Claudeforce: Inside the Salesforce and Anthropic Deal

What Claudeforce is, in one sentence

On August 26, 2026, Salesforce and Anthropic announced Claudeforce: the name they gave to the deepest stage of a partnership that was already two years old. It is not a product you buy separately, and that is worth saying up front, because the name sounds like one. It is the umbrella over a set of integrations between Claude, Anthropic's model, and the Salesforce platform.

What changed with this announcement is the direction of travel. Until now, Claude went into Salesforce: it was one of the models you could pick to work inside the CRM. With Claudeforce, Salesforce also goes into Claude. Your sales information shows up inside Claude's interface, and from there you can trigger actions that get written back to the CRM.

Marc Benioff, Chair and CEO of Salesforce, summed up the bet in the press release with a line worth quoting: "For the first time, the interface is the AI." Without the marketing: the screen a salesperson works in stops being a list of CRM tabs and becomes a conversation.

Four announcements in two years

Claudeforce did not come out of nowhere. It is the fourth announcement in a sequence, and reading them in order makes it much easier to separate what is new from what was already there.

• September 3, 2024
  AnnouncementClaude available to Salesforce customers
  What it actually enabledYou could select Claude 3.5 Sonnet, Claude 3 Opus and Claude 3 Haiku from Einstein Studio, through the bring-your-own-model feature, served via Amazon Bedrock
• October 14, 2025
  AnnouncementExpanded partnership for regulated industries
  What it actually enabledClaude becomes a preferred model in Agentforce, and Anthropic becomes the first model provider fully integrated inside the Salesforce trust boundary
• February 2, 2026
  AnnouncementSalesforce context and actions inside Claude
  What it actually enabledClaude reaches Slack conversations and triggers Agentforce 360 actions through the Model Context Protocol
• August 26, 2026
  AnnouncementClaudeforce
  What it actually enabledThe Salesforce in Claude add-on with 37 sales skills, and Claude as the default model in Slack

Sources: announcements published by Anthropic on September 3, 2024 and October 14, 2025, and by Salesforce on February 2 and August 26, 2026.

The most useful way to read that sequence: in 2024 Claude was a configuration option; in 2025 it became the recommended option for the most demanding sectors; in 2026 it also became the place where the user works.

What Claudeforce includes today

The announcement names three new pieces and confirms one that already existed. They are worth separating, because their availability is not the same.

• Salesforce in Claude
  What it doesAdd-on that brings 37 sales skills and an AI-generated panel carrying your company context. Authentication is set once at the org level, not user by user
  StatusBeta with pilot customers. Open beta announced for September 2026
• Claude Tag in Slack
  What it doesClaude takes part inside team conversations, in the channel where the topic is already being discussed
  StatusAvailable
• Slack Code
  What it doesTeams and agents write code together in the same Slack channel
  StatusAnnounced
• Claude inside Agentforce
  What it doesClaude as a model behind the Atlas reasoning engine, served through Amazon Bedrock
  StatusAvailable

Source: Salesforce press release and product page, August 26, 2026.

Salesforce also published the roadmap, and it is long: Service, Marketing, Commerce, Revenue, Field Service, Tableau, MuleSoft, Informatica, Data 360, Headless 360 and Industries are all listed as "coming soon", with no dates. Which means that today Claudeforce is, in practice, a sales tool. Everything else is announced, not shipped, and that difference matters when you are building a budget.

The 37 skills, explained

The number that traveled furthest in the coverage was 37. Before getting excited, it helps to know what a "skill" means here: it is a preconfigured task, not a new CRM feature. Someone already wrote the instructions, defined which data to read and what shape the answer should take, so the user invokes it instead of explaining everything from scratch.

Salesforce grouped them into four categories.

• Start of day
  Problem it solvesKnowing where your book of business stands before the first call
  Published examplesPipeline review, daily briefing, forecast narrative
• Deal strategy
  Problem it solvesBringing order to one negotiation already in motion
  Published examplesStakeholder mapping, close plan, objection handling
• Growth
  Problem it solvesDeciding where to spend selling effort
  Published examplesAccount planning, prospect evaluation, customer health
• Prep and action
  Problem it solvesThe work that happens before and after every meeting
  Published examplesCall prep, meeting scheduling, loss review, data hygiene

Source: Salesforce Claudeforce product page, retrieved September 7, 2026.

Looking at the list, most of these are reading and writing tasks over data that already sits in the CRM. That is good news and a warning at the same time. The good news is that they do not require redesigning processes. The warning is that their quality depends entirely on the quality of the data you have loaded.

The trust boundary: the part that actually matters

Of everything that was announced, the technical point with the most consequences got the fewest headlines. It appears in the October 2025 announcement and holds through the later ones: Anthropic is the first large language model provider fully integrated inside the Salesforce trust boundary, with all traffic contained within Salesforce's virtual private cloud.

