Salesforce in Claude puts your sales process on trial

On 15 September, Anthropic released Salesforce in Claude in beta. It’s a plugin, built with Salesforce, that brings a seller’s accounts, opportunities and pipeline into Claude. It ships with 37 prebuilt sales skills covering account research, call prep, pipeline review and CRM updates.

The practical change is where a seller works. A rep can ask Claude for the morning pipeline review, a brief for the 2pm meeting and a draft update to the opportunity after the call. Salesforce still holds the data, the permissions, the business rules and the approvals.

That split matters for anyone running a sales organisation. Every Salesforce customer on Sales Cloud Enterprise can now apply for the same model and the same 37 skills, so the model stops being an edge. What Claude finds when it reads your CRM becomes the edge, or the liability. This piece covers what shipped, how the controls work, and what to fix before you turn it on.

What Salesforce in Claude ships

Salesforce in Claude is the first Claudeforce product to ship. Claudeforce is the partnership the two companies announced on 26 August alongside Salesforce’s Q2 FY27 results. The rest of Claudeforce puts Claude inside Salesforce: it becomes the reasoning model behind Agentforce’s Atlas Reasoning Engine and the default model in Slackbot.

The plugin runs in the other direction. According to Anthropic’s launch post and Unite.AI’s report on the beta:

• It’s in beta on all paid Claude plans, on web and desktop, in chat and in Claude Cowork.
• Organisations join through a beta sign-up that Salesforce approves. Unite.AI reports it requires Sales Cloud Enterprise Edition.
• Two connectors do the work. The Salesforce connector reads records and makes updates. The Slack connector handles deal-channel summaries and account-team threads.
• The skills include a morning brief (the day’s meetings, deals closing soon, at-risk opportunities, unread threads), call prep, deal scoring against your team’s methodology, close plans, post-call summaries that become follow-up emails and opportunity updates, forecast narratives, renewal prep, QBR decks and interactive pipeline dashboards.
• Anthropic says GitLab, Siemens and Legora have deployed it, and that 7,000 Salesforce sellers use it.

Legora’s CFO, David Eckstein, described the gain in one line: “our sellers turn live data into meeting briefings in seconds instead of hours.”

Claude does the work, Salesforce keeps the rules

The design choice worth understanding is the division of responsibility. Claude reads, reasons and drafts. Salesforce decides what is allowed.

Sellers sign in with their own Salesforce credentials, so there is no second permissions model to maintain. A write from Claude is an ordinary Salesforce write. As Apex Hours’ technical breakdown points out, your validation rules, Flows, Apex triggers and governor limits all fire exactly as they do today.

Patrick Stokes, Salesforce’s President of Applications and Marketing, set out the commercial logic to CIO: “When people stop using Salesforce through the traditional human interface and start using it through an agentic interface, it dramatically increases the value of Salesforce.”

Marc Benioff went further on the Q2 earnings call that day. “This nonsense of the SaaSpocalypse, I think it’s time for it to stop,” he said, as reported by Salesforce Ben.

Read both quotes from the buyer’s side. Salesforce is betting the CRM gets more valuable as the screen matters less. That bet only pays off for customers whose data and rules are worth reasoning over.

Everyone gets the same 37 skills

Every competitor on Sales Cloud Enterprise can apply for this. They get the same model, the same connectors and the same skills. Any edge from a well-designed page layout or a team that’s quick at clicking through screens shrinks once the screens go away.

What stays different is the material Claude works from. Five things decide whether the output is useful or confidently wrong.

Data that reflects reality

A pipeline review is only as good as the close dates, amounts and next steps in the records. If reps update opportunities on Friday afternoon from memory, Claude will write a fluent forecast narrative from Friday-afternoon memory.

A methodology someone wrote down

Deal scoring runs against your team’s methodology. Claude has to be able to find it. If your qualification framework lives in a sales kick-off deck from two years ago and in the heads of three senior account executives, the scoring skill has nothing firm to test deals against.

Written exit criteria for each stage, and a clear definition of what “commit” means in your forecast, feed straight into better output. Anthropic’s help centre says a setup skill runs the first time a seller opens the plugin, learns their role and book of business, and personalises the other skills to them. Neither company has yet documented where deal scoring finds your methodology. Admin guides such as HyphenX’s setup guide say it goes in through Claude’s instructions. Wherever it ends up, someone has to write it down first.

