Salesforce's Winter '27 is live, and it switches Agentforce on by default

Salesforce just turned the agents on for you

On 12 October 2026 Salesforce's Winter '27 release reaches general availability, and the loudest part of it is a list of autonomous agents. The quieter part matters more to the people who run these systems. Agentforce, Salesforce's set of AI agents, is now enabled by default.

Until now an administrator had to switch Agentforce on. Starting with Winter '27, Salesforce auto-enables it across eligible orgs. That covers Lightning Experience with Foundations on Enterprise, Performance, Unlimited and Developer Edition, plus the Agentforce 1 Edition. New orgs are on at creation. Existing orgs are switched on a rolling basis that began in the first week of September 2026, and the on/off toggle is being removed from the Agentforce Agents page in Setup later in the release. Salesforce says the change costs nothing extra and does not alter billing.

The framing is convenience. Removing the toggle, the company's release notes say, lets teams "build, test, and deploy agents with one less step." The practical effect is that a capability which used to be opt-in is becoming opt-out, on software that sits on top of a lot of companies' customer data.

From an assistant to something that finishes the job

The pitch behind Winter '27 is a shift in what an agent is for. Earlier Salesforce agents mostly reacted: you prompted, they answered. The headline feature this time, Adaptive Experiences with Dynamic Plans, is built to run continuously. TechTimes describes it as "ambient AI" that "listens to a live case" and "updates its resolution plan mid-conversation without waiting for an explicit new prompt." Salesforce says the feature is already live with named customers, including the education software company PowerSchool with more than 550 users.

The rest of the release leans the same way, toward agents that complete whole tasks rather than help with steps. Slack Code lets people build software with AI agents inside shared Slack channels. Agentforce Contact Center adds voice and is now available in more than 30 countries. There are more than 100 prebuilt agent skills, and a third-party orchestration layer that connects Agentforce to agents running on AWS, Azure and Google. One insurance example has a voice agent taking First Notice of Loss claims across six business lines with no human representative in the loop. If you want the plain-English version of what an agent is and is not, we wrote a guide to AI agents and what they can actually do.

The numbers are Salesforce's own

Salesforce is doing something it mostly avoided in earlier agent launches. It is attaching figures to named customers. Its commerce agents are credited with a 13 percent conversion lift and a 17 percent increase in add-to-cart rates. Autonomous Scheduling with Voice, it says, turns a 15-minute booking call into seconds.

Those are real claims worth noting, and they are also the company's own marketing. TechTimes put it bluntly: "These are Salesforce-reported figures, not independently audited third-party benchmarks." The same report, citing survey data around the launch, found that only about 30 percent of organizations have fully deployed agents in production, and that 32 percent discovered agents operating outside their intended parameters only after a consequential error had already happened. That is the gap between a launch deck and a Tuesday afternoon.

What default-on actually asks of IT

Put the two halves together and the story for an IT or operations team is specific. Agents that can take actions on customer data are being switched on by default, and the switch that used to control them is going away. That is a reason to decide, deliberately, how much autonomy you want before the decision is made for you.

The governance surface is not hypothetical. Agentforce has already been the subject of a prompt-injection vulnerability that researchers called SalesBleed, the kind of flaw where hidden instructions in ordinary-looking data trick an agent into doing something it should not. We covered how the SalesBleed prompt-injection flaw worked, and the lesson travels. The more an agent is on by default and acting on its own, the more a single bad instruction can cost.

Salesforce has kept the controls, and they are worth knowing. You can disable Agentforce entirely by turning off the Einstein setting on the Einstein Setup page. You can activate or deactivate individual agents in Agentforce Builder. And you can gate who gets to run agents with the Manage AI Agents permission. None of that is new, but with default-on it moves from nice-to-have to a thing to configure this week.

There is also a dimension the feature list does not dwell on. The cost savings Salesforce points to come, in part, because agents are doing work people used to do. TechTimes noted that this "carries a workforce dimension the feature announcements do not foreground." A booking call that now takes seconds is a booking call someone used to make.

Our take

Winter '27 is a genuine product release with a genuine policy change buried inside it, and the policy change is the headline for anyone responsible for a Salesforce org. Default-on Agentforce is not a scandal. It is a default, and defaults are how most software actually gets used. But it rewards attention. Turn the agents on because you decided to, with the Einstein setting, the Builder controls and the Manage AI Agents permission set the way you want, rather than because a release flipped them for you.

On the autonomy claims, treat them the way the evidence warrants. The capability is real and shipping, the named-customer figures are plausible, and they are still unaudited vendor numbers against a backdrop where most companies have not fully deployed agents and a third have already been surprised by one. The agents can run the workflow. Whether you have governed them well enough to let them is the part Winter '27 leaves to you.