Salesforce will meter outside AI agents: what to do

If an AI agent you built or bought reads or writes your Salesforce data, how you connect it and what it costs are about to change. Salesforce says it is targeting November for three things: new security controls, registration for outside agents, and a new billing model where each successful agent call uses Flex Credits.

This post is for businesses that run Salesforce as their CRM and have connected (or plan to connect) an outside AI tool: a ChatGPT or Claude assistant, a custom agent, or anything that talks to Salesforce over its API or an MCP server. Our main source is Salesforce's help article Understand AIforce Impact, published September 17, 2026. We also checked the Agentforce pricing page, Salesforce's headless platform page, and the Flex Credits rate card dated August 31, 2026. All read on October 25, 2026. Business press picked up the plan on October 24, so expect questions from your team.

Salesforce calls the program AIforce. It was announced at TDX 2026 as "Headless 360", and the tools behind it are called the Headless Toolkit. There are three parts.

• Registration. An admin registers each outside agent. The agent gets its own identity and a narrower set of permissions. Salesforce's own comparison is giving the agent "its own badge" instead of lending it an employee's badge.
• Metering. Every successful call from a registered agent, over MCP or the regular API, counts as a Headless Platform Interaction (HPI). These are tracked in Digital Wallet and use Flex Credits.
• Security controls. New controls arrive with registration so admins can see and limit what each agent does.

The price per HPI is not public. The Flex Credits rate card lists Headless Platform Interaction with its credit amount as "TBA". The help article says agent usage is not being metered yet and that Salesforce will give 30 days' notice before metering starts.

What is public is the price of the credits themselves. The Agentforce pricing page lists Flex Credits at $500 per 100,000 credits in US dollars. So the math will be: number of successful agent calls, times the HPI credit amount Salesforce announces, times the credit price. Until that middle number is published, anyone quoting you a monthly cost is guessing.

• Traditional integrations. Salesforce says integrations that are not agents keep their current pricing and security. Your form-to-lead sync or accounting connector is not affected by this change on its own.
• Non-production orgs. Only active production orgs are metered. Sandboxes, scratch orgs and Developer Edition orgs are not.
• Agent calls before the notice. Nothing is metered until Salesforce gives the 30-day notice.

Salesforce's timing depends on when you bought.

• Customers who bought on or after September 17, 2026 must register every agent that uses Salesforce APIs within 3 months of Salesforce's notice that agent identity is available.
• Existing customers register and move their agents over at renewal.
• Any use of a Salesforce MCP server needs registration within 3 months of that notice, for every customer, regardless of when you bought.

To get ready, the article says you turn on Flex Credit billing (Salesforce Foundations, currently listed at $0, is one way), register each agent with its new identity, update each MCP client and API agent to use the new sign-in credentials, then reconnect.

Most small and mid-size Salesforce customers do not know how many agents already touch their org. A personal ChatGPT connection here, a sales tool's AI feature there. The registration deadline is a good reason to find out, and to tighten access while you are at it. Here is the work we would do:

• Inventory every connection. Pull the list of connected apps and API users from Salesforce Setup, then ask each team which AI tools they have connected. Mark each one as an agent (it decides what to do) or a traditional integration (it runs fixed steps).
• Give each agent its own identity. One identity per agent, with only the objects and fields it needs. A support agent that reads cases should not be able to edit opportunities. Our post on AI agent permissions and human approval covers how we set those limits.
• Count calls before they cost money. Log how many calls each agent makes in a normal week. When Salesforce publishes the HPI rate, you can price it in minutes instead of waiting for the first bill.
• Cut wasted calls. An agent can re-read the same record several times in one task. Caching, batching and having the agent ask for exactly the fields it needs all lower the count. Moving a fixed, repeat job out of the agent and into a normal integration may take it out of metering altogether.
• Test in a sandbox. Sandboxes are not metered, so do the re-registration and credential changes there first.
• Set an alert. Once metering starts, watch Digital Wallet usage weekly for the first month.

Typical projects here are a lead-qualifying agent that writes to Salesforce, a support assistant that reads case history, or an internal agent that answers pipeline questions over MCP. For CRM setups beyond Salesforce, our post on HubSpot credits and AI agent costs covers the same question for HubSpot.

• Shared logins. If an agent signs in as a real employee today, it will need its own identity. Plan the switch so the agent does not stop working when that person leaves or changes their password.
• Tools you did not build. A vendor's AI feature may count as an outside agent. Ask each vendor whether they will register their agent and who pays for its calls.
• MCP access by edition. Salesforce's headless page says MCP access is included with Enterprise Edition and above and uses your existing API limits. If you are on a lower edition, check what you have before building on it.
• Retries. An agent that retries failed calls in a loop will cost more once metering starts. Make sure retries have a limit.

If your only connections are traditional integrations, such as a form tool creating leads or your accounting system syncing invoices, there may be nothing to register. Confirm that with your inventory and move on.

If one person on your team uses an AI assistant to look up a few records a day, registering it is mostly an admin task. A full project is not needed. Just make sure the connection has its own identity and the narrowest access.

This week, list every AI tool that connects to your Salesforce org and who owns each one. Then check your contract renewal date, because that decides your deadline if you are an existing customer.

If you want help, our free automation audit reviews your Salesforce connections and tells you which ones need registration, which can become plain integrations, and where calls can be cut. See what our systems integration and MCP server development projects include, or read more on our Salesforce integration page.