In plain terms: when an Agentforce agent uses Claude, your data does not leave the perimeter Salesforce administers to be processed on Anthropic's own infrastructure. The model runs on Amazon Bedrock, inside that perimeter. Salesforce's developer documentation states this explicitly for Anthropic and Amazon models.

That is why the 2025 announcement targeted four specific sectors: financial services, healthcare, cybersecurity and life sciences. These are the ones that cannot move customer data to a third party without substantial legal work. The two customers Salesforce and Anthropic named at the time fit the pattern: CrowdStrike in cybersecurity, and RBC Wealth Management in wealth management. Rohit Gupta of RBC Wealth Management said in the release that the deployment saved advisors significant time.

One necessary caveat. Traffic staying inside the Salesforce perimeter is not the same thing as automatic regulatory compliance. Your company is still accountable for how the data is handled, and in Latin America that means reviewing local rules in every country where you operate. Zarasa does not provide legal advice: this one is settled with your legal team.

Which Claude models you can use today

Here is a concrete, verifiable detail that barely shows up in the coverage: Salesforce's developer documentation lists eight Claude versions available in Agentforce, across the model's three families.

• Claude Haiku
  Versions available4.5
  When it fitsThe fastest and cheapest. High volume of simple queries
• Claude Sonnet
  Versions available4.5, 4.6 and 5
  When it fitsThe balance between cost and capability. The workhorse for most agents
• Claude Opus
  Versions available4.5, 4.6, 4.7 and 4.8
  When it fitsThe most capable and the most expensive. Long reasoning and hard cases

Source: Salesforce developer documentation on supported models, retrieved September 7, 2026.

Two things follow from that list. First, "preferred model" does not mean "only model": the same documentation lists models from OpenAI, Google, Amazon and NVIDIA. You get to choose. Second, model choice hits the invoice directly, because cost per query varies widely between families. Running Opus where Haiku would do is one of the most expensive and least visible ways to waste an AI budget.

The numbers, read carefully

The 8.1 million hours figure traveled the furthest and is the easiest to misread, so it is worth pausing on. It is an internal Salesforce measurement of Slackbot usage among its own employees, annualized, and it grew more than 2x quarter over quarter. It is not a customer outcome, it is not audited by a third party, and it is not a promise of what you will get. It shows Salesforce uses what it sells, which counts for something, but it is not a business case.

There is something equally informative in what they did not publish. The Claudeforce release contains no pricing, no edition or license requirements, no deal figures between the two companies, and no individual customer names. If someone shows you a precise Claudeforce ROI calculation, it did not come from the official sources.

As ecosystem context, there is a fourth announcement that is not Salesforce's but affects anyone implementing: on March 12, 2026, Anthropic launched the Claude Partner Network with an initial USD 100 million investment for 2026, free membership and its own technical certification. That is the signal that the partner channel will be the main path to adoption.

What this means for your company

It depends a good deal on where you stand today. Four situations, and what changes in each.

If you already run Salesforce

The immediate change is the interface, not the data. Your objects, permissions and automations stay where they are. What gets added is a new way to query them and to trigger actions. The prerequisite work is not technical: it is checking that the data feeding those 37 skills is complete. A close plan generated on an opportunity with no stage, no amount and no linked contacts will not help anyone.

If you are evaluating Salesforce

Model choice has stopped being a configuration detail. It is worth putting into the evaluation which models each platform offers, whether processing stays inside its perimeter, and what happens to sensitive data. Two years ago nobody asked this question; today it decides architectures.

If you operate in a regulated sector

This is where the announcement carries the most weight. Banks, insurers, healthcare providers and fintechs in Latin America kept hitting the same wall: AI was interesting, but the risk function would not approve moving customer data. The trust boundary argument does not remove the analysis, but it changes where that conversation starts.

If your team lives in Slack

This is the most available piece and the lowest-risk place to start. Claude inside Slack does not touch your data model and does not require a project. It is also the best way to find out whether your team actually adopts this kind of tool before investing in something larger.

How to start without rushing

Our recommendation, after reading all four announcements in full, is to treat Claudeforce as what it is today: a sales beta with an ambitious roadmap. That suggests an order of operations.

• First, data quality. Audit the fields that feed forecast, stages and contacts. That work pays off with or without AI, and nothing else works without it.
• Second, Slack. It is available, needs no project, and gives you a real read on team adoption.
• Third, one narrow agent in Agentforce. One use case, measured, on the cheapest model that delivers the expected result. Scale after that, not before.
• Fourth, the open beta. When it ships, run it with a small group of reps and compare them against the rest before rolling it out.

What we would not do is rewrite this year's plan around an announcement whose larger half still says "coming soon". The two-year sequence shows both companies are moving seriously, and it also shows that months pass between an announcement and general availability.