Salesforce AIforce: Your CRM Is Becoming a Data Layer for AI Assistants

If your sales team lives in Slack and your marketers draft in an AI assistant, the CRM is probably the last place anyone updates. On September 15, 2026, at Dreamforce in San Francisco, Salesforce announced a product built around that habit. AIforce is meant to let people read, change and act on Salesforce data from wherever they already work, without opening Salesforce at all.

What Salesforce announced

Salesforce describes AIforce as a live interface layer that carries the data, workflows, business logic, permissions and governance held inside Salesforce to any AI interface. At launch that means three surfaces, according to the company's announcement:

• Salesforce in Claude, part of what Salesforce calls Claudeforce: a prebuilt MCP (Model Context Protocol) server inside Anthropic's Claude, launching with 37 prebuilt sales skills that Salesforce says cover prospecting through pipeline hygiene. It is in beta for all customers after pilots with Deloitte, GitLab and Legora.
• Slackforce: Salesforce context inside Slack, including Slack CRM for creating accounts, logging notes from calls and updating records by prompt, and Slackforce Surfaces for live, interactive views.
• Agentforce Coworker: an assistant inside Lightning, Salesforce's own interface, which reasons across accounts, activity and history. Salesforce says 100,000 users activated it within 35 days of launch.

The release also says AIforce is powered by a Headless Toolkit that exposes the platform through MCPs, APIs, plug-ins, skills and developer tools, so teams can build their own AI experiences on the same data. In a press briefing before the announcement, reported by Salesforce Ben, Patrick Stokes, Salesforce's President, Applications and Marketing, put the reasoning plainly:

The CRM becomes something assistants query

Read past the product names and the change is architectural. For most teams the CRM has been both the database and the screen. AIforce separates the two. The record, the permission and the workflow stay in Salesforce, while the question and the action can arrive from Claude, from Slack or from Lightning. Claude is one surface among several here, and the same layer is meant to serve whichever interface a company connects.

That has a practical consequence. When a salesperson opens a record, they see its gaps: the empty industry field, the duplicate contact, the opportunity that has sat in one stage for months. When an assistant answers a question such as which accounts are ready for a renewal conversation, those gaps disappear inside a confident paragraph. The answer can be no better than the quality of the records, and the person reading it has fewer cues that something is wrong.

The governance model shifts too. Salesforce says every AIforce request runs on the customer's existing permissions and business rules, so an agent sees only what the person asking can see, and every action routes back through Salesforce. It also says business data used to answer a question is not retained by the model provider, which it calls Zero Data Retention. That is a sensible design, but it moves the weight onto permission sets that many organisations have not reviewed since they were first configured. An assistant that can reach everything a user can reach will find the records that user was never meant to have.

What to prepare this quarter

None of this requires a purchase decision today. It does reward work that belongs in any AI implementation plan, and doing it before a pilot starts is cheaper than doing it after a pilot surfaces the problems.

Data hygiene

Start with the fields an assistant is most likely to be asked about: lifecycle stage, lead source, owner, industry, country, last activity and opportunity stage. Check how many records have them filled, whether picklists are used consistently, and how duplicates are merged. Agree definitions in writing. If marketing and sales mean different things by a qualified lead, an assistant will pick one meaning and state it as fact.

Permissions

Audit profiles and permission sets with the question an assistant raises: if this user asked for everything they can see, would that be acceptable? Pay particular attention to shared admin accounts, integration users, agency and contractor logins, and anyone who changed role without losing old access. Because Salesforce says AIforce inherits these rules rather than adding a new permissions model, the audit is the control.

Which workflows to pilot

Choose work that is frequent, low-risk and easy to check. Reasonable candidates:

• Pre-meeting account summaries, where the rep can verify the answer against the record.
• Pipeline hygiene prompts, such as finding opportunities with no next step or an overdue close date, which matches the sales skills Salesforce says it launched with.
• Logging calls and notes from Slack, where the gain is that the update actually happens.
• Read-only questions for marketing, such as which campaigns influenced a set of open opportunities, before anyone is allowed to write.

Leave bulk updates, pricing, contract changes and anything that emails a customer until you have seen how the assistant behaves on the easy cases.

Governance

Decide who owns the pilot, which users join it, what the assistant may change, and how changes made through a prompt are reviewed. Confirm the Zero Data Retention terms and data residency in your own contract rather than relying on a keynote. Salesforce's release says availability may vary by region and is governed by customer agreements. For teams in the EU and UK, bring the data protection lead in early, because CRM records are personal data.

What remains unannounced

Several things a buyer needs were not in the announcement. Salesforce's release says pricing and packaging are subject to change and that customers should base purchasing decisions on products currently available. A live blog of the keynote by The Letter Two reported Patrick Stokes saying on stage that the Sales Cloud skill for Claude was available in open beta for free from that day. The release itself describes Salesforce in Claude as in beta but states no price, so treat free access as reported rather than confirmed. Apart from Agentforce Coworker, which Salesforce says is available to customers now, no general availability dates were given.

The marketing and commerce pieces were thin. Salesforce says Salesforce in Claude will add skills for service, marketing, commerce and industries later, and the same live blog listed a Commerce Cloud shopper agent and agentic commerce search among the keynote updates, and, with a question mark, a marketing campaign agent, without further detail. Until those are documented, plan around sales use cases and treat marketing work through an assistant as a roadmap item.

Finally, only three surfaces launch now. Salesforce says more will be announced soon, and its release lists Anthropic, AWS, Google and Microsoft interfaces among AgentExchange partner integrations, but does not name ChatGPT, Gemini or Microsoft Teams as AIforce surfaces. If your organisation standardises on another assistant, the Headless Toolkit may be the route, but whether it works as well as the prebuilt integrations is something to test, not assume.