Security and deployment
You choose where it runs.You hold the keys.
Collections Hub ships as containers you run on your own servers or cloud account, behind Microsoft sign-in, with AI services billed to your own keys in the region you pick. The models that read every call afterwards run locally with no internet access, and the voice engine can run on self-hosted speech and language models where a deployment calls for it.
Reference deployment
Standard topology,drawn honestly.
What runs inside your boundary, what it calls out to on your keys, and what never leaves.
What stays with you
Customer records, recordings, transcripts and the evidence chain live in your database and storage. Personal-data detection, intent, sentiment and beep timing run on four local models with no internet access.
What calls out, on your keys
Live speech recognition, voices and language models run on Azure OpenAI and Azure Speech under your own subscription and region by default. The voice engine can use other providers, or self-hosted models, per deployment.
How it lands
Containers on your servers or cloud account, behind Microsoft Entra ID, with one contract for your core banking and loan data. Nothing starts dialling until an admin switches outbound on.
Deployment models
Three shapes,one product.
| Model | Where it runs | What it suits |
|---|---|---|
| Your data centre | Your servers and your change control, with AI services on your own keys. | Institutions whose records have to stay in the building. |
| Your cloud account | Your subscription, your region, your keys, your network policy. | Teams already running a governed cloud estate who want the same isolation without the hardware. |
| Self-hosted models | The voice engine on self-hosted speech and language endpoints, set up per deployment. | Environments where sending call audio to a cloud AI service is not an option. |
Controls
Data, identityand evidence.
Data
- Residency
- Records, recordings, transcripts and the evidence chain stay in your database and storage.
- Redaction
- Transcripts masked as they are written; recordings beeped over identity and card values. The original plays only for roles cleared for it, and every play is logged.
- Encryption
- Borrower phone numbers, emails and addresses are encrypted in the database, with a key-rotation script. TLS at the edge.
- Outbound events
- No customer data leaves through a webhook until someone reviews the destination and a signed test delivery succeeds. Payloads carry no phone numbers or transcripts.
Identity and access
- Single sign-on
- Microsoft Entra ID. No local passwords, guest and personal accounts refused, and a first-time user gets no role until invited.
- Authorisation
- Role and scope checks on every module and every export. Routes that are not registered are denied, and the build checks the coverage.
- Separation
- Tenant isolation enforced by Postgres row-level security. A deployment will not start with isolation, append-only audit or data encryption switched off.
- Session evidence
- Every export and every playback of a raw recording is logged, with who and when.
Models
- Your keys
- Azure OpenAI and Azure Speech on your own subscription by default. The voice engine can use other providers, or self-hosted models, per agent.
- Local where it matters
- Personal-data detection, intent, sentiment and beep timing run on four pinned models inside your deployment, with no internet access.
- Versioned
- The agent version is recorded on every call, and model versions on every post-call result and planning decision.
- No training on your data
- The engine's own models learn from your book inside your deployment. Cloud AI services run under your own agreement with the provider.
Assurance
What to ask us forbefore you sign.
Ask for the architecture review early. What can be attested depends on the deployment and on who hosts it, which is a real distinction rather than a hedge, so we walk through it for your setup.
What the review covers
- Data flows and sub-processors
- Identity and access
- Tenant isolation
- Redaction and retention
- DPDP subject rights
- Model and prompt versioning
Every service that touches a customer of ours runs on our keys, in our region, under our contract.
Questions
What security teamsask first.
Does any customer data leave our infrastructure?
Records, recordings, transcripts and the evidence chain stay in your database and storage. Live calls use speech and language services on your own keys, in your chosen region, so audio and text go only to the provider you contract with. If that is not allowed, the voice engine can run on self-hosted models instead, scoped per deployment.
Can it run without cloud AI services?
The post-call models already run locally with no internet access. For live calls, the voice engine can be pointed at self-hosted speech and language models. We scope that per deployment, because it changes the hardware you need.
Which telephony does it use?
Yours to choose. The voice engine works with Twilio, Exotel, Plivo, Vobiz, Telnyx, Vonage and Cloudonix, or with your own Asterisk. WhatsApp runs on the WhatsApp Business Platform.
How do people sign in?
With Microsoft Entra ID. There are no local passwords, guest and personal accounts are refused, and a first-time user gets no access until they are invited or approved.
Do you train on our data?
The decision engine's models learn from your book inside your deployment. Cloud AI services run under your own agreement with the provider, on your keys, not ours.
Send us yourarchitecture review.
We would rather answer the long questionnaire early than discover a blocker in month three. Send it over and we will fill it in before the first call.