What Claudeforce can't do yet

Every capability in the previous guide is real and sourced. This one is about what isn’t yet, because a rollout decision made on the announcement alone is missing information a careful admin or architect would want first.

---

Cost isn’t quantified

Salesforce in Claude’s usage ties to API calls billed separately through an Anthropic contract. Coverage of the announcement is consistent on one point: consumption is “certainly not zero,” and nobody’s published a number that lets you estimate what a pilot, let alone a full rollout, actually costs. If a vendor conversation gives you a confident total cost of ownership figure this early, ask what it’s based on, since the public information doesn’t yet support one.

The honest move here is instrumenting a small pilot cohort and measuring your own actual usage before committing to a budget built on someone else’s guess.

---

The beta window collides with a Salesforce release cycle

Salesforce in Claude’s open beta is expected in September 2026, the same window as several scheduled Winter ‘27 production release upgrades. Testing a new AI integration and absorbing a platform release at the same time means it’s harder to tell which change caused which behavior if something breaks. Sequencing your pilot around, rather than directly inside, a major release window is a reasonable precaution, not excess caution.

---

Vendor concentration is a named concern, not a hypothetical one

Architects covering this partnership have flagged a real structural point: Claudeforce deepens mutual dependency between Salesforce and Anthropic specifically. If your organization’s AI strategy assumes model flexibility, the ability to switch reasoning models without re-architecting how your CRM data gets exposed to them, that assumption is worth revisiting once a large part of your workflow is built specifically around this one integration.

This isn’t a reason to avoid it. It’s a reason to keep an explicit escape hatch: know what re-platforming to a different model would actually require before you’re several skills deep and find out the hard way.

---

The skill list is 91% unconfirmed

Salesforce in Claude ships with 37 prebuilt sales skills. Three are named anywhere in public sources: meeting prep, deal health review, and pipeline review. If a specific skill matters to your use case, whether it exists yet is an open question, not a safe assumption, until you can confirm it directly with your account team or see it in the pilot itself.

---

“Governed” doesn’t mean governed the way you’d assume

Salesforce’s messaging emphasizes that actions are governed, meaning they route through existing permissions and business logic. That’s accurate and also easy to over-read. Governed means it respects the rules you’ve already set, not that it makes good judgment calls where those rules are silent. An overly broad permission set is still an overly broad permission set once an agent is the one using it. See the admin pilot checklist for what that actually requires you to check.

---

Vendor productivity claims aren’t independent proof

Public discussion of Claudeforce has cited a large productivity figure, millions of hours of annualized gains, drawn from Salesforce’s internal use of Slackbot, a related but different product. It’s a real, sourced number, and it’s also self-reported by the vendor about a different tool, not an independently verified result for Claudeforce specifically. Useful as directional context. Not a substitute for measuring your own pilot.

---

When treating this guide as complete would be a mistake

Reading this once, months from now. Beta products change. A limitation named here in September 2026 may be resolved by the time you’re evaluating a rollout. Check the current official documentation before making a final decision, and treat this guide as a snapshot of what was known at the time it was written, not a permanent verdict.

Assuming “unconfirmed” means “false.” The other 34 skills likely exist and likely work reasonably well. Unconfirmed means exactly that: not yet something you can verify from a source outside the vendor’s own marketing. That’s a reason to check before you build a workflow around a specific one, not a reason to assume it doesn’t exist.

---

What to try next

Come back to this list after your own pilot. The gap between what you find here and what you actually observe running the thing is more useful than either one alone.