Dreamforce 2026's Answer: AI Agents, Have They Finally Moved from Demo to Revenue?

AI has spent so much money—has it finally started to "make money"?

Over the past two years, the most gratifying main narrative around AI in US stocks has been almost entirely about lining up to collect money on the infrastructure side: Nvidia printing cash, TSMC running at full tilt, Broadcom quietly getting rich off custom chips, with capital flooding through power and optical modules into the entire physical world.

The whole investment logic is simple and brutal—as long as large models keep getting bigger, compute is the most certain hard currency.

But standing in the fall of 2026, after hundreds of billions of dollars in capital expenditure has been poured into data centers, the market's patience is visibly narrowing, and people are starting to ask: when exactly will this compute be converted into real revenue on financial statements?

This watershed moment already tore open a crack in the US stock market in mid-September. On September 14, affected by discussions that AI development could be slowing down, chip stocks that had been sprinting for two years came under collective pressure, yet two security giants, CrowdStrike and Palo Alto, surged more than 13% in a single day, and even Salesforce and ServiceNow, which had been range-bound for a long time, saw a long-awaited return of buying.

Right at this critical juncture, Dreamforce 2026 also kicked off.

Setting aside all those dazzling keynotes and conceptual packaging, the entire conference was really just answering one core question for everyone: Can AI Agents actually turn demos into cold, hard cash?

1. Farewell to the "Chat Buddy": AI Starts "Getting Work Done"

Enterprise AI over the past two years, to put it bluntly, was mostly just an advanced version of Copilot.

An employee asks a question, and AI helps summarize documents, generate emails, write a piece of code, or organize meeting content. It can improve efficiency, but there must be a real person sitting in front of the computer feeding instructions step by step.

This year, Salesforce finally stopped telling this "co-pilot" story.

Take Hunter, which handles customer acquisition. In the past, Copilot could at best help you polish a cold email. Now the design is to let it make its own plans, dig through data, and keep following up for weeks. On the eve of Dreamforce, it expanded Agentforce's Agent products all at once, including Casey for customer service, Paige for IT and HR, Carter for e-commerce shopping, Hunter for sales, and Marshall for back-end supply chain processes.

It is designed to work across systems, across steps, and even across time continuously, thereby connecting the entire chain: find potential customers, make plans, keep following up, adjust strategy based on new customer information, ask salespeople for approval when needed, and then continue pushing the task forward.

This process can last for weeks or even longer. Although Hunter is still in the Pilot stage and is expected to reach GA only in November this year—it is not yet a large-scale, mature commercial product—the underlying business logic has already changed.

In the past, the relationship between people and AI was more like constantly sending prompts; in the future, what enterprises want to buy may be a set of digital labor that can read data on its own, call tools, execute tasks, and leave audit records.

Once a product changes from a "feature" into "digital labor," the ROI calculation becomes completely transparent.

Whether a chatbot writes well or not easily remains a subjective experience; but if a customer service Agent can directly reduce tickets, a sales Agent can generate pipeline, and a back-office Agent can reduce manual operations, then enterprises can directly calculate how much they paid for it and how much work it completed on their behalf.

Only at this point does AI truly touch the threshold of business.

2. Farewell to Seats: Charging by "Amount of Work Moved"

This is also the most noteworthy part of this year's Dreamforce.

Demos at launch events can always be edited to sound spectacular. To judge whether commercialization has landed, there are only two indicators—whether there is real cash flow, and whether there are real invocations.

Salesforce's latest quarterly earnings report provided a set of very noteworthy data. Among them, Agentforce ARR crossed $1.5 billion, up 240% year over year; combined ARR for Agentforce and Data 360 approached $3.9 billion, up more than 210% year over year. At the same time, Agentforce and Slack have cumulatively completed 7 billion Agentic Work Units, of which 3.2 billion were completed in the second quarter alone, up 97% quarter over quarter.

Although this $1.5 billion ARR includes Slackbot and other AI assets, and is somewhat dressed up by statistical scope, it at least shows that enterprises are no longer just "trying it out."

In customer cases disclosed by Salesforce, for example, about 50% of Engine's chat inquiries can already be fully resolved by its Agent; about 60% of Perk's sales pipeline is built by sales Agents; about 70% of Autism Queensland's administrative requests are handled by Agents; Hibbett's AI already participates in about 90% of its core shopping process; and when Anthropic uses Fin, about 79% of customer service conversations that Fin touches can be automatically resolved.

The statistical scopes of these data points differ, and they all come from Salesforce's official disclosures, so they cannot yet prove that the entire Agent industry has matured. But they at least represent that the core of everyone's attention has truly shifted from "how smart is it" to "how much work has it carried for me."

More critically, Salesforce's own charging method is also beginning to adapt to this change.

