Authorisation for the actions that matter
Zero trust for AI voice agents.
We make an authorisation layer that sits between an AI voice agent and the tools it can call. Before a refund, credential reset or payout change, Ringfence checks who the caller is and what they’re allowed to do.
Prototyping. Seeking design partners.
Illustrative simulation · fictional customer and actions
Daniel asks to change his payout destination. Press Play to follow the request.
Decision trail
lookup_accountInternal record resolution. No private details disclosed.update_payout_account(•••• 5471)Identity, account permission and exact-action approval required.Ready. No action executed.
Open the narrated demo → Reviewer login required
The problem
A convincing caller can ask for a consequential action.
Connecting a voice agent to business tools creates a new place to enforce permissions. A plausible story, a familiar voice or an instruction in a prompt cannot authorise a transaction.
The tool call carries the consequence
Refunds, password resets and bank-detail changes need checks against the customer’s actual account and policy. Sensitive reads need protection too.
Identity is only part of the decision
Knowing who called does not grant access to every account or permission for every change. Approval must cover the specific action and its parameters.
Social engineering crosses channels
Urgency, authority claims and requests to skip verification can arrive through a real human or a synthetic voice. The action boundary needs a consistent policy.
Builders need an integrated workflow
Platforms provide tools for authentication and agent behaviour. Teams still need to connect caller verification, business permissions, independent approval and execution evidence.
The mechanism
Every sensitive action earns its approval.
Ringfence evaluates tool requests against customer-owned policies and returns allow, deny or step-up. Risk signals can tighten a check; they cannot grant permission.
Intercept the request
A webhook or SDK wrapper puts the gate between the agent and its protected tools. Unconfigured tools default to deny.
Check the caller
Track what the caller has proven in the session, then check their permissions for the requested account and action.
Approve the exact change
Step up through an enrolled device or an appropriate channel on record. Bind approval to the intended transaction and expiry.
Enforce and record
Recheck before execution. Keep the decision, challenge result and execution outcome together in an evidence trail.
Checks at the tool boundary
The gate runs when an agent requests an action. It does not require a classifier round trip on every spoken turn.
Evidence with the decision
Associate the matched policy and challenge outcome with the relevant transcript and audio segment where available. Keep approval distinct from whether the action actually completed.
Part of your security stack
Connect identity to execution.
Ringfence brings voice-session context, scoped step-up and tool decisions into one workflow, alongside the controls you already use.
| Layer | What it contributes | How it fits |
|---|---|---|
| Identity & authorisation | Authentication, permissions and approval primitives. | Inputs and alternatives to consider when enforcing customer-specific actions. |
| Fraud & voice detection | Signals about caller identity, fraud risk and synthetic audio. | Inform risk; a voice score alone does not authorise a transaction. |
| Testing & monitoring | Find weaknesses and observe behaviour before and after launch. | Complement checks that run before a protected action executes. |
| Ringfence | Caller assurance, per-action approval and a decision trail. | A policy enforcement point on the voice agent’s tool-call path. |
Build with us
For teams giving voice agents real responsibilities.
Agencies and voice-agent builders
Help shape the gate around the tools your clients depend on. Our integration targets include Vapi, Retell, Bland and LiveKit; integrations are in development.
Start with an authorised assessment
We’re seeking design partners for a free, scoped red-team assessment using controlled test data and actions. Identify which tools need stronger checks, then shape the enforcement workflow with us.
Early access
Put your sensitive tools behind a policy.
Tell us what your agent can do and where approval gets difficult. We’re seeking builders to shape the prototype and its red-team workflow.
Request early access
Contact details are being set up. You can explore the illustrative scenario above or open the narrated demo with reviewer credentials.
Explore the previewThis page does not collect or store your information.