Salesforce adds tools to catch AI agents that quietly give wrong answers

Salesforce announced new tools on October 27, 2026 that watch AI agents for quality problems and rising costs, then step in when something goes wrong. The tools are part of Agent Fabric, the company's control system for managing AI agents across different platforms. The goal is to catch a wrong answer or a spending spike before a customer complains or the bill grows.

The problem Salesforce describes

An AI agent is software that uses AI to handle a task, such as answering questions about an order. Salesforce says agents can fail silently. They give wrong answers and rack up costs before anyone notices. It also says fewer than half of organizations with agents test them on live traffic. Once a problem is found, fixing it can mean tracking down the owner and getting approval to restrict the agent while it keeps making mistakes.

What is new

• Agent Evals creates tests before launch, then checks agents in live use against quality targets set by the company.
• If results fall short, a preset response can run, including an Agent Kill Switch that takes the agent offline, with human approval if policy requires it.
• API Evals tests the connections agents depend on, so a change to one is caught before it reaches customers.
• Model Wallet sets hard spending limits that stop requests before costs pass the budget, and a FinOps Dashboard shows all AI spending in one view.

Salesforce also offers Cost Optimization Policies. It says they cut token usage by up to 75%, based on test scenarios. Tokens are the units AI services use to measure and bill usage. That figure comes from Salesforce's own tests, not from customer results.

How it might work in practice

Salesforce gives an example. A retailer switches its order-support agent to a cheaper AI model. Agent Evals runs scheduled tests and finds the agent is giving wrong information about delayed orders. It asks for approval to quarantine the agent, because the retailer wants a person involved. After approval, the team improves the agent's instructions and reruns the tests. Only then does the agent go back to customers.

Other updates and open questions

Salesforce also announced a Custom Agent Scanner, which finds and lists agents in internal and third-party systems. It added shared rules for API gateways from Azure, Apigee and Kong. ASICS Digital is helping shape the Agent Evals beta, which suggests the feature is still early. The material does not say when it will be widely available or what it will cost.

What to do

Any business that lets AI answer customers should ask two questions. Who checks the answers, and what stops the bill from growing unnoticed? Salesforce says its tools address both, though only its own customers can use them.