Salesforce Launches FinOps Dashboard and Model Wallet: Stopping AI Cost Overruns Before They Hit Customers

On October 27, 2026, Salesforce announced a new expansion of Agent Fabric — its AI agent governance platform launched in September 2025 — this time focused on two concrete problems companies face when deploying Agentforce or other AI agents at scale: runaway language-model costs and agent failures in production. Per the official announcement (salesforce.com/news), Agent Fabric now handles more than 1.2 billion requests per day, a volume Salesforce says justifies adding dedicated financial-control and oversight tools.

What's Available Today (October 27, 2026)

• FinOps Dashboard: a consolidated view of AI spending by team or cost center, aimed at budget owners.
• Model Wallet: hard spending caps that block requests before an allocated budget is exceeded, rather than alerting after the fact.
• Cost Optimization Policies: reduces token consumption via semantic caching, payload compression, and response trimming, with a reduction of "up to 75%" based on test scenarios cited by Salesforce.
• Open Policies for Federated API Gateways: centralized discovery, enforcement, and tracking of API policies (IP allowlisting, rate limiting) across Azure, Apigee, and Kong from one console.
• Agent Kill Switch: a manual shutdown for an agent in production, available today.

What's Coming Later — Don't Confuse It With Immediate Availability

The announcement clearly separates several pieces still in development: Agent Evals and API Evals (generating tests before deployment and continuously evaluating agents, LLMs, and MCP servers in production) enter beta in November 2026, with general availability announced for January 2027. Custom Agent Scanner, which catalogs the agents, models, and MCP servers present in a company's environment (including behind a firewall or at third parties), is in beta today, but its general-availability version doesn't arrive until January 2027 — as does the automatic triggering of the kill switch by Agent Evals (Triggered-Agent Kill Switch).

Why This Matters Beyond Salesforce Customers

Salesforce cites two third-party statistics to justify the announcement: fewer than half of organizations reportedly evaluate their AI agents on live traffic (as opposed to controlled test environments), and 96% of enterprises reportedly run their agents on more than one platform. Both figures come from third-party sources cited by Salesforce in its release, not from an internal study — treat them as a product pitch backed by external data, not first-party Salesforce research. The underlying point still holds for any SMB deploying AI agents, Salesforce or not: without a hard spending cap and an emergency stop button, an agent that loops or over-calls a token-billed model can run up an uncontrolled bill before anyone notices.

What This Actually Means for Your Decision

• If you already use Agentforce: the FinOps Dashboard, Model Wallet, and Agent Kill Switch are available today — check licensing prerequisites and whether manual activation is required with your Salesforce contact, since the release names no associated pricing.
• The "up to 75%" token-reduction figure comes from Salesforce's own test scenarios, not an independent audit — verify it against your own traffic before building it into an ROI calculation.
• If you're not on Agentforce: the underlying problem (uncapped token costs, no emergency stop button) is generic to any AI-agent deployment built on usage-billed LLMs, regardless of vendor.
• Continuous production-evaluation features (Agent Evals, API Evals) aren't available yet — don't plan an AI rollout that depends on them before January 2027 at the earliest.

Deploying or planning AI agents and want to avoid cost overruns and silent production failures? Our security and governance offering sets up spending caps, access controls, and emergency-stop procedures before go-live.