Docs
Book a demo
Banking

How Bank Muscat kept customer data in the bank while adopting Gemini

Bank Muscat let employees use Gemini while customer data stayed under the bank’s control. Protecto found sensitive values in Arabic and English with 99% recall, and the proof of concept worked in four weeks.
99%
Recall on sensitive data, incl. Arabic

Get the full case study

We email the PDF to your work inbox, so you can forward it to your team.

By submitting you agree to Protecto’s privacy policy.

Bank Muscat wanted employees to use Google Gemini for everyday work and for questions over internal banking knowledge. Customer and financial data had to stay under the bank’s control. Protecto ran inside the bank’s environment, found sensitive values in Arabic and English, and sent Gemini tokens instead of real data.
99%
Recall on sensitive data, incl. Arabic
96%
Precision, so little is over-masked
85%
Semantic accuracy retained after masking
4 weeks
To a working proof of concept

The challenge

Gemini was useful, but customer data had to stay in the bank

Bank Muscat wanted Gemini for drafting, research, summaries and questions over internal documents. Sending customer records straight to a cloud model was ruled out: data residency rules and the bank’s own policy required customer and financial data to stay under bank control. Employees also could not be asked to clean every document by hand before using AI.

  • Mixed Arabic and English text with transliterated names and Arabic numerals
  • Messy real-world input: typos, spacing problems and numbers fused with the next word
  • Many entry points: prompts, pasted text, uploads and SharePoint content
  • Useful answers: blunt redaction would leave Gemini guessing
  • DLP and file permissions could block a file or let it through, but could not decide what each AI answer may reveal
Banking
99%
Recall on sensitive data, incl. Arabic
The solution

Protecto inside the bank, in front of Gemini

Protecto ran in the bank’s Oracle Cloud environment and sat between Gemini and the bank’s data, covering direct employee use and content retrieved from SharePoint.

01

Detect in Arabic and English

  • Finds names, account numbers, card numbers and IDs in prompts, uploads and documents
  • Handles transliterated names, Arabic numerals and values fused with other text
02

Tokenize and keep the context

  • Replaces each value with a consistent, format-preserving token
  • Gemini can still search, summarize and answer over the protected text
03

Apply policy for each recipient

  • Gemini receives tokens only
  • Real values return only to the roles the bank’s policy names, and each release is logged
04

Run inside the bank

  • Deployed in the bank’s own cloud environment
  • Detection, token mappings and keys stay under bank control

Want the full story? Get the case study PDF in your inbox and share it with your team.

Get the case study

The results

Results

Bank Muscat chose Protecto after evaluating several vendors on detection quality, answer quality and deployment inside the bank.

Key achievements

  • 99% recall on sensitive data in mixed Arabic and English text
  • 96% precision, so little is masked by mistake
  • 85% semantic accuracy retained after masking
  • A working proof of concept in four weeks, with little effort from the bank’s internal team
  • Employees can use Gemini over internal content without cleaning documents by hand

More case studies

Browse all case studies

API
Detection and masking for telemetry
Enterprise cloud and software

Protecting Sensitive Telemetry for AI and Analytics

A large technology company wanted to use operational telemetry and partner data for internal AI, analytics and model improvement. Customer identifiers and other sensitive fields complicated the use of raw data.
Read the case study
100%
Compliance Achieved
AI Agent

How a Fortune 100 Giant Put Secure Guardrails on its AI—Without Slowing It Down

See how a Fortune 100 leader put guardrails on multi-agent AI—enforcing zero-trust data policies, retaining AI reasoning, and meeting GDPR compliance without added latency.
Read the case study
13M
Long-form texts processed daily
SaaS

Helping SaaS Leader Process 13M Texts Daily for AI Training

Learn how a leading SaaS company achieved 90% cost reduction in data masking while protecting sensitive PII/PHI across 13M daily texts to power their AI innovation.
Read the case study
Book a demo

See Protecto on your own data

Book a demo to see how Protecto finds and masks sensitive data before it reaches any AI model.

Book a demo
Talk to sales
New: run any AI model without exposing real values.
See how it works