ADI AGENTS · TRANSACTIONS
Voice-authenticated operations, in seconds
Transfers, enquiries and payments executed inside the app, with authentication and a full record of every step.
The operational jump
From information to execution
Both models handle the same request. What changes is who completes the operation.
An assistant that only informs
ADI transactions agent
Results in production
Impact that is measured
Measured on our own voice stack, with no chained third-party APIs.
How it runs
From instruction to receipt
Authenticated customer
The conversation starts inside the app or on an authenticated channel, with the session already validated.
Request in free speech
The customer says what they want to do, with no menu and no numbered option to pick.
Permitted action
The operation resolves to a registered enum. What is not on the list does not happen.
Confirmation and policy
The agent confirms details and calls your endpoint, which operates under your authorisation policy.
Operation completed
Transfer, enquiry or scheduling executed through HTTP Tools against your APIs.
Verifiable event
A receipt to the customer and an event signed with HMAC-SHA256 to your system.
Capabilities
Executing without opening a gap
Closed action catalogue
Transfer, check balance, schedule a payment. What is not registered does not happen, whatever the request.
No credentials in the agent
The agent never receives the end customer's credentials; it calls an endpoint authenticated by API key or header.
Correctly pronounced amounts
The agent reads "one thousand two hundred thirty-four reais and fifty centavos", not the raw string.
Encrypted call end to end
SIP TLS 5061 with SRTP, including the telephony layer.
VAD barge-in
The customer interrupts and the agent yields the turn, as in a real conversation.
An event per operation
Every executed action produces a signed webhook, traceable in your system.
Feasibility
Connects to what you already have
Core, transactional APIs and telephony connected through a documented interface. No core changes.
Onboarding in 1 to 3 weeks. With an existing SIP trunk, the voice channel goes live in 2 to 5 days.
Inside the rules
Transacting without leaving policy
Authorisation on your side
The endpoint called already operates under your policy; the agent does not decide permissions.
Guardrails in code
Deterministic filters before and after the model, three of them mandatory on any agent.
Per-turn traceability
What the agent said, on which version, under which controls, and which action it executed.
Tokenised PII
Personal data is tokenised before any call to a model provider.
The full cycle
One engine, five agents
One compliance approval covers all five stages. You start with the one that hurts most and expand with no new project.
Origination
Captures, pre-qualifies and converts in free speech, on web or voice.
View →Onboarding
Sign-up and activation with deterministic identity validation.
View →Collections
Negotiates inside the defined policy and logs the commitment in the CRM.
View →Support
Resolves on first contact, with context across all three channels.
View →FAQ
Frequently asked questions
Can the agent execute an operation we did not authorise?
No. Every available action is registered as an enum validated outside the model. What is not on the list does not happen, whatever the customer asks or however they ask it.
Does the agent have access to customer credentials?
No. It calls an endpoint of yours, authenticated by API key or header, which already operates under your authorisation policy.
How do we audit each operation?
Every operation produces a verifiable event, delivered by a webhook signed with HMAC-SHA256, and the interaction stays reconstructable with the agent version and the active controls.
Is voice secure enough to transact?
The call travels encrypted, including the telephony layer, over SIP TLS 5061 with SRTP. Session authentication happens before any amount is mentioned.
Why does 800 ms matter?
Above that threshold the customer notices the wait between instruction and confirmation and starts repeating themselves. The voice stack is our own precisely so latency is controlled in one place.