A bank’s AI platform cleared OSFI B-13 in six weeks.
An enterprise AI platform on Amazon Bedrock, with zero data egress: no data leaves the bank.
The problem
“We want AI. Legal says one byte of customer data off-prem and it’s a newspaper story.”
Canadian Tier-1 bank
What was at stake
- The bank’s internal team had built a RAG prototype (an assistant that answers from the bank’s own documents) on OpenAI’s API six months earlier.
- It worked technically, but compliance had stopped the production rollout cold.
- The data was leaving the building, and no amount of terms-of-service language was going to clear that with OSFI.
What we built
Everything inside the bank’s VPC.
Masking, retrieval and the model all run in the bank’s own environment, and every prompt is logged for the auditor.
Before · the bank’s own prototype, on OpenAI’s API
Bank users
RAG prototype
OpenAI API
Data left the bank’s environment. Compliance stopped the production rollout.
After · rebuilt on Amazon Bedrock, inside the bank’s VPC
Bank users1,300 queries a day
Bank VPC · zero data egress
Zero-trust accessPer-team IAMBedrock guardrailsCloudTrail + CloudWatch
PII maskingAmazon Comprehend23 entity types detected and masked. Reversible for authorized workflows only.
Hybrid retrievalAmazon OpenSearchChunkingEmbeddingsHybrid searchTuned for financial documents.
Amazon BedrockClaude, the reasoning modelOnly masked text reaches the model.
AnswerBack to the userMasked values restored only for authorized workflows.
Audit bucket · S3Every prompt logged
CloudTrail + CloudWatchAudit trail meeting regulator expectations
AuditorRead-only access
Simplified from the production design. 1,300 queries a day, six weeks after the rebuild began.
- 01Multi-layer RAG on Bedrock: chunking, embeddings and hybrid search tuned for financial documents.
- 02Comprehend detects and masks 23 PII entity types, reversible for authorized workflows only.
- 03Zero-trust access: per-team IAM, Bedrock guardrails, CloudTrail and CloudWatch audit trail.
- 04Every prompt logged to an S3 bucket the auditor can read, read-only.
Timeline
Six weeks from rebuild to a cleared review.
Six months before we startedBefore we arrivedThe bank’s team builds a RAG prototype on OpenAI’s API. Compliance stops the rollout.
Six weeks to clearanceRebuild on BedrockMulti-layer PII pipeline, Claude, OpenSearch, and every prompt logged to S3.
OSFI B-13 clearedServing 1,300 queries a day.
In productionThe engagement runs 2024 to the present.
Outcomes
What the bank got.
Compliance
6 weeks
for the AI platform to clear its OSFI B-13 review
Serving 1,300 queries a day within six weeks of the rebuild.
Cost
61%
lower AI running cost than the GPT-4 version would have cost
Data
0
data egress. Customer data never leaves the bank.
Stack
BedrockClaudeOpenSearchComprehendStep Functions
Tell us what you're building.
We reply within one business day.
First call: 30 minutes, on a Tuesday or Thursday, with two engineers, not a sales rep. If we’re the wrong fit, we’ll say so and name a firm that is.