Every big software platform now sells AI agents for the office, a marketplace to buy them from, and a control room to watch them in. The products are real, and so is the cancellation wave the analysts predict. Here is the state of play, with every vendor number labeled as the vendor’s own.
The shortest honest summary of ai agents enterprise news is this: the building phase is over, and the governing phase has begun. Over the last eighteen months, Salesforce, Microsoft, ServiceNow, Snowflake, Oracle and SAP have each shipped the same three products in the same order. First a studio where companies build AI agents. Then a marketplace where they buy ready made ones. Then a control plane where IT watches what those agents do and what they spend. An AI agent, in plain terms, is software that takes a goal, works out the steps, and carries them out across other programs with limited human checking. If the term itself is new, our explainer on what agentic AI is covers the basics before the news.
Why the rush to govern? The first wave of deployments taught a hard lesson. An agent that acts on its own can also spend on its own, make mistakes on its own, and touch data it was never meant to see. So the loudest launches of 2026 were not smarter agents. They were registries, gateways, signed credentials and kill switches. This roundup covers what each platform shipped, which deployment numbers come from independent research and which come from the seller, and the governance questions that now decide whether a deal closes.
| The enterprise agent race at a glance | |
|---|---|
| Salesforce | Agentforce builder and AgentExchange marketplace. Control: Command Center |
| Microsoft | Agents built in Copilot Studio. Control: Agent 365 registry and control plane |
| ServiceNow | AI Agent Orchestrator and agent teams. Control: AI Control Tower |
| Snowflake | Cortex Agents, CoWork and CoCo. Control: Cortex AI Gateway |
| Oracle | AI Agent Studio for Fusion, free to customers. Control: testing, validation and security in the studio |
| SAP | Joule Assistants and Joule Agents. Control: existing authorizations and audit trails |
Every platform shipped the same three products
Follow ai agents enterprise news for a month and one pattern jumps out: the launches repeat. Builder, marketplace, control room. The logos change, the stack does not.
Salesforce started the race and still shouts the loudest. It launched Agentforce in October 2024 and pushed it harder than any product in its history. Agentforce 3, released in June 2025, added the Command Center, a screen where managers watch agent activity and results, built in support for MCP, the open standard that lets agents plug into outside tools, and an expanded AgentExchange with services from more than 30 partners and over 100 ready made industry actions. Salesforce said 8,000 customers had signed up to deploy Agentforce by that point. At Dreamforce in October 2025 the platform took a new name, Agentforce 360, and a wider pitch: the whole company reorganized as an agentic enterprise.
Microsoft built its offer around the software most offices already rent. Companies create agents in Copilot Studio, coordinate teams of them with multi-agent orchestration, and reach them inside Teams and Outlook. At Ignite in November 2025, Microsoft added the missing layer: Agent 365, a registry and control plane that inventories agents built in Copilot Studio, on its Foundry platform, by partners and from open source projects, and governs them in one place. In October 2026 it went a step further and gave agents their own accounts inside Microsoft 365, so an agent can hold its own credentials, be mentioned in a document, and keep working between requests. That rollout began through the Frontier early access program, a useful reminder that even the biggest launches still arrive in stages.
Oracle moved earliest of the suite vendors. Its AI Agent Studio opened on March 20, 2025, at CloudWorld in London, free for Fusion Applications customers, on top of more than 50 ready built agents already inside the suite. In October 2025, Oracle added an AI Agent Marketplace where partners sell agents that Oracle has validated. The pitch is convenience with a leash: build inside the same system that holds the finance and HR data, and let the studio’s testing and validation tools check the agent before it goes live.
SAP sells context. Its Joule assistants and agents sit on the data that runs factories, warehouses and payrolls. At SAP Connect on October 8, 2026, the company expanded six Joule assistants for supply chains with specialized agents across planning, manufacturing, logistics and asset operations, with general availability planned for the fourth quarter of 2026. SAP named Takeda as an adopter of its planning assistant and said one customer, Doehler, had won back about a third of the time its logistics team used to lose to manual troubleshooting. Both are SAP’s own customer claims, and we print them as exactly that.
Underneath the rivalry, the platforms quietly agreed on the plumbing. MCP, short for Model Context Protocol, is an open standard that gives agents one common way to connect to outside tools and data. Salesforce built it into Agentforce 3, SAP wired it into Joule Work alongside the Agent2Agent protocol for agents that talk to each other, and Snowflake’s gateway manages access to more than 100 MCP servers at once. The models still differ. The connectors, increasingly, do not. For buyers, that sameness matters more than any single launch, because it decides whether an agent bought from one vendor can work inside another vendor’s stack, or gets locked out of it.
Snowflake moves the fight to the data
The most telling item in recent ai agents enterprise news came from a data company, not an app company. Snowflake built its agents where the company data already sits. Cortex Agents answer questions and take steps over governed data, Snowflake CoWork gives business users an agent front end, and CoCo does the same job for developers. Snowflake says CoCo already runs in more than 7,100 accounts, a figure from the company’s own earnings commentary.
Then, on July 28, 2026, Snowflake announced Cortex AI Gateway, and the announcement read like a security product, because it is one. The gateway sits between agents and the models, tools and data they reach. It governs Snowflake’s own agents and third party ones such as Claude Code and Cursor, supports more than 100 MCP servers, keeps an end to end record of agent activity, and attributes AI spend to teams with limits attached so runaway costs get stopped early. It builds on Natoma, the agent connectivity company Snowflake acquired in May 2026, and it launched with identity integrations from Okta, 1Password, SailPoint and Saviynt. Read the fine print and the story gets more honest, not less: Snowflake’s own announcement marked the gateway as heading to public preview, with most partner integrations in private preview. The gate is real. Parts of it are still being built.
