Skip to content
Mimir Lab
← Back to blog

tool breakdown

Claudeforce: what Salesforce announced, what it costs, and what it didn't say

· crm, salesforce, anthropic, claude, mcp, smb

In short: on 26 August 2026 Salesforce and Anthropic announced Claudeforce. The CRM goes into Claude as a plugin with 37 sales skills, and Claude becomes the default model for Agentforce Vibes, Agentforce Coworker and Slack. The plugin’s price is not published. The next day the stock closed up 22.6% (stockanalysis.com).

I read the press releases, price lists and analyses to separate what was announced from what was narrated. Below are the facts, with sources; my expectations are in a section of their own, and the two don’t mix.

Update, 23 September 2026. At Dreamforce (15–17 September) Salesforce in Claude moved to an open beta; Patrick Stokes told the keynote audience the plugin is available from the AppExchange (The D*AI*LY Brief, 20 September). There is still no published price, and nothing at Dreamforce changed the two-meter structure described below. Salesforce’s own product page, checked on 23 September, still says “available to select pilot customers today, with plans for an open beta starting in September”, tags Sales as “BETA” and lists no price and no required Claude plan. The stock has given back part of the jump: USD 237.58 at the 23 September close, 10% below its 3 September peak but still 15.5% above the pre-announcement close (stockanalysis.com). Everything below is as of 3 September unless marked otherwise.

What they announced

The announcement landed on the same day as Q2 fiscal 2027 results. The press release is headlined “The #1 AI Meets the #1 AI CRM”; the earnings release mentions neither Anthropic, nor Claude, nor Claudeforce.

Salesforce in Claude. A plugin with 37 ready-made sales skills: pipeline review, forecast narrative, meeting prep, lead triage, sfdc hygiene check. It reads Salesforce, Slack and any connector active in Claude (Salesforce Ben). At announcement it went to selected pilot customers, with the open beta planned for September 2026 and skills beyond sales from late 2026 (press release, product page). General availability date: not disclosed.

Claude inside Salesforce. Reasoning model for the Atlas Reasoning Engine, default for Agentforce Vibes and Agentforce Coworker, served via Amazon Bedrock inside Salesforce’s Trust Boundary (press release). What’s new is the default: Claude has been selectable in Salesforce since 3 September 2024 and was already the “preferred model” for regulated industries since 14 October 2025.

Slack. Claude becomes the default model, powers Slackbot and Claude Tag, and Anthropic is a “founding partner” of Slack Code. According to Salesforce, Slackbot has delivered 8.1 million annualised hours of productivity, more than double the previous quarter (press release).

Mutual adoption. Salesforce says it uses Claude as the default AI for its employees and is rolling out Claude Code and Claude Enterprise to all its developers and knowledge workers (press release).

Headless 360. Exposes data, workflows, agents and governance as APIs, MCP tools and CLI commands instead of screens. Apex Hours dates it to TDX in March 2026, with the hosted MCP server in beta since July 2026 (Apex Hours); the Claudeforce product page, however, still lists it as “coming soon”.

The product page lists twelve “coming soon” items, including Service, Marketing, Commerce, Revenue Cloud, Field Service, Tableau, MuleSoft, Informatica, Data 360 and Industries: none are available yet. Enterprise Frontier Safeguards, with customer-held encryption keys and logs in the customer’s cloud, arrive “in phases beginning later this fall 2026”.

The causes

January’s crash explains most of the rest. On 30 January 2026 Anthropic released 11 plugins for Claude Cowork and software stocks fell: a Jefferies trader coined “SaaSpocalypse”, roughly USD 285 billion of market value was wiped out in a day, the S&P Software & Services index fell 25% between 12 January and 23 February 2026, and Salesforce and ServiceNow dropped about 7% (The Batch).

On the other side of the table sits a supplier that has changed order of magnitude in the meantime. Anthropic closed its Series H on 28 May 2026: USD 65 billion raised at a USD 965 billion post-money valuation, with a revenue run-rate that, Anthropic writes, had already passed USD 47 billion in May.

Patrick Stokes, Salesforce’s President of Applications & Marketing, put it bluntly, as Apex Hours reports: when people stop using Salesforce through the traditional interface and move to an agentic one, that “dramatically increases the value of Salesforce”.

