In fintech, the bar for "AI support" is higher: answers must be grounded, actions must be controlled, and everything sensitive needs a human in the loop. Done right, an agent cuts response times on the 80% of routine questions while never touching the 20% that needs judgment. Here's the pattern.
Where agents help
Account and transaction FAQs ("why was I charged a fee," "how do I update my payout method"), KYC/onboarding status questions, dispute intake, and routing to the right specialist — all grounded in your approved policy docs and answered with citations.
Where agents must stop
Moving money, changing limits, closing accounts, or giving financial/investment advice. These are approval-gated or handed to a person, full stop.
The prompt
You are the support agent for a fintech app. Answer only from approved policy
and help documents; cite the source. Do not give financial, tax, or investment
advice. Never confirm or perform transfers, refunds, limit changes, or account
closures yourself — propose the action for human approval or escalate.
If a message looks like fraud or a security concern, escalate immediately.
Tone: precise, reassuring, no jargon.Tools with controls
- Knowledge search — grounded, cited answers.
- Custom HTTP tool to read (not write) account/transaction status, behind stored-secret auth and rate limits.
- Approval-gated tools for anything that writes — the agent drafts the call; a human approves it in the cockpit before it executes.
- Escalate for fraud, disputes, and advice requests.
The workflow
- Trigger — conversation on web, app, or email.
- KB search → LLM — grounded answer + intent classification.
- Manager branch — routine info → answer; money-movement or risk → gated/escalation path.
- Gated action — e.g. fee reversal under threshold → drafted, approved, then executed; over threshold → human.
- Audit — every step, tool call, and approval is recorded as a reviewable run.
Why the gates are the feature
The approval gate isn't friction — it's the thing that lets you deploy at all. It means an agent can work the queue at speed while your team retains control over every consequential action, with a trail to show exactly what happened and who approved it.
As always: confirm your specific regulatory and data obligations with your compliance and security teams — the platform gives you the controls; your policies decide how they're used.
Next: tool calling with MCP for the read/write tool split and build a triage workflow with AI for the routing.