Revenium launches Guardrails to block rogue AI calls

Revenium has launched Guardrails, a set of runtime controls for AI spending and model access designed to stop unapproved AI calls before they reach a provider.

The launch adds a new control layer to Revenium's existing tools, which track AI spending and measure whether that spending delivered a useful result. Guardrails instead decides in real time whether a call should proceed, based on rules set by a customer's engineering or operations team.

Users can apply a rule to a specific organisation, product, agent, model or task type. Teams can choose to send an alert or block the call entirely before it reaches the provider.

That means a business can restrict use of a newly released model until its pricing and intended use have been reviewed internally. This can be done without changing application code if the call runs through Revenium's software development kit.

One example is model-level blocking for a new release such as Claude Fable 5. In that case, a team can place the model in enforce mode so any call from an agent or application is stopped until approval is given.

Guardrails also includes administrative features for policy management. Users can create a rule from a filtered view of spending by employee, carry that scope over automatically, retain a history for each rule and give colleagues read-only access where needed.

If a call is blocked, the person who created the rule can attach a message explaining why. This is intended to show developers the reason for the block rather than leave them with a failed request and no context.

Jason Cumberland, chief product officer and co-founder of Revenium, described the feature as a way to turn internal policy into an immediate technical control.

"Teams don't want to wait for a budget review to decide whether a brand-new AI model belongs in their stack," Cumberland said. "With Guardrails, that decision is a rule instead of a policy nobody reads. Point it at a model like Claude Fable 5, set it to enforce, and the answer is already built into the workflow."

The release comes as businesses try to manage AI costs that can change quickly with model choice, prompt design and usage volume. Many teams already monitor spending after the fact, but post-hoc reporting does not prevent a costly model or workflow from being used in production before someone spots the change.

Wider update

Guardrails is part of a broader set of updates across Revenium's platform. The company has also introduced smarter cost-risk alerts, automatic explanations for spending spikes, clearer labelling between billed and metered usage, and deeper employee-level analysis.

The cost-risk alerts are intended to flag cases where spending per call rises faster than usage, which may indicate a model swap or prompt change. They also watch for unusual spending patterns from a single agent, product or API key compared with that item's normal behaviour.

For sudden jumps in daily AI spending, the platform now aims to explain the change automatically by linking it to the people and usage behind the increase. This is meant to reduce the need for teams to investigate across multiple dashboards.

Another part of the release is a distinction between provider invoices and Revenium's own observed usage. Numbers across the product are now labelled as either billed or metered, and a reconciliation view shows where the two match and where instrumentation may lag actual spend.

Employee-level usage analysis has also been expanded. Filtering now covers model tier, provider and vendor, while benchmarking can be done against team norms and cost per million tokens. Data can also be exported as a CSV file.

John Rowell, chief executive officer and co-founder of Revenium, said Guardrails marks a shift from reporting past activity to intervening before spending occurs.

"Every dashboard we've built has been about knowing what already happened," Rowell said. "Guardrails is the first one that decides what happens next. A rule can stop a call before the provider ever sees it, which is a different kind of control than a chart that updates the next morning."