Your Bedrock guardrail isn't blocking bad requests. It's blocking the word "finance" — and your customers are paying for it.
AWS's own guidance on Denied Topics makes the failure mode embarrassingly clear: teams write a topic called "Finance" and then act surprised when the assistant refuses to tell a customer when their direct deposit clears. The guardrail did exactly what it was told. The problem is what it was told.
The subject is not the sin
A banking assistant has to talk about money. That's the job. The thing that should be refused is a specific action — a personalized recommendation to buy, sell, or hold a security — not the entire domain of finance. AWS's AnyCompany Bank scenario shows all four failure modes in one fictional deployment:
- Finance blocked a direct-deposit question.
- Legal blocked a mortgage-modification request.
- Fraud blocked a customer reporting their card was skimmed.
- Competitors blocked customers asking comparison questions.
Every one of those is a legitimate customer doing exactly what the assistant exists for. A guardrail that blocks your own service path isn't safety — it's a broken product with a policy label on it.
Renaming isn't fixing
Here's the part most teams will miss: renaming "Finance" to "Investment Recommendations" changes nothing if the definition still covers "anything related to investing or personal finance decisions." AWS's own worked example shows the intermediate definition still matching mutual-fund education, historical financial news, and requests to view holdings. The label got sharper; the boundary didn't.
The fix has to run through the whole topic — name, definition, examples, and neighboring topics — as one consistent statement of what the user is trying to get the assistant to do. AWS's template is worth stealing: [action] for [specific objects] by [qualifying constraint]. "Asking for a personalized recommendation to buy, sell, or hold specified investments for the user's own returns" is a boundary. "Finance" is a vibe.
The details that actually bite
- Definitions are evaluated semantically, not by keyword. Five paraphrases of one idea leave the boundary unguarded; examples naming specific countries or products drag those entities into the topic's context.
- Longer definitions aren't better. AWS explicitly warns they can become more sensitive to the wording of neighboring topics. Filling the 1,000-character limit is not precision.
- Overlapping topics interfere. Two individually plausible descriptions can combine to block inputs neither would block alone. Review the policy as a set, not a pile of sentences.
- Keep a must-pass list next to your must-block list. The fraud example is the cleanest illustration: investigating a fraud case is out of scope; reporting fraud for human routing is the assistant's job. If you only write down what to refuse, you'll refuse the work too.
The honest caveat
AWS is also straight about this: results aren't fully deterministic, and the guidance offers no measured accuracy improvement. Better framing is a supported authoring practice, not a guarantee. The operational question stays the same — does the policy refuse the intended behavior while preserving the interactions the assistant exists to support? If you can't answer that with a must-pass list in hand, your guardrail is a guess with a dashboard.
Go deeper
Read the full breakdown — the AnyCompany Bank case study, the revised topic definitions, and the four authoring problems AWS flags — on Agentic Affairs:
Best practices for Amazon Bedrock Guardrails: Define actions, not domains
Originally published on Agentic Affairs (https://agentic-affairs.com)



