Dreamforce 2026: Can AI Agents Turn Demos into Revenue?

During the gold rush, those selling shovels made a fortune, but if miners keep striking no gold, the shovels will eventually stop selling.

Written by Jim, MSX MaiTong

Edited by: Frank, MSX Maotong

After spending so much money on AI, is it finally starting to make money?

Over the past two years, the most compelling theme in U.S. equities around AI has been almost entirely centered on infrastructure: NVIDIA printing money, TSMC running around the clock, Broadcom quietly profiting from custom chips, with capital flooding throughout the physical world via power and optical modules.

The entire investment logic is straightforward— as long as large models continue to grow, computing power remains the most certain form of hard currency.

But standing in the autumn of 2026, after hundreds of billions of dollars in capital expenditures have been poured into data centers, market patience is visibly narrowing, with growing questions about when this computing power will translate into real revenue on financial statements.

This turning point had already begun to show signs on the U.S. stock market in mid-September. On September 14, amid discussions that AI development momentum might slow, chip stocks that had been surging for two years collectively faced pressure, yet cybersecurity giants CrowdStrike and Palo Alto each surged over 13% in a single day, and even long-stagnant Salesforce and ServiceNow saw a welcome return of buying interest.

At this very moment, Dreamforce 2026 has also kicked off.

Setting aside the dazzling Keynotes and conceptual packaging, the entire conference was really just answering one core question for everyone: Can AI Agents truly turn demos into real, tangible money?

FollowOfficial X for US stock market updates

Jointhe official community to discuss trending topics

Two years ago, enterprise AI was, in essence, 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 notes—it can boost efficiency, but there must still be a live person sitting at the computer, providing step-by-step instructions.

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

For example, Hunter, who is responsible for client acquisition, previously used Copilot to merely polish a cold email; now, its design enables it to create plans, analyze data, and follow up consistently over several weeks. Just before Dreamforce, it expanded Agentforce’s Agent suite to include Casey for customer service, Paige for IT and HR, Carter for e-commerce shopping, Hunter for sales, and Marshall for supply chain backend processes.

It is designed to work continuously across systems, across steps, and even across time, seamlessly connecting the entire process: identifying potential customers, creating plans, following up consistently, adjusting strategies based on new customer information, requesting sales approval when needed, and then moving the task forward.

This process can last several weeks or longer, and although Hunter is still in the pilot phase and not expected to reach general availability until November, it is not yet a fully mature commercial product, the underlying business logic has already changed.

In the past, the relationship between humans and AI resembled constantly issuing prompts; in the future, what companies may want to purchase could be a digital workforce capable of autonomously reading data, invoking tools, executing tasks, and maintaining audit trails.

Once a product transitions from "feature" to "digital labor," the ROI algorithm becomes completely transparent.

How well a chatbot performs is often subjective; but if a customer service agent reduces tickets, a sales agent generates pipeline, and a backend agent reduces manual tasks, companies can directly calculate how much they paid for it and how much work it accomplished on their behalf.

AI has only now reached the threshold of real business.

This is also one of the most worthwhile highlights of Dreamforce this year.

Demos at press events can always be edited to look spectacular; to determine whether commercialization has truly been achieved, there are only two metrics—whether there is real revenue and real usage.

Salesforce's latest quarterly earnings report revealed a set of highly noteworthy figures: Agentforce ARR surpassed $1.5 billion, a 240% year-over-year increase; combined Agentforce and Data 360 ARR approached $3.9 billion, up more than 210% year-over-year. Meanwhile, Agentforce and Slack have collectively completed 7 billion Agentic Work Units, with just the second quarter accounting for 3.2 billion—a 97% quarter-over-quarter increase.

Although this $1.5 billion in ARR includes Slackbot and other AI assets, with some statistical framing, it at least indicates that enterprises are no longer just "experimenting."

As shown in Salesforce’s customer case studies, Engine can fully resolve approximately 50% of chat inquiries via its Agent; Perk generates about 60% of its sales pipeline through sales Agents; Autism Queensland handles around 70% of administrative requests via Agent; Hibbett’s AI is involved in approximately 90% of its core purchasing processes; and when Anthropic uses Fin, about 79% of customer service conversations touched by Fin are resolved automatically.

These statistics vary in methodology and all come from official Salesforce disclosures, so they cannot yet be used to prove that the entire Agent industry has matured; however, they at least indicate that the core focus has genuinely shifted from “how smart it is” to “how much work it helps me get done.”

More importantly, Salesforce's own pricing model is beginning to adapt to this change.

