Cloud AI, without the personal data
Run Blits.ai in your own data center or private cloud and still use the best cloud models for language, speech and voice. The privacy shield replaces every name, account number and id by a placeholder before anything leaves your environment, and puts the real values back before your customer sees the answer.
How it works
Three steps, all inside your own environment. Your customer notices nothing.
1. Detect
Before a message leaves your environment, the shield finds the personal parts: a name, an IBAN, a card number, a national id, an email address, a date of birth. It uses exact checks, such as the IBAN and card number checksums, so an order number is never mistaken for an account.
2. Replace
Each personal value becomes a typed placeholder: [PERSON_1], [IBAN_1]. The cloud model knows it is dealing with a person and an account number, and can reason about them, but never sees which. The same value keeps the same number for the whole conversation.
3. Restore
The model answers with the placeholders. The shield puts the real values back before the answer reaches the chat, the voice or a connected system such as your CRM. Your customer reads and hears a normal answer.
What the cloud sees
One turn in a banking assistant. Left, the conversation as your customer has it. Right, what the cloud model actually received and returned.
Your customer
Customer: Hi, I am Jan de Vries. Can you link my IBAN NL91 ABNA 0417 1643 00 to my profile?
Assistant: Thank you Jan de Vries. I have linked NL91 ABNA 0417 1643 00 to your profile. Would you like a confirmation by email?
The cloud model
Customer: Hi, I am [PERSON_1]. Can you link my IBAN [IBAN_1] to my profile?
Assistant: Thank you [PERSON_1]. I have linked [IBAN_1] to your profile. Would you like a confirmation by email?
The same applies to everything else the model needs: the conversation history, results from your systems, knowledge base content and customer profile data all pass through the shield. A tool the assistant calls, such as a lookup in your core banking system, receives the real value again, because it runs inside your environment.
Everything the model needs passes through the shield
A placeholder in the question alone is not enough. Personal data also arrives through the history, through your systems and through the knowledge base, so every path to the cloud model is covered.
The message
What the customer types, or the transcript of what they said, is shielded first. The model, the sentiment check and the guardrails all work on the shielded text.
The conversation history
Earlier turns are sent to the model with placeholders. The same person keeps the same number for the whole conversation, so the model can refer back.
Your systems
A lookup in your core banking system or CRM receives the real values, because it runs inside your environment. Its result is shielded again before the model reads it.
The knowledge base
Retrieved documents and the search query are shielded too, so a document that mentions a customer never leaks through the back door.
The verified profile
When a customer is verified, the platform knows their name, address, email and phone. Those exact values are shielded wherever they appear, in any spelling.
Long term memory
Memories extracted from conversations are built on the shielded text. A memory that would still contain personal data is never stored.
The standard set
Ships with the platform, kept up to date, verified with check digits and context. You switch each type on or off per tenant and decide what the model may see instead: a placeholder, a safe hint, or the value.
| Type | What the model sees | How it is recognised |
|---|---|---|
| Bank account | [IBAN_1] | IBAN with the international checksum verified, so an order number is never mistaken for an account. |
| Card number | [CARD_1] | 13 to 19 digits that pass the Luhn check, with or without spaces. |
| Phone number | [PHONE_1] | Valid phone numbers for the language of the conversation and international numbers. |
| National id | [NATIONAL_ID_1] | Dutch BSN, Belgian national number, German tax id, Saudi national id, Emirates ID and Qatar ID, each with its own check. |
| Email address | [EMAIL_1] | Any email address. |
| Date of birth | [DATE_OF_BIRTH_1] | A date next to a birth cue such as born, geboren, date of birth or geburtsdatum. Other dates stay readable. |
| Passport number | [PASSPORT_1] | A number next to a passport cue. |
| Address | [ADDRESS_1] | Street names with a house number in Dutch, German, English, French, Spanish, Italian and Arabic, and the verified profile address. |
| Person name | [PERSON_1] | The verified profile name, names after cues such as my name is, ik ben, ich heiße or اسمي, and first names from the built in lists for seven languages or your own lists. |
| Sensitive answer | [SENSITIVE_1] | Any answer to a dialog question you mark as sensitive is shielded as a whole. |
Add your own patterns on top, for example a policy number or an internal customer id, and give them their own placeholder. Words that must never be hidden, such as your product names, go on an allow list; words that must always be hidden go on a deny list.
What your administrators see
The shield is not a black box. Every turn can be checked, and personal data stays out of sight unless someone with the right permission has a reason to look.
Placeholders in the conversation log
The log shows [PERSON_1] and [IBAN_1] with their type. The real values stay in the encrypted vault of the conversation, on your own database.
Reveal with a reason, written to the audit log
A user with the permission Reveal personal data can reveal one message after entering a reason, for example a complaint number. Every reveal is written to the audit log with who, when and why.
Proof per turn
The agent trace and the decision log record what the model received, which types were replaced, how long detection took and whether the final check caught anything. Never the values themselves.
Show what the AI saw
In the test chat an administrator can open a panel that shows, per turn, exactly what the cloud model received and answered. Real channels never expose it.
Built for your own infrastructure
Detection never calls the cloud
The shield runs inside the Blits.ai backend on your servers. No text is sent anywhere to decide what is personal. It adds a few milliseconds to a turn.
Values stay in your vault
The link between a placeholder and its real value is stored encrypted with the conversation, on your own database, and is erased with it. Nobody at Blits.ai or at a model provider can look it up.
Proof per conversation
Every turn records what the model saw and which types were replaced, never the values themselves. Administrators see placeholders in the conversation log; revealing a value needs a dedicated permission and is written to the audit log.
Voice
Speech needs two honest choices, and you make both per assistant.
When the assistant speaks
To read an account number aloud, the voice provider has to know it. You choose: speak the real value, speak only a safe hint such as "the account ending in 6300", or use a self hosted voice for that sentence.
When the customer speaks
Cloud speech recognition hears the audio before any text exists, so the shield cannot act there. The assistant shows a short notice next to the microphone, and says it once on a phone call, asking customers to type sensitive details. For the strictest environments, Blits.ai also runs speech recognition on your own servers.
Deployment
Built for teams that run the platform on their own infrastructure, with as little extra as possible.
No extra container
Detection runs inside the Blits.ai backend on your servers. No separate model service to size, patch or keep available.
On by default on premises
An on premises distribution switches the shield on for every tenant with a single environment setting. In the managed cloud a tenant switches it on in the Privacy shield menu.
Works with what you have
Assistants that used the earlier masking keep masking, now through the shield, with their own patterns carried over. Old conversations stay readable.
Questions about the privacy shield
See the shield on your own data
Bring a transcript from your own customer service. We show you what the cloud would see.
Book a demo