Stage definitions that mean one thing

In a lot of organisations, “Stage 4” means verbal agreement in one region and procurement started in another. A human sales leader knows to adjust. Claude’s slip-risk analysis will treat both as the same thing and average across them.

Permissions you’d be happy to see used

Claude reads whatever the signed-in seller can see. Profiles and sharing rules tend to widen over the years, so reps often have access to accounts, notes and fields they never open. A person ignores them. An assistant asked to summarise everything on an account won’t.

Audit sharing rules and field-level security for the pilot group before rollout.

Automation built for one click at a time

Many Flows and triggers were written assuming a person edits one field on one record. A post-call summary that updates the stage, rewrites the next step and adds three contacts in one pass can fire approval Flows, send notifications nobody expected, or hit governor limits. Test it in a sandbox first.

We’ve seen the same thing outside sales. We’re building an invoicing and payroll system for a staffing firm, and the hard part hasn’t been the software. It’s been the rules that only one person knew. Their finance lead knew a new pay group needed setting up whenever a letter A appeared at the bottom of her spreadsheet. Staff are never paid for under four hours, except on one patrol contract that bills a flat fee for a two-hour round. None of it was written down anywhere a system could read.

So the first weekly runs flagged those cases as exceptions, and she cleared them by hand. On one of those calls we put it simply: “this shouldn’t be an exception, this should be a rule.” Each rule we write down becomes something the system applies on its own, and one less thing she checks by hand every week. Your sales process works the same way. Whatever lives only in your best account executive’s head, Claude can’t use.

What to do before you switch it on

A four-week pilot is enough to learn whether your organisation is ready. This is the sequence we’d run.

1. Week 1. Pick one team and two read-heavy skills, such as the morning brief and call prep. Leave CRM updates off. Apex Hours recommends starting with a permission set scoped to specific objects and fields, read-only, then adding write access skill by skill.
2. Weeks 1 and 2. Write down the sales process in a form Claude can use: stage exit criteria, the qualification framework and forecast categories. This is usually the most useful document a sales team produces all year, with or without AI.
3. Week 2. Audit permissions for the pilot group. Run Claude-style multi-field updates against a sandbox and watch which Flows and triggers fire.
4. Week 3. Turn on post-call updates with seller approval left on. Track how often sellers edit Claude’s proposed changes before approving them. A high edit rate on a particular field tells you exactly where your process is unclear.
5. Week 4. Compare against the previous month: time spent on CRM admin, opportunity field completeness, and how far forecast calls moved in the final week. Then decide scope.

Keep the approval default on for longer than feels necessary. Every approval is a governed checkpoint, and the edit data it produces is the cheapest process diagnostic you will get.

What is still unclear

A beta launch leaves open questions. These are the ones we’d put to Salesforce and Anthropic before committing budget.

• Pricing. Neither company has published it for the plugin. Apex Hours describes the underlying model as consumption-based and tied to API calls. Model the cost before the pilot grows.
• Full skill list. Anthropic’s announcement describes the skill areas without naming all 37. Salesforce has said more prebuilt skills are coming. Check the setup guide for what your org actually receives.
• Data handling. Claude inside Agentforce runs on Amazon Bedrock within the Salesforce Trust Boundary. The plugin brings Salesforce data into Claude instead, so your Claude plan’s terms apply. Anthropic says it doesn’t train on customer data by default on Team and Enterprise plans. Regulated firms should still run their own legal and security review of that route.
• Stability. Beta terms mean skills, limits and edition requirements can change.

Where Exponential Partners fits

Salesforce in Claude will make well-run sales teams faster. It will also expose a messy CRM quickly, because every gap in the data now shows up in a brief someone reads before a client meeting.

That gap is where we work. We map how deals actually move through your organisation, tighten the process and data Claude will read, set up the pilot, and train your sellers to use it inside their own pipeline. Where the process needs tooling that Salesforce doesn’t provide, our build and automation work covers that too.

If you’re weighing a pilot and aren’t sure your CRM is ready, we’ll give you a straight answer. Schedule a Conversation.