Salesforce in Claude, and what it actually changes

What actually shipped

On 15 September 2026, Anthropic released Salesforce in Claude in beta. It is a plugin, built with Salesforce, that brings a seller’s accounts, opportunities and pipeline into Claude under their existing Salesforce permissions.

It follows Claudeforce, the expanded partnership the two companies announced on 26 August 2026, which had put Salesforce in Claude in front of a small set of pilot customers. The September release opened it to any organisation Salesforce approves through beta signup.

The framing worth holding onto is that this is not a new CRM, and not an autonomous agent working your pipeline overnight. It is Claude, reading and writing your Salesforce records with your permissions, with a set of prebuilt workflows on top.

What the 37 skills cover

The 37 skills map to the work an account executive repeats every week rather than to anything exotic:

• Daily briefings. What is on the calendar, what changed on those accounts since last contact.
• Account research and call preparation. Assembling history, recent conversations and open opportunities into one brief.
• Deal scoring and close planning. Grading open opportunities and setting out what has to happen next.
• Post-call CRM updates. Turning what was said into the fields and next steps the record needs.
• Pipeline dashboards and forecasts. Coverage, movement and the stage-level view leadership asks for.

Anthropic has indicated more skills will follow, so 37 is an opening set rather than a ceiling. The number is also the least interesting part of the release. A skill here is a packaged prompt and workflow over your own data, which means the quality of any one of them depends far more on the records underneath than on the skill itself.

Who sees what, and what Claude can change

This is the part that decides whether a revenue leader can say yes, and the design is conservative in the right places.

Salesforce remains the system of record. Sellers sign in with their own Salesforce credentials, and Claude reads only what those permissions already allow, so it cannot surface a record the seller could not open themselves. By default, Claude asks the seller to approve each proposed change before it is written. On Team and Enterprise plans, the data is not used for model training.

Two things follow from that. Your existing Salesforce sharing model is the security model, which means any over-broad profile you have been meaning to tidy is now over-broad in a much more accessible interface. And the approval step is a default rather than a law of physics, so the question of who is allowed to relax it belongs in your rollout plan rather than in a later incident review.

What it takes to turn on

• A paid Claude plan. The beta is available across all paid tiers.
• The latest Sales Cloud enterprise edition. This is the gate most teams will hit first.
• A Salesforce admin. Access is requested through AgentExchange, then connected once for the whole organisation.
• An OAuth setup step. The Salesforce admin creates an External Client App and passes the consumer key and secret into Claude’s connector settings. The Help Center guide sets out the sequence.

None of that is difficult, but it is an admin project rather than a per-rep toggle, and it needs a Salesforce owner who will schedule it.

The part that is genuinely valuable

The strongest case for this is unglamorous : the assembly work before and after a customer meeting is real, recurring, and almost entirely undifferentiated.

A rep preparing for a call opens the account, scrolls the activity history, finds the last two call notes, checks which opportunities are open and at what stage, and then goes looking for the Slack thread where someone mentioned the procurement delay. That is fifteen to twenty minutes per meeting that produces nothing a customer ever sees. Multiply it across a full calendar and it is the single largest recoverable block of time in an AE’s week.

The same applies in reverse. The reason CRM hygiene decays is not that reps are lazy, it is that updating the record happens after the value has already been captured, when the next call is starting. Moving that write to a reviewed, one-click step is a structural fix to a behavioural problem, which is a better class of solution than another reminder.

This is consistent with the pattern across this category. As we set out in AI is a layer, not another tool, the durable gains come from AI running across the CRM, the data provider and the sequencer rather than from a standalone product bought alongside them. Salesforce in Claude is that layer, applied to the CRM.

What it inherits from your CRM

Here is the limit, and it is the same limit every tool in this category runs into.

A pipeline dashboard generated from your Salesforce is only as accurate as your stage definitions. If “Discovery” means a real conversation for one rep and a connected call for another, a faster dashboard reports the same inconsistency with more confidence and better formatting. Deal scoring has the same dependency: a score built on fields nobody fills consistently inherits that inconsistency.

We have made this argument with numbers elsewhere. Moving MQL to SQL conversion at Locus took no additional lead volume and no new tooling; it took a written definition of qualified, enforced at stage exit. The same held on named client work. The constraint was never how quickly someone could assemble the information, it was whether two reps would grade the same lead identically, which is the subject of the sales qualification framework.

That is not an argument against the plugin. It is an argument about sequence. The published research on AI in outbound points the same way: the tools compress research, drafting, triage and admin, and leave targeting, message quality and qualification untouched. We collected what could be traced back to a named source in the 2026 AI outbound benchmarks.

Who should turn it on now

Turn it on now if you are running Sales Cloud enterprise edition, your stage definitions are enforced rather than aspirational, and your AEs carry enough meetings that prep time is a real line in their week. The return is immediate and it recurs.

Wait if your CRM is currently a reporting fiction. Every hour you spend rolling this out before the definitions are fixed buys you a faster route to the same unreliable answer. Start with a pipeline audit and a written qualification standard, then come back to it.

It changes nothing about which CRM you should be on. Salesforce and HubSpot were each implemented in the SDR and revenue operations teams at Locus, GoComet and Landmark Group. If that question is still open, the trade-offs are in HubSpot versus Salesforce, and this release is not on its own a reason to move. HubSpot teams have a comparable release in the HubSpot Prospecting Agent. The connectors for HubSpot, Apollo and Outreach, and Anthropic’s Sales plugin, are covered in Claude for sales teams.

One prediction worth making. The teams that get the most from this will be the ones that treat the approval step as a control worth keeping rather than friction worth removing, because the value of a CRM everyone trusts is higher than the value of a CRM that updates itself.

References

• Salesforce in Claude. Anthropic, 15 September 2026. Beta release, 37 prebuilt skills, permissions model, approval-before-write default, availability on paid plans.
• Set up Salesforce in Claude for your organization. Claude Help Center. AgentExchange request, External Client App and OAuth sequence, per-member Salesforce sign-in.
• Sales Cloud in Claude. Salesforce AgentExchange listing.
• Salesforce and Anthropic Announce Claudeforce. Salesforce, 26 August 2026. The partnership this release sits inside.

Related reading

The wider argument this sits inside is AI is a layer, not another tool, and the tool-by-tool view is the best AI tools for sales development. The stack-level decision is the B2B sales tool stack, chosen by workflow. Setting the CRM up so it is worth summarising is HubSpot and Salesforce setup.

The discipline this depends on is the sales qualification framework and the rate it is judged on is MQL to SQL conversion. The reporting layer it feeds is GTM analytics and sales reporting, and the fastest way to find out whether your stages are trustworthy is a pipeline audit.