Salesforce in Claude: What Beta Access Actually Gets You

Legora's sellers now turn live Salesforce data into meeting briefings in seconds instead of hours, according to the legal AI company's own CFO. That line, not a roadmap slide, is what Anthropic and Salesforce are pointing to as the actual payoff of Salesforce in Claude, a plugin that moved from a small pilot into open beta on September 15. Roughly 7,000 sellers are already running it in production, with GitLab, Siemens, and Legora named as early customers.

Salesforce in Claude pulls a seller's accounts, opportunities, and pipeline into a Claude conversation, under the same permissions that seller already has inside Salesforce. Connecting a large language model to a CRM isn't new by itself, Salesforce has published an API for that since long before anyone called it agentic AI. What's new here is the second half of the announcement: the plugin ships with 37 pre-built skills for the specific work an account executive does every day, instead of a blank chat window pointed at a data source and left for the seller to figure out.

How Salesforce in Claude actually connects

The setup is admin-first. An organization connects Salesforce to Claude once, and individual sellers then sign in with their own existing credentials rather than getting a separate login to manage. Claude reads only the data that seller's permissions already allow, so a rep who can't see another territory's pipeline in Salesforce can't see it through Claude either. Anthropic says Team and Enterprise plans don't use customer data to train models, and by default any change Claude wants to make, updating an opportunity, logging a call, creating a follow-up task, stops for the seller's approval before it writes back to Salesforce. That default can be adjusted on Team and Enterprise plans, so it is worth confirming how your own org has it set rather than assuming every write is gated.

That approval step is worth dwelling on, because it's the difference between an assistant and an autonomous agent making changes to a system of record. Salesforce in Claude reads freely and drafts freely, but out of the box every write action stops at a confirmation, which is the same pattern showing up across most enterprise AI rollouts this year: broad read access, gated write access.

The 37 skills, and what they replace

The skills cluster around the tasks that eat an account executive's day outside of actually talking to customers. Pre-call preparation pulls account history, recent Slack mentions, and past emails into one briefing instead of a rep manually checking four tools before a meeting. Pipeline review and forecast narratives turn raw opportunity data into the kind of write-up a sales manager expects, with slip-risk flagged rather than buried in a spreadsheet. Deal scoring runs an opportunity against the team's own methodology and proposes a close plan. Post-meeting, a skill takes a call transcript and turns it into a Salesforce update and a follow-up email, the two things that reliably get delayed until the next morning, if they happen at all.

None of these are technically difficult tasks. What made them slow before was that the information lived in four or five different places, Salesforce, email, Slack, call recordings, and someone had to manually stitch it together every single time. That stitching work is exactly what the 37 skills are built to remove.

Why 37 pre-built skills matters more than the integration

A generic chat interface connected to a CRM has been technically available for a while, and it has mostly underwhelmed. The hard part was never wiring up the data pipe, it was knowing what a good pipeline review or a good call-prep brief actually looks like for a specific sales motion. Anthropic shipping opinionated, pre-written skills instead of a general-purpose agent is a concession that specificity beats flexibility for this kind of work. A rep doesn't want to prompt-engineer a forecast narrative from scratch every Friday. They want a skill that already knows the shape of a good one and asks for the numbers.

This mirrors a pattern showing up across other AI tooling right now: bare model access is rarely the bottleneck anymore, the packaging around it is. A tool that ships with the workflow already encoded tends to get adopted faster than one that hands a user a blank canvas and calls it flexibility.

What this means if you're evaluating your own AI stack

If a sales team inside your company has a "we need a custom AI build for CRM prep" project sitting in a backlog, this beta is worth checking against it directly. The pitch isn't a novel capability, transcription-to-CRM-update and pipeline dashboards have existed as point solutions for a while, it's that the capability now ships as a subscription feature inside a tool sellers might already have open, rather than a bespoke internal build someone has to maintain. For a smaller team without a dedicated ops function, that's the more realistic path to the same outcome.

What's still unclear

The beta launched to a specific customer set rather than a general rollout, and Anthropic hasn't published standalone pricing for Salesforce in Claude separate from existing Claude plans. There's also no independent, third-party review yet of how the 37 skills perform against a messy real-world pipeline versus the clean demo scenarios in the announcement, adoption numbers from Anthropic and Salesforce measure usage, not necessarily quality. Anyone evaluating this for their own team should treat the beta status as real: expect rough edges before wider general availability, and confirm what admin setup and data residency terms actually apply to your org before connecting a production Salesforce instance.

Sources: Anthropic, "Salesforce in Claude" and Unite.AI, "Anthropic Releases Salesforce in Claude Plugin With 37 Sales Skills".