Agentforce has launched Flex Credits. Under the current public pricing, 100,000 Credits cost $500, with each standard Agent Action consuming 20 Credits—approximately $0.10. Businesses can pay directly based on usage, such as updating a record, processing a workflow, or executing an action.

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

After all, the core business unit of traditional SaaS is the seat: for 1,000 employees, you sell 1,000 accounts.

The ultimate problem Agent aims to solve is enabling fewer people to accomplish more work—if a task that previously required 10 people can, in the future, be handled by just 3 people plus a team of Agents, software companies relying solely on seat-based pricing may face an awkward dilemma: the more successful the AI, the fewer human seats are needed.

Therefore, Salesforce is rushing to implement usage-based pricing, essentially seeking to establish a new pricing unit for the post-SaaS era.

If this change takes effect, it will affect more than just Salesforce.

It is well known that over the past few years, the most important keyword in the AI sector has been CapEx.

Training models requires GPUs, which require data centers, and data centers require power, network connectivity, optical modules, and storage. Therefore, as hyperscalers continue to increase capital expenditures, the entire infrastructure chain will continue to benefit.

Once Agent is fully commercialized, AI will gain another value chain: connecting models to enterprise workflows, and integrating with enterprise data, identity, permissions, security, and real business systems.

Companies like Salesforce and ServiceNow manage workflows; platforms like Snowflake connect enterprise data; and security vendors such as CrowdStrike, Palo Alto Networks, SailPoint, and Varonis face increasingly complex challenges as the number of agents grows.

Inside a future enterprise, there will likely be not just tens of thousands of employees, but also thousands or even tens of thousands of "non-human identities" running simultaneously—these Agents can read internal documents, call APIs, modify CRM systems, send emails, manipulate code, and even participate in transaction processes.

By then, the questions businesses face will become: Who is it? What can it see? What can it do? Who does it represent? And can its actions be traced if something goes wrong? Therefore, as Agents get closer to real production environments, capabilities once seen as backend—Identity, Permission, Data Governance, and Runtime Security—will increasingly move to the core of AI applications.

This is why the cybersecurity stock rotation on September 14 is significant: AI’s attack surface continues to expand from human accounts, devices, and servers to an increasing number of agents with autonomous execution capabilities—CrowdStrike and Palo Alto are benefiting not from short-term sentiment, but from the severe spillover of AI’s expanding attack surface.

Meanwhile, this commercialization path is still far from being fully realized, with two major hurdles clearly ahead:

• Margin erosion: Every time an Agent acts, it incurs real costs in reasoning, retrieval, and cloud resources—after accounting for these expenses, can there still be the attractive high margins typical of traditional SaaS?
• Self-cannibalization: Can the incremental usage offset the decline in manual seats? If not, it’s just taking from one pocket to pay the other.

There's another change at Dreamforce 2026 that might make this test even more interesting.

Salesforce's newly launched AIforce is opening up data, workflows, business logic, permissions, and governance capabilities previously confined to the CRM interface, extending them to new AI interfaces such as Claude, Slack, and Agentforce Coworker. AIforce initially offers 37 pre-built sales skills; Agentforce Coworker has achieved 100,000 user activations within 35 days of launch—though activations do not directly equate to actively engaged or paying customers.

This means the competition in enterprise software could shift once again.

Previously, Salesforce’s most important feature was its CRM interface; in the future, users may not even need to open Salesforce at all, but instead directly ask Claude or Slack to invoke Salesforce’s underlying data and workflows via AI. In other words, the interface is disappearing, but the backend data, permissions, processes, and business logic may become even more valuable.

If we look at this AI market cycle over a longer timeframe, the past few years have already gone through two very distinct phases.

• The first phase is training AI, with the biggest beneficiaries being GPUs, advanced processes, and ASICs;
• In the second phase, AI, data centers, power, optical communications, and storage begin to absorb massive capital expenditures;
• Meanwhile, the signals released so far for Dreamforce 2026 mark the market's official collision with the third phase wall: turning computational power into productive force;

This path has not yet been fully realized, let alone has software already "taken the baton" from semiconductors. But at least, for the first time, enterprise AI has begun to show a increasingly complete business loop—someone purchases an Agent, the Agent begins to perform tasks, the tasks generate usage, and the usage translates into revenue for the software company.

Overall, the business loop has begun to take shape, though it’s still bumpy.

But the most honest logic in the business world never changes—those who can truly convert the anxiety businesses feel over computing power into actual profits on their balance sheets will be the winners of the next round.