Real deployments are narrower than the demos
Set the ai agents enterprise news next to the independent research and a colder, clearer picture forms. The deployments that survive share one shape: high volume, repetitive, easy to score. Internal IT requests. Software provisioning. Support triage. The wins are real, and they look nothing like the keynote demo of an agent running a whole department.
ServiceNow runs its own platform inside its own company and publishes the results. In its agentic workforce announcement, the company said its internal agents automate 97 percent of software provisioning requests, cut service desk volume by nearly 40 percent, and solve 85 percent of routine internal IT requests on their own, with AI Control Tower, available since May 2025, watching over the fleet. Those are ServiceNow’s numbers about ServiceNow: believable in direction, unaudited in size. Salesforce tells similar stories about its customers, including 1-800Accountant resolving 70 percent of administrative chats by itself during tax season and Grupo Globo lifting subscriber retention by 22 percent. Again, seller’s numbers. Useful as proof the pattern can work somewhere, useless as a promise it will work for you.
The independent checks cut the other way. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, killed by escalating costs, unclear business value and inadequate risk controls. It also estimates that only about 130 of the thousands of companies selling agents offer the real thing, a habit it calls agent washing: old chatbots and automation wearing a new badge. A 2025 study from MIT’s NANDA initiative, which reviewed some 300 public AI initiatives, found that 95 percent of organizations saw no measurable return from their generative AI pilots. That study covered AI pilots broadly, not agents alone, and its headline number has been argued over ever since. The direction, though, matches every other dataset we have: most pilots stall, and the minority that scale started narrow. Our tour of agentic AI use cases shows the same shape from the buyer’s side of the table.
Governance is now the product
The loudest argument in ai agents enterprise news is no longer about how smart agents are. It is about who answers for them. Five questions now show up in nearly every enterprise deal. Does the agent act under its own identity or borrow a person’s? Can it touch only what the task needs? Is every action recorded? Does the bill arrive per agent, with a cap? And how fast can someone switch it off?
October 2026 supplied the freshest exhibits. ServiceNow announced AI Passport, a signed digital credential that states what each agent is authorized to do and gets verified as the agent works, with NVIDIA as launch partner and availability expected later in 2026. Days later, its personal agent Arc went live under the renamed Super Intelligence Control Tower, with NVIDIA’s OpenShell layer enforcing the rules while the agent runs. Microsoft’s agent accounts attack the same problem from the identity side: an agent with its own credentials can be audited, limited and removed like an employee who left the company.
Analysts expect casualties from this fight. In May 2026, Gartner predicted that by 2027, 40 percent of enterprises will demote or decommission autonomous agents after incidents reveal governance gaps nobody had closed. Its diagnosis is precise: companies govern agents as a binary choice, locked down or fully trusted, instead of matching the control to the freedom each agent actually has. The remedy it recommends, tiered control by autonomy level, is the same thinking behind our guide to AI governance in a business context. And when several agents share one job, the oversight problem multiplies, which is why the multi-agent ai news keeps circling back to a single question: who supervises the team?
What to watch next
The next stretch of ai agents enterprise news comes with dates attached. ServiceNow’s AI Passport is expected in the fourth quarter of 2026. SAP’s new supply chain agents are due for general availability in the same quarter. Snowflake’s gateway moves from announcement toward public preview. Microsoft’s Ignite conference runs from November 17 to 20, and the company has signaled more agent news there. Watch less for new capabilities and more for the boring words: general availability, named customers, published prices. The gap between announced and shipped is where this market currently lives.
Conclusion
That is the state of ai agents enterprise news as October 2026 closes. The builders are built, the shelves are stocked, and the control room is the product now. Enterprises are not short of agents; they are short of proof. Proof of identity, proof of limits, proof of value. The vendors that win the next year will be the ones whose agents survive an audit rather than a demo, and the projects that survive will be the narrow, measured ones that were governed like employees from day one.
Frequently asked questions
Which companies lead the enterprise AI agent market?
The same names that already run the office software market. Salesforce sells Agentforce, Microsoft builds agents in Copilot Studio under the Agent 365 control plane, ServiceNow pairs its AI Agent Orchestrator with AI Control Tower, Snowflake runs Cortex agents next to the data, and Oracle and SAP build agents into Fusion and Joule. Each now sells a builder, a marketplace and a governance tool.
Are enterprise AI agents delivering real results?
Some are, in narrow jobs. Vendors report strong numbers in support triage and internal IT, and those figures are the vendors’ own. Independent research is colder: Gartner predicts over 40 percent of agentic AI projects will be canceled by the end of 2027, and MIT research found most organizations saw no measurable return from generative AI pilots.
What worries enterprises most about AI agents?
Control. Companies ask who an agent acts as, what it may touch, whether every action is recorded, what it costs, and how fast it can be switched off. Gartner predicts that by 2027, 40 percent of enterprises will demote or decommission autonomous agents after incidents expose governance gaps.
Written by
Ahmad Jajja
Ahmad covers AI, chips and the companies behind them. For this roundup he read each platform’s own announcement and labeled every vendor number as the vendor’s own.
Reviewed by
Danish Mustafa
Danish reviews laptops, phones and home tech, and fact-checks the numbers in our AI coverage. Here he checked every date and figure against the original releases.