Then there’s the income statement. The Motley Fool calculates that gains on strategic investments contributed USD 2.53 of the USD 5.90 of adjusted earnings per share, that USD 2.7 billion of unrealised gains in the quarter are attributed to the Anthropic stake, and that excluding those gains adjusted earnings grew about 16%. The earnings release limits itself to a note: guidance assumes the value of the strategic investment portfolio stays unchanged.

The effects

On 27 August 2026 Salesforce closed at USD 252.05, up 22.58% from USD 205.62 the day before (stockanalysis.com); CNBC’s headline said “almost 23%” and the best day since 2020. The same day the IGV software ETF rose 8%, ServiceNow 10% and Adobe 6% (The Daily Upside).

Two details from the same daily price history put the jump in context. Volume that day was 55.5 million shares, about four times the six-month median of 13.0 million. And the rally didn’t start on the news: from its six-month low of USD 150.12 on 22 June, the stock had already climbed 37% by the close of 26 August (stockanalysis.com).

Update, 23 September 2026. The close peaked at USD 264.43 on 3 September. Dreamforce week didn’t help: from USD 259.43 on 14 September to USD 242.85 on 17 September, three down days in a row (−6.4%). On 23 September the stock closed at USD 237.58: −5.7% from the 27 August close, −10.2% from the peak, still +15.5% above 26 August (stockanalysis.com).

The quarter’s numbers, from the earnings release: revenue USD 11.3 billion, up 11% year on year; Agentforce ARR above USD 1.5 billion, up more than 240% year on year; 3.2 billion Agentic Work Units in the quarter, up 97% quarter on quarter. Benioff, in a CNBC interview quoted by The Daily Upside: “This is not the SaaSpocalypse”.

Analysts are split. For Keith Bachman (BMO Capital Markets), the big AI labs are already large Salesforce customers and frontier models consume CRMs more than they replace them; for John DiFucci (Guggenheim), AI poses a material risk to many software vendors but is “not likely to be a death knell” (The Daily Upside).

Pricing: what’s known and what isn’t

The price of Salesforce in Claude is not published. The press release only says that “pricing and packaging are subject to change”, and the product page lists neither prices nor required editions.

What is published sits on two separate price lists, both checked on 3 September 2026. I rechecked the Agentforce one on 23 September: unchanged, except the per-conversation option, which no longer shows a price.

Agentforce price list:

Item Price
Flex Credits USD 500 per 100,000 credits (USD 0.005 per credit)
Standard action 20 credits = USD 0.10
Voice action 30 credits = USD 0.15
Per-conversation option USD 2 per conversation on 3 September; on 23 September “for existing Agentforce Conversations customers”, price via account executive
Agentforce User License USD 5 per user per month, credits extra
Sales, Service, Field Service add-on USD 125 per user per month (Industries: USD 150)
Agentforce 1 Editions from USD 550 per user per month, with 2.5 million credits a year per org (1 million in 2025)
Salesforce Foundations USD 0, with Agentforce Builder, Prompt Builder, Agent Script, Coworker and Vibes included

Claude price list:

Plan Price
Free USD 0
Pro USD 17 per month billed annually (USD 20 monthly)
Max from USD 100 per month
Team USD 20 per standard seat, USD 100 premium (USD 25 and 125 billed monthly)
Enterprise USD 20 per seat plus usage at API rates

API rates, in USD per million input and output tokens: Sonnet 5 at 2 and 10, Opus 5 at 5 and 25, Haiku 4.5 at 1 and 5.

Two price lists, two vendors, two meters. VentureBeat reports that the plugin is billed on consumption, tied to API calls rather than seats, and that inference is contracted separately with Anthropic. Forrester (Kate Leggett and Ted Schadler, 28 August 2026) writes that on top of the per-seat and consumption charges already on the bill come Claude Cowork licences “priced by consumption band”, that “you cannot purchase them today in a single contract”, and asks customers: “Are you ready for a surge in mystery costs from Anthropic and from Salesforce?”. The Register recalls a Gartner warning from October on consumption models: credits often require upfront commitments and rates can change unilaterally; in the same article Salesforce COO Miguel Milano claims that “fifty percent of the bookings came from customers refilling the tank”. The Motley Fool estimates that only about 5% of Salesforce’s sales and service seats are on the premium tiers that include AI, which cost 60% to 80% more. According to VentureBeat, Salesforce expects to spend about USD 300 million on Anthropic tokens in 2026.

