Salesforce Launches AIforce at Dreamforce 2026 as Agentforce ARR Passes $1.5B

Salesforce opened Dreamforce in San Francisco on 15 September 2026 with a new umbrella product, AIforce, that is meant to put Salesforce data, workflows and permissions inside whatever AI interface a customer already uses, starting with Anthropic's Claude. Marc Benioff called the sell-off in software stocks "crazy nonsense" and the launch "an interface revolution", according to TheNextWeb. The conference came three weeks after Salesforce reported that Agentforce annual recurring revenue exceeded $1.5 billion, up over 240%, a figure that now counts more products than it used to.

What happened

The Dreamforce announcements fell into three groups.

Interfaces. AIforce is described as a live layer that connects Salesforce's data, business logic and security model to any AI interface or agent, with an SDK and an "AIforce Max Edition" bundle promised. Its first surfaces are Claudeforce, which exposes Salesforce inside Claude through an MCP server with 37 prebuilt sales skills and is in open beta; Slackforce inside Slack; and Agentforce Coworker inside Salesforce's own Lightning UI. Salesforce said Coworker was enabled by 100,000 users in its first 35 days, per TheNextWeb. A Claude Code plug-in for Salesforce developers ships more than 40 skills.

A model. Koa is Salesforce's first CRM reasoning model, post-trained on Nvidia's Nemotron 3 Super, with Salesforce holding the weights and running inference on its own infrastructure. Salesforce claims about three times fewer errors than leading models on its own CRM benchmark. It is in pilot with six customers, with general availability in US regions expected in winter.

Named agents and pricing. Salesforce packaged its agents as job titles: Casey for service, Hunter for outbound sales, Piper for inbound leads, Paige for IT and HR, plus commerce and supply chain agents, according to Salesforce Ben. The company claimed, per TheNextWeb, that Casey had handled 5 million resolved conversations and Hunter had generated $500 million of pipeline last quarter. On pricing, SiliconANGLE reported that Salesforce now charges for Agentforce service agents only when a customer's issue is resolved, not when the agent hands the case to a human.

Governance featured heavily. TechTarget noted that many announcements were about guardrails and infrastructure rather than features, and quoted Constellation Research's Liz Miller: "The governance piece will be where the mistakes are made and prevented."

The numbers

Why it matters

Two shifts are worth separating from the branding.

First, Salesforce is conceding that the user interface may not be Salesforce. Putting CRM inside Claude via MCP, with the customer's own permissions enforced, is a bet that the system of record survives even if people stop opening its screens. That is a defensive move, and an honest one. It also makes MCP a first-class enterprise integration path rather than a developer toy, which matches what I described in the 2026 MCP ecosystem.

Second, pay-per-resolution pricing for service agents is the cleanest example yet of the outcome-based pricing that vendors have talked about for two years. It moves risk to the vendor: if the agent fails and escalates, the customer does not pay for that attempt. It also hands the vendor the definition of "resolved", which is where every argument will happen.

What sceptics say

The Agentforce ARR growth rate is real but not like-for-like. From Q2 FY27, Salesforce folded Slackbot and Headless 360 into the metric, so the 240% compares a broader basket today with a narrower one a year ago. Total revenue grew 11% including an acquisition contribution, which is a long way from the AI line's growth.

Adoption breadth is also thinner than the keynotes suggest. Salesforce Ben put Agentforce adoption at around 12% of customers as of earlier this year and noted that the stock had fallen by roughly a third since the company's run of more than a dozen acquisitions began in 2025. The usage figures from the stage (Coworker users, resolved conversations, pipeline) are company claims without independent audit, and "pipeline generated" is not revenue.

My view, as opinion: the AIforce strategy is sensible, and Claudeforce is the most useful thing announced, because it meets users where they already are. But outcome pricing only works if the customer can measure the outcome independently. Without that, it is per-conversation pricing with better marketing.

What it means for engineering and DevOps teams

• Treat Claudeforce-style integrations as production integrations. An MCP server with access to CRM records needs the same review as any API integration: scopes, audit logs, data retention. Salesforce says the products are "built with zero data retention"; verify that in your contract, not the keynote.
• Instrument "resolved" yourself. If you buy per-resolution pricing, log your own resolution signal (reopen rate, follow-up contact within N days) and reconcile it against the bill monthly.
• Expect agent sprawl across vendors. Salesforce, Microsoft and ServiceNow each shipped an agent registry or control plane this autumn. Decide which one is authoritative for your organisation before each team picks a different one. Non-human identity is the right lens.
• Budget for consumption. Agentic Work Units and resolutions are metered. Put them through the same FinOps process as cloud spend.

FAQ

What is Salesforce AIforce?

AIforce is a layer announced at Dreamforce 2026 that connects Salesforce data, workflows and permissions to outside AI interfaces and agents. Its first surfaces are Claudeforce (inside Anthropic's Claude), Slackforce (inside Slack) and Agentforce Coworker (inside Salesforce Lightning).

How is Agentforce priced in 2026?

Salesforce uses several models, including per-user and consumption pricing. At Dreamforce 2026 it said Agentforce service agents are charged when a customer issue is resolved rather than when the case is handed to a human, according to SiliconANGLE.

What is Agentforce ARR?

It is Salesforce's annual recurring revenue figure for its AI agent products. It passed $1.5 billion in the quarter to 31 July 2026, but from that quarter it also includes Slackbot and Headless 360, so growth rates are not directly comparable with earlier quarters.

Related: Seats Are Dying. What the 2026 SaaS Repricing Looks Like From the Compute Layer, What Is MCP (Model Context Protocol)? and The 2026 SaaS Selloff.