Give your staff an approved AI chat. Protecto masks customer names and account numbers before a prompt reaches the model, then restores them for employees allowed to see them.
SOC 2 Type II · ISO 27001 · Runs in your bank’s cloud
Bank Muscat lets employees use Gemini on bank content, with Protecto in its own Oracle Cloud environment
Recall on sensitive data, including Arabic
Precision
To a working proof of concept
Give employees one approved chat run by the bank, connect it to ChatGPT or Gemini, and put a data layer between the chat and the model. Customer details become tokens before the prompt leaves the bank, and real values come back only for staff whose role allows it.
An approved chat interface, signed in with your corporate identity, so nobody needs a personal account.
Prompts, uploaded files and retrieved documents pass through Protecto before they reach the model.
Names, account numbers, card numbers and IDs become consistent tokens. The model sees tokens only.
The answer comes back with real values for employees whose AD or Okta group allows it. Others see tokens.
Chat history stays on your infrastructure. DLP and web controls keep covering email, files and personal AI accounts.
A relationship manager gets a complaint from a long-standing client. She pastes the email into an AI chat and asks for a polite reply. The email holds the client’s full name, account number and a note about a missed loan payment. All of it goes to the model provider in one block of text, and it stays in the chat history on the provider’s servers.
This is the customer data problem in generative AI for banks: the tools work, but the data can’t go where they run.
Staff who can’t use an approved tool use a personal account, and the bank loses sight of it entirely.
They stop the provider training on your prompts. Customer details are still sent and stored under its retention terms.
DLP looks for a card number in an attachment. Here the data is plain text inside an API call.
Only one keeps customer data away from the model provider and still gives staff useful answers.
| Approach | Customer data reaches the provider? | Answers still useful? | Staff use it? |
|---|---|---|---|
| Ban AI chat | No, on paper. Staff move to personal accounts. | No approved tool to use | Not the approved tool |
| Enterprise ChatGPT or Gemini licenses | Yes. Real values are sent and stored under the provider’s retention terms. | Yes | Yes |
| DLP and redaction | Partly. DLP was not built for AI traffic. | Often not: “[REDACTED] missed a payment on [REDACTED]” | Teams push back when work breaks |
| Approved chat with Protecto | No. The model receives tokens only. | Yes. Tokens keep the meaning of the data. | Yes |
Independent benchmark: UT Dallas and DataXpert measured a higher F1 for Protecto than for AWS Comprehend and Microsoft Presidio.
Protecto sits between your chat interface and the model and acts at four points in every conversation. Who sees real values comes from the AD or Okta groups you already manage.
Names, account numbers, card numbers and IDs are replaced before the prompt leaves the bank.
Uploaded files and SharePoint passages are masked before they enter the prompt.
When the assistant calls a customer lookup, Protecto resolves the token at the tool boundary.
Employees allowed to see customer data get real values. Everyone else sees tokens.
The employee types
Draft a reply to Fatima Al Balushi about the late fee on account 0471 5520 1938.
ChatGPT or Gemini receives
Draft a reply to <PERSON>Kq7Rm2</PERSON> about the late fee on account <BANK_ACC_NO>8830 1146 2075</BANK_ACC_NO>.
The answer, by role
Dear Fatima Al Balushi, thank you for raising the late fee on account 0471 5520 1938. We have reviewed it and…
Dear <PERSON>Kq7Rm2</PERSON>, thank you for raising the late fee on account <BANK_ACC_NO>8830 1146 2075</BANK_ACC_NO>. We have reviewed it and…
Unmasked: her AD group allows customer data. The reveal is logged.
Stays masked: this role works with trends, not named customers.
On-premises · Private cloud · Air-gapped · SaaS. When Protecto runs in your environment, token mappings, keys and logs stay on your servers.
Masking decides what the model sees. Where Protecto runs decides where the real data lives.
Protecto deploys inside your environment. Bank Muscat runs it in its own Oracle Cloud environment.
With a bank-run chat, the full conversation stays on your infrastructure. The model provider only holds masked values.
Detection is tested in English, Arabic, French, German, Spanish, Dutch and Japanese.
“Protecto enabled us to deliver data security, privacy, and compliance for our AI features at scale across 5,000+ enterprise clients, strengthening customer trust in our AI automation solutions.”
DLP still guards email, endpoints and file transfers. Protecto covers what DLP was not built for: customer data moving into AI prompts, files and tools.
| Control you already run | Keeps doing | Protecto adds |
|---|---|---|
| DLP, such as Microsoft Purview | Guards email, endpoints and file transfers | Customer data inside AI prompts, uploaded files and retrieved documents |
| Web gateway or CASB | Steers staff away from personal AI accounts | An approved chat good enough that staff don’t look elsewhere |
| Active Directory or Okta | Defines who each employee is | Through Privacy Gateway for AI Chat, decides who sees real customer values |
| SIEM | Runs your security monitoring | A log of every reveal of an original value |
Protecto doesn’t watch what employees type into public AI websites on their own devices. Keep your web controls for that, and give staff an approved chat that does the job.
Start with one team and one use case, check the answers on your own data, then widen.
Step 1
Complaint replies, policy questions and summaries are common starts. Choose which entities are masked and which roles see real values.
Step 2
Connect your chat interface, your model accounts and your AD or Okta groups.
Step 3
Check answer quality and masking on your own data, in your languages.
Step 4
Connect sources such as SharePoint so answers draw on bank documents, masked on the way in.
Enterprise plans stop the provider from training on your prompts and add admin controls. Customer details you send are still transferred to the provider and stored under its retention terms. With Protecto in the path, the provider receives tokens instead of real values.
No. DLP keeps guarding email, endpoints and file transfers. Protecto covers customer data moving into AI prompts, files and tools, which DLP was not built to handle.
Protecto works through the chat interface your bank runs, not public websites. Most banks give staff an approved chat and use web controls to steer people away from personal accounts.
Protecto is model agnostic. You can connect ChatGPT models, Gemini, Claude or models you host yourself, and pick the one that gives your teams the best answers.
Protecto Gateway is designed to detect and mask supported entity-based sensitive data, including common personal, health, payment, account, government, and customer identifiers, before protected chat content reaches the model provider. Custom entities, such as an internal customer ID format, can be added.
Yes, through Privacy Gateway. You keep each user’s masking and unmasking access in AD or Okta groups, and Gateway issues a Protecto auth key with that access at login. If you use Protecto Vault directly, it’s a headless API, so your team can build the integration around it.
Gateway latency depends on prompt size, entity density, selected model, synchronous versus asynchronous masking, and whether tool arguments require resolution. During deployment, Protecto measures median and tail latency against the customer’s representative prompt sizes and workflows.
LibreChat integration is available today, including sign-in with your corporate identity. Open WebUI compatibility is in validation.
See Protecto mask customer data in a live chat, on prompts like the ones your teams write.
A third-party study compares how accurately Protecto, Microsoft Presidio, and AWS Comprehend find personal and health data in unstructured text.
By submitting you agree to Protecto’s privacy policy.
