Salesforce wants every outside AI agent that touches its platform to stop at the front desk and sign in first.
The company plans to require customers to register external AI agents that use its APIs or Model Context Protocol (MCP) servers. It is calling the system Agentic Identity, and customers will also be able to monitor what those agents access. Pricing for all this is still unknown, which is the part customers care about most.
How Agentic Identity would work
The requirement is scheduled to launch in November 2026. It is part of a wider set of upgrades Salesforce is rolling out under its Headless Toolkit and AIforce banners.
“Headless” describes software that reaches into a platform directly, without going through the usual on-screen interface. An AI agent pulling customer records through an API is a good example. No human is clicking anything.
MCP works like a universal adapter. It gives AI agents a standard way to connect to outside tools and data, and Salesforce runs its own MCP servers for that purpose.
Under the new rules, each registered agent gets its own unique identity and its own set of permissions. Those permissions will be narrower than what a human user typically receives.
The rollout comes in stages, and the deadlines depend on who you are:
New customers who buy Salesforce services after September 17, 2026, must register their external agents within three months of being notified that Agentic Identity is available.
Existing customers get more runway for API-based agents. They will need to comply when their contracts come up for renewal.
Anyone using Salesforce’s MCP servers, new or existing, must register within three months of receiving notice.
The tooling arrives slightly ahead of the deadline. Salesforce’s Winter ’27 release, available from October 12, 2026, will include API Catalog features that customers can use to register their agents.
The Flex Credits question
Registration does more than tighten security. It also lays the groundwork for Salesforce to charge for third-party AI activity on its platform.
The company is moving toward a billing model that uses Flex Credits for headless platform interactions. Every time a registered outside agent works with Salesforce, that activity could eventually draw down credits.
How much it will cost has not been decided. Salesforce has not set exact pricing structures for these interactions.
Agent interactions will not be billed until Salesforce updates its Flex Credit rates. That update is expected to come after a 30-day notice to customers.
Why Salesforce is doing this now
The stated problem is what the company calls “agentic sprawl.” As AI agents multiply inside organizations, tracking which bot is touching which data gets messy fast.
Agentic Identity is meant to bring accountability to that crowd. Each agent becomes a named, permissioned entity with a traceable activity record.
The timing also tracks the growth of the third-party AI ecosystem. That ecosystem includes tools from major players such as Anthropic and OpenAI, whose agents increasingly plug into enterprise systems like Salesforce.
What this means for customers and investors
The move has drawn a positive reaction from investors. The logic is simple: Salesforce is turning outside agents from free riders into metered users.
Customers have a more mixed view, and pushback is anticipated. The main concern is fees. Businesses that built workflows around external AI tools may now face costs they did not budget for.
The staggered timeline matters here. For existing customers, compliance is tied to contract renewal. That effectively turns each renewal into a negotiation over AI agent access and cost.
There are several dates worth watching. The Winter ’27 release arrives October 12, 2026, with registration tools. The policy for new customers kicks in for purchases after September 17, 2026. Agentic Identity itself launches in November 2026.
The most important signal has no date yet: the Flex Credit rate update. Once Salesforce issues its 30-day notice, customers will finally see what registered agents cost to run.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.