An example built on my own assumptions — read it as plain arithmetic: if plugin actions were metered like standard Agentforce actions at USD 0.10, a seller running 50 a day would spend about USD 100 a month over 20 working days; add a Claude Team seat at USD 25 and you reach about USD 125 per user per month, on top of the Sales Cloud licence. No source confirms the plugin is metered this way.

ROI: vendor claims and independent evidence

Every Claudeforce number comes from the vendor side, with Salesforce calling itself “Customer Zero”. The press release cites 8.1 million annualised hours from Slackbot; Apex Hours reports that 83% of Salesforce’s workforce uses it; VentureBeat describes a seller’s morning routine going from about 10,000 clicks to about 30 seconds. Results from external pilot customers: none found.

For Agentforce customers the metrics are still Salesforce’s own (metrics page): Grupo Falabella reports tripling its WhatsApp support volume in three months with 60% autonomous resolution; Engine reports 50% of cases handled autonomously and 15% shorter handling times.

The most-cited study is Forrester’s Total Economic Impact of Agentforce Field Service, 29 April 2026: 195% ROI, USD 8.7 million net present value, USD 13.2 million of benefits against USD 4.5 million of costs over three years. Read it together with the headline: the study was commissioned by Salesforce and rests on five interviewees, with a composite organisation of USD 500 million in revenue.

An independent, non-commissioned ROI study of Agentforce or Claudeforce: I didn’t find one. In my dossier, that absence weighs the most. The only precedent in a regulated industry is RBC Wealth Management, which in 2025 used Claude in Agentforce to prepare advisor meetings, with no published figure (Salesforce).

Implementation

The mechanism is standard MCP, the Model Context Protocol. The Salesforce Hosted MCP Server, in beta since July 2026, exposes four tools: Discover, Describe, Dispatch and Dispatch Read-Only. It requires API v67.0 or later; with Winter ’27 that constraint goes away (Apex Hours).

Authentication is per-user OAuth through External Client Apps with the mcp_api scope: every call runs as the user who makes it (Apex Hours). According to VentureBeat, an admin connects the org once and everyone stays within their own permissions; Stokes: “if you don’t own that record, if you don’t have permission to see that record, the MCP server doesn’t either”. Writes go through the usual controls: validation rules, Flows, Apex triggers and CPU limits behave as before (Apex Hours). Salesforce claims “zero data retention” on the Sonnet, Opus and Haiku models (product page).

Inference runs on Bedrock inside the Trust Boundary, in Salesforce’s VPC (Apex Hours). The Claude platform generally offers EU/EEA regional processing (Claude); the inference region for this deployment is not disclosed, and I found no comment on the Swiss FADP. According to licensing guides, Agentforce requires Enterprise Edition or higher; the requirement for the plugin itself is not disclosed, and neither is the language of the 37 skills.

The most concrete recommendation I read comes from MarketScale: read-only copilots are a training issue, agents that write become a controls-and-audit-trail issue. Map the write scope and the human approvals before the pilot.

Potential

Headless 360 exposes the CRM as a set of governed tools over APIs and MCP (Apex Hours). From there on, where users work depends on the assistant in front of them. Today it’s Claude; trade press reports Agentforce 360 inside ChatGPT since 14 October 2025 and Agentforce Sales inside Gemini Enterprise, in open beta, since 22 April 2026. The same pattern, again per trade press, repeats elsewhere: HubSpot with a CRM connector for Claude since 29 July 2025, SAP with Claude as its primary reasoning engine since May 2026, ServiceNow with a deal in early 2026. The Motley Fool notes that the deal doesn’t lock Salesforce customers into Claude alone.

The installed base is large: more than 25,000 companies use Agentforce in 124 countries (metrics page), with 7.0 billion Agentic Work Units delivered to date (earnings release). At Dreamforce the GM of Agentforce, Mark Wakelin, put it at 30,000 customers live (The D*AI*LY Brief).