Agentforce has already launched Flex Credits. In the current public price list, 100,000 Credits cost $500, and one standard Agent Action consumes 20 Credits, or about $0.10. Enterprises can directly pay based on usage when an Agent updates a record, processes a workflow, or executes an action.

This may be more important than the $1.5 billion ARR itself.

After all, the most core business unit of traditional SaaS is the seat: 1,000 employees means selling 1,000 accounts.

But the problem Agents ultimately want to solve is precisely to let fewer people complete more work. If a business that previously needed 10 people to handle can in the future be handled by only 3 people plus a batch of Agents, then if software companies continue to rely purely on seat-based charging, a very awkward problem may emerge: the more successful AI is, the fewer human seats there are.

Therefore, Salesforce's rush to promote usage-based billing is essentially trying to find a new pricing unit for the post-SaaS era ahead of time.

3. Undercurrents Beneath the Surface: Security, Interfaces, and the Gross Margin Test

If this change holds, it will not affect only Salesforce.

As is well known, the most important keyword in the AI market over the past few years has been CapEx.

Training models requires GPUs, GPUs require data centers, and data centers require power, networking, optical modules, and storage. Therefore, as long as Hyperscalers continue to raise capital expenditure, the infrastructure chain can continue to benefit.

Once Agents are truly commercialized, AI will gain another value chain: entering enterprise workflows from models, then connecting enterprise data, identity, permissions, security, and real business systems.

For example, companies like Salesforce and ServiceNow control workflow; platforms like Snowflake connect enterprise data; and security vendors such as CrowdStrike, Palo Alto Networks, SailPoint, and Varonis will face increasingly complex problems as the number of Agents grows.

In the future, an enterprise may not only have tens of thousands of employees, but also run thousands or even tens of thousands of "non-human identities" at the same time. These Agents can read internal documents, call APIs, modify CRM records, send emails, operate code, and even participate in transaction processes.

By then, the questions enterprises face become: Who is it? What can it see? What can it do? On whose behalf is it acting? If something goes wrong, can it be traced? So the closer Agents get to real production environments, the more capabilities that used to seem back-office—Identity, Permission, Data Governance, Runtime Security—will move closer to the core of the AI application layer.

This is also why the rotation into cybersecurity stocks on September 14 is worth watching. AI's attack surface is expanding from human accounts, devices, and servers to more and more Agents with autonomous execution capabilities. What CrowdStrike and Palo Alto are benefiting from is not short-term sentiment, but the dramatic spillover of AI's attack surface.

At the same time, this commercialization path is still far from fully proven, and there are still two hard hurdles ahead:

• Gross margin erosion: Every time an Agent acts, behind it is real spending on inference, retrieval, and cloud resources. After deducting this, can it still leave behind the attractive high gross margins of traditional SaaS?
• Self-cannibalization: Can the incremental revenue brought by usage outrun the pace of shrinking human seats? If not, it is just moving money from one hand to the other.

Dreamforce 2026 has another change that may make this test even more interesting.

Salesforce's newly launched AIforce is opening up Salesforce's data, workflow, business logic, permissions, and governance capabilities—originally locked inside the CRM interface—to new AI interfaces such as Claude, Slack, and Agentforce Coworker. Claudeforce initially provides 37 pre-built sales Skills; Agentforce Coworker gained 100,000 user activations within 35 days of launch—of course, activation numbers cannot be directly equated with sustained active users or paying customers.

This means competition in enterprise software may undergo yet another shift.

In the past, the most important thing about Salesforce was that CRM interface; in the future, users may not even need to open Salesforce, but instead directly ask AI inside Claude or Slack to call the data and workflows behind Salesforce. In other words, the interface is disappearing, but the data, permissions, processes, and business logic in the back end may actually become more valuable.

If we look at this round of the AI market over a longer horizon, the past few years have in fact gone through two very clear stages.

• The first stage was training AI, and the biggest beneficiaries were GPUs, advanced process nodes, and ASICs;
• The second stage was building AI, as data centers, power, optical communications, and storage began to absorb massive capital expenditure;
• And the signals released by Dreamforce 2026 so far mark the market officially hitting the wall of the third stage: turning compute into productivity;

This path has not yet been fully proven, let alone can it be said that software has already "taken the baton" from semiconductors. But at least, for the first time, enterprise AI has shown an increasingly complete commercial loop—someone buys Agents, Agents start completing work, work generates usage, and usage turns into revenue for software companies.

Overall, the commercial loop is already vaguely taking shape, even though it is still very bumpy.

But the most honest logic in the business world will never change—whoever can truly convert enterprises' anxiety about spending on compute into profit on the balance sheet will be the winner of the next game.

Followthe official X for US stock news

Jointhe official community to discuss hot topics