Applying different redaction rules per Azure OpenAI deployment for retail banking and wealth management
Treat redaction as a policy bound to each Azure OpenAI deployment and enforced at runtime on both prompts and responses. Classify the data each deployment handles, write a separate versioned rule set for each, place an enforcement layer in the request path that resolves the rule set from the deployment being called, and record the deployment, rule version and action taken (allow, redact or block) for every interaction.
Bind each rule set to a deployment and enforce it in the request path
- Caller sends a request addressed to a deployment
- Enforcement layer resolves the bound rule set and version
- Prompt is evaluated: allow, redact or block
- Redacted prompt is forwarded to the Azure OpenAI deployment
- Response is evaluated with the same rule set
- Rule, version, deployment and action are recorded
Test rule behavior before each release
- Seed test prompts with synthetic values for every data class the deployment handles
- Confirm redacted output preserves enough context for the task to remain useful
- Test responses separately, including cases where the model repeats or reconstructs a value
- Test unbound and misconfigured deployments to confirm requests fail closed
- Re-run the full suite whenever a rule set version changes
- Store results with the rule set version they validated
Why one redaction setting fails across banking workloads
Retail banking and wealth management teams usually run separate Azure OpenAI deployments for different jobs: a customer service assistant, an advisor support tool, an internal document summarizer. The data each one touches differs. Retail workloads see account numbers, card numbers, national identifiers and transaction details. Wealth workloads see portfolio holdings, client financial circumstances and advisor notes.
A single organization-wide setting has to pick one point on a spectrum. If it is strict enough for the customer service assistant, it strips the holdings and context an advisor tool needs to be useful. If it is tuned for the advisor tool, it lets identifiers through to a customer-facing deployment where nothing justifies their presence. Neither outcome is defensible to a model risk management reviewer.
Native platform features help with adjacent concerns. Microsoft describes Azure OpenAI as offering private networking, regional availability and content filtering. Those address network exposure and content categories. A redaction rule is different: it is defined by your own data classification and must differ by deployment, so it needs to be bound and enforced by a layer you control.
Classify data by deployment and define separate rule sets
Start with an inventory per deployment, not per organization. For each deployment, list the data classes that can appear in prompts, in retrieved context and in model responses. Include sources the user does not type directly, such as documents attached for summarization.
The identifier examples in the SEC's Regulation S-P final rule are a useful seed for national and government identifiers: Social Security numbers, driver's license or identification numbers, alien registration numbers, passport numbers, taxpayer identification numbers and biometric records. Add the classes specific to your business, such as account and card numbers, portfolio holdings, net worth and advisor notes. Then decide an action for each class in each direction. The examples below are design choices, not requirements.
- Customer service assistant (retail): Block prompts containing national identifiers. Redact full account and card numbers in prompts and responses, retaining only a masked form if the workflow needs one.
- Transaction inquiry assistant (retail): Allow transaction descriptions needed for the answer. Redact counterparty identifiers and account numbers. Block any response that echoes an unmasked card number.
- Advisor support tool (wealth): Allow portfolio holdings the advisor is entitled to see. Redact client names and national identifiers. Redact net worth unless the use case requires it.
- Internal document summarizer: Redact advisor-note identifiers before summarization. Block documents that contain national identifiers or biometric records and route them for review.
Evidence for examiners and model risk reviewers
Reviewers will want to know which rule applied to a given interaction, not that a rule existed. For each interaction, retain the deployment identifier, the rule set and version, the data classes detected, the action taken (allow, redact or block), the direction (prompt or response) and a timestamp. Keep exception approvals and test records alongside, so a reviewer can trace from a single interaction to the rule version and the tests that cleared it.
This is where enforcement and after-the-fact review differ. A review of sampled transcripts can suggest a rule may have been missed. Evidence generated by the enforcement step itself shows what the policy was and what it did at the time. NIST AI 600-1 lists governance and incident disclosure alongside pre-deployment testing, and records of this kind support all three. NYDFS issued guidance on cybersecurity risks arising from artificial intelligence on October 16, 2024, and teams subject to NYDFS should read that guidance directly when mapping evidence to their obligations.
Set retention and data residency for this evidence deliberately. Evidence about redacted interactions should not itself become a store of the sensitive values the rules removed.
Where Trussed fits
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-approval actions applied to individual interactions. Enforcement is at runtime, not reviewed after the fact, and it generates audit-ready evidence mapped to frameworks including NYDFS and NIST. It deploys self-managed in the customer cloud, with configurable retention and data residency.
Trussed runs in production in regulated enterprises, with customers in insurance, pharma, healthcare and education. The pattern in this guide does not depend on any one product. Whatever layer you choose, evaluate it against the same questions: can it bind distinct rule sets to distinct deployments, does it cover prompts and responses, and does it produce per-interaction evidence of the policy applied and the action taken.
Per-deployment redaction path
Classify
Map data classes to each deployment and use case.
Bind
Attach one versioned rule set to each deployment.
Enforce
Apply allow, redact or block on prompts and responses at runtime.
Evidence
Retain the rule, version and action for each interaction.
Review your deployment-level redaction design
Walk through your deployments, data classes and evidence requirements with a Trussed expert.
Talk to an Expert