM licenses. The incumbents are not simply competing with startups. They are, in effect, competing with their own business model.

And they are not the only ones claiming a share of the value being created. Enterprises keep part of it as a margin improvement when automation reduces headcount. Model providers such as Anthropic and OpenAI capture another portion through API usage generated by agents.

The infrastructure layer has already secured its position, with NVIDIA projecting an order backlog approaching $1T through 2027, meaning the hardware economics are largely locked in regardless of how software pricing evolves. Gartner estimates that agentic AI could account for roughly 30% of enterprise application software revenue by 2035, up from a small share today, leaving software vendors, incumbents, and startups alike competing for whatever remains after the other layers take their cut.

Pricing is changing, work is changing with it

Most analysts believe outcome-based pricing is where the industry is heading, but implementing it is harder than it sounds. Under the traditional per-seat model, vendors earn money based on how many employees use the software, not on how much work the software actually gets done.

In an outcome-based model, revenue depends on results, contracts reviewed, tickets resolved, code written, or leads generated. That shift benefits customers but makes revenue less predictable for vendors.

If AI systems perform poorly, the vendor earns less, not just the customer. It also means software companies have to track performance in ways their products were not originally designed to measure. Deloitte expects pricing models to keep evolving for several years, with subscriptions, usage fees, and outcome-based pricing all coexisting well into the decade.

What the report does not answer

NVIDIA’s report documents adoption, and the February repricing documents market anxiety. GTC documents that the hardware layer has committed capital regardless of how the software debate resolves. What none of these data points addresses is the human cost sitting underneath the pricing story.

What is being automated now is knowledge work: the junior lawyer reviewing contracts, the financial analyst processing reports, the software tester running checks, the customer support agent answering queries. These are not low-skill jobs; they are the entry-level roles through which people historically built careers in law, finance, and technology.

What usually happens is quieter than a headline layoff. Instead of sudden cuts, companies stop replacing people who leave, freeze hiring, or ask smaller teams to do the same work in the name of efficiency. When layoffs do happen, they are often tied to AI spending. Atlassian, for example, cut about 10% of its workforce in March 2026 and said the restructuring would help fund its AI investments. Variations of this pattern are starting to appear across the tech industry.

The deeper structural problem is one that the report does not model. The junior lawyer who reviews 1k contracts learns, through that process, how to become a senior partner. If agents do that reviewing, the pipeline of experienced lawyers a decade from now may be thinner than anyone has accounted for. The same logic applies across finance, software development, and broader business operations. Automating the entry level does not just change who does the work today. It changes who is capable of doing the harder work decades from now.

So what is happening beneath all the stock movements and analyst forecasts is a renegotiation of three things at once: how software gets priced, who does cognitive work, and what a career in a knowledge profession looks like over the next decade.

The NVIDIA report and the GTC conference are data points in all three conversations simultaneously. None of those negotiations has concluded, and the contract renewal cycle, conducted quietly across thousands of enterprise agreements over the next few years, will do more to answer them than any forecast produced this week.

Quick Bits, No Fluff

• AI boom, zero GDP: Futurism says Goldman now sees basically no measurable boost to U.S. economic growth from the 2025 AI spending spree, despite hundreds of billions poured in.

• Sora gets shut down: TechCrunch reports OpenAI is pulling the plug on Sora’s standalone app, ending one of the creepiest AI video experiments to land on people’s phones.

• Arm wants the stack: The Verge says Arm’s first in-house CPU will land in Meta’s AI data centers this year, a sign the AI infrastructure fight is moving deeper into custom silicon. 

Thursday Poll

3 Things Worth Trying

• Gumloop: No-code AI workflow builder for turning repetitive knowledge work into automations across your tools.

• Zapier Agents: Agent layer for connecting apps, triggering actions, and testing where AI can replace manual workflow glue.

• Dust: Enterprise AI workspace for building internal assistants that work across docs, apps, and team knowledge.

The Toolkit

• Pika: Fast, social-first AI video tool built for creators who want quick scenes, effects, and remixable clips without pro editing skills.

• Runway: Higher-end AI video platform for more controlled, cinematic generation and real production workflows.  

• CapCut: Consumer-friendly editing stack with AI video tools, captions, speech features, and templates already built for short-form distribution.

Rate This Edition