Salesforce said in its October 27, 2026 announcement that Agent Fabric’s spending dashboard, Model Wallet and cost-optimisation policies are generally available. For budget owners and teams running AI agents across providers, the offering combines spending visibility with controls that can block further model requests.
Before treating a wallet as a ceiling on your AI bill, check three things: which requests pass through it, how their cost is calculated and what happens when a limit is reached.
Which spending falls inside the controls?
Model Proxy acts as a gateway between applications and AI providers. MuleSoft’s cost-management guide says requests that bypass it are outside these controls. The implication: a central dashboard is not, by itself, evidence that every AI service in a company is covered.
A Model Wallet associates requests with a caller, such as an application or agent, and sets a budget against a provider. Limits can use US dollars or tokens—the units models process—with daily, weekly or monthly resets.
Using wallets requires an Anypoint Platform account, a configured proxy, an identity provider issuing signed caller credentials, and policies on each relevant proxy. The wallet documentation also distinguishes spending control from permission to access the proxy: a wallet does not grant or deny that access.
For the separate question of agent identity and delegated access, see RohitAI’s coverage of IBM’s agent-identity controls.
Two ways a low total can mislead
Outside-proxy traffic is one gap. Missing model prices are another: administrators enter input and output rates, and models without configured prices count as zero dollars in both spending reports and dollar-based wallet budgets. Token-based budgets still count their usage.
Dollar reporting also starts only when rates are configured; it does not price earlier usage retroactively. Before relying on the total, map the workflows covered and assign someone to keep prices current for every production and fallback model.
What happens when a budget runs out?
Once a wallet’s allowance is exhausted, Model Proxy prevents subsequent calls to the affected provider’s models. But MuleSoft documents two qualifications: enforcement is approximate, allowing a brief overshoot, and a configured fallback can send the request to another model.
The wallet page gives no maximum overshoot or exact blocking delay. That means the setting should not be treated as a guarantee that the provider’s invoice can never exceed the configured amount.
For a customer-facing workflow, the practical choice is whether an approved backup model should continue the work or whether interruption is acceptable. Check the backup’s budget and suitability, not just the original route. A budget limit is also a service-continuity decision.
Price the whole deployment
MuleSoft lists packages from $2,000 a month, billed annually, with features and included credits varying by package. This is a platform starting price, not an all-inclusive quote for this Agent Fabric setup. Ask for your required features and expected credit consumption to be priced, separately from model-provider charges.
Optimisation can include reusing an earlier answer to a similar request. However, semantic caching requires supporting services, and its settings include answer expiry and identity-related matching options. For personalised or changing information, decide when reuse is acceptable before counting potential savings.
If buying primarily to reduce costs, compare expected avoided model charges with the extra platform, implementation and supporting-service expenses. Existing MuleSoft customers should use their incremental quote, rather than automatically treating the advertised starting price as a new expense.
Separate available controls from future features
Salesforce’s availability schedule puts its Agent Evals and API Evals testing tools in beta in November 2026, with general availability planned for January 2027. The evaluation-triggered Agent Kill Switch is also scheduled for January. Those are delivery plans, not features to assume are available now.
The current containment guide describes human review before quarantine and policies on each protected proxy. Its detection checks also make extra model calls, increasing token usage. This is distinct from the future evaluation-triggered intervention.
Availability wording deserves care. August 31 release notes already described wallets and budget management, while the undated product FAQ still calls some cost controls forthcoming. The October 27 statement is Salesforce’s dated general-availability announcement, not evidence of first-ever customer access. Confirm access for your package and environment.
Before committing, ask the platform team for a list of covered workflows and model rates, and have each business owner agree whether work may stop or switch models. If an absolute company-wide invoice ceiling is essential, the documented wallet behaviour does not establish one.