Two things happened in the same stretch of weeks.
Salesforce will auto-enable Agentforce in Winter ’27 for every org that already has access. The rollout started in the first week of September. Later in the release, the on/off toggle comes off Setup. Salesforce says there is no extra cost. Admins can still turn Einstein off, or deactivate an agent.
Then Salesforce and Anthropic announced Claudeforce. The first product is Salesforce in Claude: a plugin with 37 prebuilt sales skills, live CRM context, and governed actions from inside Claude. It is in pilot now, with open beta this month. Claude is also the default reasoning model inside Agentforce and Slack.
The names can be confusing, so to be clear:
Now we can talk about what actually changed.
A toggle, a plugin and 37 skills do not change how a brand sells, serves or decides. They change how easy it is to look busy. The hard part is still process, change and governance: which work an agent is allowed to finish, who works differently on Monday, and who is accountable when an agent sends, scores or spends.
Salesforce is not ignoring that. The product has permissions, Einstein controls, and “governed action.” The gap is organisational. If Einstein is already on, auto-enable puts Agentforce Builder in front of admins. That is not the same as a sales team using it, a service team trusting it, or a marketer letting it act on live campaigns.
We have watched this pattern on platforms for years. Paid for, switched on, still not earning their keep. The gap is rarely the platform. It is the layers around it. (We set that out here: 10 signs your digital platform is underdelivering.)
Claudeforce makes that sharper, not easier. If the CRM interface is no longer the only place work happens, data quality, permissions and the operating model matter more, not less. An agent that can update pipeline from Claude will do it with whatever records it can see. If nobody owns those records, the agent does not create a data problem. It industrialises one.
So what does the next 90 days look like if the agents are already on?
Find out what auto-enable actually switched on. Which Agentforce, Claude and Slack surfaces people are already using. Name the two or three workflows an agent is allowed to finish. Be explicit about who owns data quality, permissions and the commercial outcome.
Write the policy: what can act without a human, what stays in draft until someone publishes. Choose a tool architecture. Agentforce, Salesforce in Claude and Slack are not all required on day one. Write the risk appetite down so the board can stand behind it.
Take one priority workflow and map it end to end: data, permissions, checkpoints, human approvals. Train the people who will work differently on Monday against that workflow, not a product demo. Measure completed governed work, not “agents enabled.”
That is not a software project. It is the same digital transformation work brands have needed around CRM, content and experience for decades. The agents just made it urgent.
If the agents are on and the organisation is not, maybe this is the 90 days you need: AI Foundations.