Pros and cons

Pros Cons
Per Salesforce: permissions inherited from the CRM, no new authorisation model Plugin price not disclosed; two separate contracts (Forrester)
Per Salesforce: writes subject to validation rules, Flows and Apex triggers Forrester: “new power users” may emerge, hammering the MCP server to see data they shouldn’t see today
Per Salesforce: inference inside the Trust Boundary via Bedrock, zero data retention on Sonnet, Opus and Haiku Inference region for the EU and Switzerland not disclosed
Open standard (MCP) and one-step org connection (VentureBeat) The Register/Gartner: upfront commitments and unilaterally changeable rates on consumption models
Models swappable on paper (Motley Fool) Dave Michels (TalkingPointz, on No Jitter): the primary interface moves to Anthropic; Apex Hours: “deep mutual dependency”

Forrester closes with three recommendations: sellers in regulated industries with compliance-bound processes should wait; partners and system integrators should go all in now; marketing, service and commerce should watch, listen and learn (Forrester).

My expectations

From here on it’s my opinion, deliberately kept apart from the facts above.

I read it as a distribution decision. Claude had been selectable inside Salesforce since September 2024; on 26 August Salesforce accepted that its own screens are optional. It keeps the data, the permissions and the meter, and lets the screen go. The market rewarded the trade, though four weeks on it has given back about a third of that gain.

In my view the SaaSpocalypse thesis only changes shape: in January the market punished whoever risked losing the interface; in August it rewarded whoever stopped defending it to sell data and metering. I bet that by the end of 2027 a plugin inside Claude, ChatGPT or Gemini will be normal for any enterprise SaaS, with consumption units replacing seats as the metric investors watch. I’m not betting on per-seat prices falling in 2026.

I expect two stable meters. On price, I bet on an announcement around Dreamforce (15–17 September 2026) or at the opening of the beta, shaped as “included for orgs already licensed for Agentforce, MCP calls on credits, inference billed separately”. I half believe it; if I’m wrong, the most likely alternative is a Claude Enterprise bundle resold by Salesforce. General availability in the first half of 2027. (As of 23 September: the beta opened at Dreamforce, the price still hasn’t — see the update at the top.)

For a Swiss SMB with 10 to 200 employees I expect a small direct effect in 2026 and a large indirect one. My impression, without hard data, is that few SMBs in Ticino are on Sales Cloud; those that are will see the beta from September, without a price and with two contracts to sign. But the pattern gets copied fast, and at that point the CRM’s interface stops being the product: what remains is clean data, the permission model and a set of skills. The cheapest preparation is data hygiene and a review of sharing rules. Cost: time, zero licences.

I also bet that governance incidents will arrive before ROI numbers: the first public problems will be about data exposed through sharing rules that the friction of the old interface kept hidden, more than wrong writes to the CRM. The mitigation I’d recommend: a two-day audit of sharing rules and field-level security before switching the plugin on.

On ROI, I expect vendor-side evidence only for two or three quarters. The rule I’d apply: one team, minutes per task measured before and after for 30 days, a spending cap. Claude Tag already has per-org and per-channel caps with alerts at 75% and 95% (Claude); I’d ask for the same on the plugin.

That leaves the financial entanglement. Salesforce is at once Anthropic’s investor, customer and distributor: USD 2.7 billion of unrealised gains in the quarter (Motley Fool) and about USD 300 million of tokens planned for 2026 (VentureBeat). My reading: decisions about which model is “default” have a price, optionality exists on paper, and roadmap money follows Claude. I expect the preferred-alliance label to hold for all of fiscal 2027.

What we still don’t know

As of 3 September 2026, none of these had been disclosed:

  • The price of Salesforce in Claude.
  • The Claude plan required to use the plugin.
  • The Salesforce edition required for the plugin.
  • Whether plugin calls consume Flex Credits or Agentic Work Units.
  • The general availability date and how pilot customers are chosen.
  • Inference regions and data localisation for the EU and Switzerland.
  • The languages of the 37 skills.
  • The names of external pilot customers and their results.
  • Any independent, non-commissioned ROI data.
  • Anthropic-side terms: rate limits and spending caps for the plugin.

Update, 23 September: the open beta started at Dreamforce; every item above is still undisclosed on Salesforce’s product page.

Sources

https://www.salesforce.com/news/press-releases/2026/08/26/salesforce-and-anthropic-announce-claudeforce/

What it means