Salesforce adds AI spending limits, with automated shutdown controls due in January

Salesforce is promising to stop failing AI agents, not just flag their mistakes. Its October 27, 2026 Agent Fabric announcement pairs spending limits available now with evaluations designed to trigger intervention. But the full automated shutdown path remains ahead: agent evaluations enter beta in November, and the triggered kill switch is scheduled for January 2027.

Agent Fabric is Salesforce’s shared management layer for governing agents across platforms. The expansion brings quality checks, budget enforcement and predefined responses into that layer, aiming to connect detection with action.

From a failed check to a controlled stop

Agent Evals is designed to generate tests before deployment, then evaluate agents, language models and MCP servers against defined quality thresholds in production. A failed evaluation can initiate a predefined intervention, including the Agent Kill Switch, with approval when the organization’s policy requires it.

Salesforce illustrates this with a hypothetical retailer switching an order-support agent to a cheaper model. Tests detect inaccurate delayed-order answers and request approval to quarantine the agent. The team then revises its instructions and retests before restoring it. This is an intended workflow, not a customer result.

API Evals addresses a different failure point: changes to the application programming interfaces an agent depends on. Salesforce says it generates functional and performance tests from API specifications to catch breaking changes before production.

Budget caps arrive before the evaluation suite

The FinOps Dashboard and Model Wallet give budget owners a consolidated view of AI spending and hard limits that stop requests before costs exceed budget, Salesforce says.

Salesforce says its Cost Optimization Policies can reduce token usage by up to 75% through semantic caching, payload compression and response trimming. That figure comes from company test scenarios, not a guaranteed deployment saving. It measures token usage rather than a demonstrated reduction in a customer’s total AI bill.

Discovery and policies reach beyond Salesforce

• Custom Agent Scanner lets teams build scanners for internal systems, environments behind firewalls and third-party platforms. It catalogs agents, models, MCP servers and APIs without registering each asset manually. It is in beta now, with general availability planned for January 2027.
• Open Policies for Federated API Gateways is generally available. Teams can discover, enforce and track policies across Azure, Apigee and Kong gateways, applying a rule such as IP allowlisting or rate limiting across them in one action.

For buyers, the delivery split is consequential. Salesforce cautions that purchasing decisions should rest on fully released capabilities. The immediate purchase case rests on available controls; evaluation-driven intervention remains a scheduled addition.