The news
The contest over enterprise AI agents shifted in September 2026 from which model is smartest to which vendor owns the screen where employees work. Anthropic, Salesforce (CRM), OpenAI and Meta Platforms (META) each launched products between September 10 and 28 aimed at that layer.
On September 15, Anthropic released Salesforce in Claude in beta, a plugin with 37 skills that brings accounts, opportunities and pipeline data into Claude for tasks such as call preparation, pipeline reviews and CRM updates. According to Anthropic, sellers sign in with Salesforce credentials, Claude reads only what their existing access allows, and by default it asks for approval before writing changes. Anthropic said 7,000 Salesforce sellers already use it, with GitLab, Siemens and Legora named as early adopters.
The skills go beyond lookups. Unite.AI reported that the plugin can assemble a morning brief of meetings, closing deals and at-risk opportunities, prepare for calls using Salesforce, Slack and email, score deals against a team's sales method and turn call transcripts into follow-up emails. It requires the latest enterprise edition of Sales Cloud, and by default customer data on Claude's Team and Enterprise plans is not used for model training.
The plugin builds on an expanded partnership announced in August, under which Claude became a reasoning model for Salesforce's Atlas engine and the default model for Slack, Unite.AI reported. On September 17, at its Dreamforce conference, Salesforce launched AIforce, a platform combining its data, applications, permissions and agents. It said Agentforce has more than 30,000 customers, and it introduced a CRM reasoning model called Koa built on Nvidia's Nemotron models, plus a MuleSoft registry for managing agents built on Microsoft Azure, AWS, Google and Salesforce, according to Yahoo Finance.
Model makers are building their own routes in. OpenAI opened its Agents API to public beta on September 10, letting developers build agents on a managed Codex harness while OpenAI handles session orchestration, context compaction and recovery. On September 28, Meta announced an enterprise AI platform including its Muse agent, a business agent and developer APIs, and named MongoDB chief executive CJ Desai as chief enterprise platform officer reporting to Mark Zuckerberg, VentureBeat reported. Meta did not disclose pricing.
The numbers
- Sales skills in Salesforce in Claude
- 37
- Salesforce sellers using the plugin, per Anthropic
- 7,000
- Agentforce customers, per Salesforce at Dreamforce
- More than 30,000
- Agentic work units processed by Agentforce, per Yahoo Finance
- 7 billion
- Meta enterprise AI pricing disclosed
- None
Why CEOs should care
For CIOs, the practical decision is where agents should live. Running them inside the system of record, as Salesforce's Agentforce does, keeps data and permissions close to the source. Running them inside a model vendor's workspace, as Salesforce in Claude does, gives employees one assistant across many tools. Either choice creates lock-in. Set a default per function, and require that any agent inherit the user's existing permissions and ask before writing data, the model Anthropic and Salesforce adopted for this plugin.
CFOs should watch for paying twice. A sales team could hold Claude seats, Salesforce agent licenses and new platform fees at the same time, each billed differently. Build a single inventory of AI agents and their costs, and judge vendor claims about productivity, such as Salesforce's statement, reported by Yahoo Finance, that its internal agents generated $500 million in pipeline in the prior quarter, against measured results in your business.
CISOs and risk leaders should treat each new agent surface as a new access path to customer data. Agents that pull from CRM, email and Slack together widen what a single compromised account can reach. For newer entrants, check vendor staying power: VentureBeat noted Meta has discontinued earlier enterprise software, so contracts should include data export and transition terms.
The bigger picture
Control of the interface decides who captures the margin. Model developers want employees to work inside their assistants and call business software as tools. Application vendors want the reverse, and Salesforce is hedging by putting Claude inside its products, putting its data inside Claude, training its own Koa model and offering a registry for agents from rival clouds. Meta's arrival adds a well-funded competitor whose past enterprise software efforts were discontinued, which could pressure prices but adds execution risk.
What’s next
Watch for general availability and pricing of Salesforce in Claude, the terms OpenAI sets when its Agents API leaves beta, and Meta's first enterprise pricing and customer names. The number of agents companies register across vendors in tools like MuleSoft's registry will also show whether multi-vendor agent fleets become the norm.
