Financial Services Agents: Compliant Copy, Product Explainers & Human Review

In production, a financial-services Agent usually does three jobs only: explain products and processes from effective documents, read-only status lookups on an allowlist, and route high-risk wording into human review before send. It does not: give investment advice, promise returns, auto-approve credit/disbursement, or change fees without approval. GeonAI delivers inside your compliance boundary; /agents (e.g. Data Privacy Consultant) are capability references only—not a public trial of all 363+ presets. Email [email protected], subject “Enterprise Agent inquiry”. KB guardrails: /blog/enterprise-rag-knowledge-base-agent. Private deploy is common in finance: /blog/deepseek-private-deployment-guide.

Nail four boundaries in the SOW first

Without these four, “compliant copy / product explainer / human review” stay verbal and fail acceptance disputes.

1) Compliant copy = corpus + hard refuse—not softer tone

A workable compliance layer needs three assets:

1. Approved corpus: FAQs, risk-disclosure templates, complaint scripts—with version and effective date; the Agent cites, it does not “polish into a new sales pitch”
2. Don’t-say list (hard refuse): capital guarantee, promised return ranges, insider hints, competitor disparagement, undisclosed fees—on hit, fixed refuse + handoff; no rewrite around the block
3. Mandatory disclosure blocks: for wealth/insurance/loan products, append the line’s standard risk notice (template ID in the audit log)

2) Product explainers: effective-version filter + field-level allowlist

Prospectuses, fee sheets, and policy wordings change often. Retrieval must carry product_id + effective_from/to; obsolete revisions may be auditable but must not answer clients. Fix chunk metadata: product code, channel (app/branch/phone), audience tags, dual-recording required?, suitability required?.

• Auto-answer OK: definitions (cooling-off), document checklists, steps, public fee tables, branch hours
• Draft only: personalized illustrations (“if you hold 500k for a year…”)—numeric outputs enter review
• Refuse → human: suitability conclusions, tax planning, cross-border compliance, complaint adjudication, unpublished internal quota rules

Illustrated calculations are a top failure mode: models turn assumptions into promises. Prefer controlled formulas or fee-table lookup, watermark “assumptions,” and default to human review. RAG refuse/citation detail: /blog/deepseek-rag-knowledge-base.

3) Human review as a state machine—not “call someone when it breaks”

Coexistence means every outbound reply has a state. Minimal machine: `draft` → `auto_allowed` (low-risk) or `pending_review` → `approved` / `edited` / `rejected` → `sent` (channel + message id). Reviewer UI must show: client question, proposed reply, hit doc IDs/versions, which rule fired, one-click edit-and-send.

1. Reviewers via SSO; log who/when/what changed
2. Rejects need reason codes (wrong promise / missing disclosure / stale doc / tone)—feed don’t-say and corpus
3. Peaks: review only high-risk intents; sample the rest; rule changes need compliance-signed versions

Audit fields to lock before go-live

• session_id / customer_token (masked) / channel
• intent_label, risk_flags[], refuse_reason
• retrieved_doc_ids[], doc_versions[], disclosure_template_id
• output_state, reviewer_id, edit_diff_hash, sent_message_id
• model_route, prompt_template_version (hash sensitive payloads if needed)

Retention and access follow internal control and local regulation; engineering first ensures complete fields, tamper evidence, session replay. General enterprise checklist (classification + audit + internal access): /blog/enterprise-ai-compliance-overview.

Integrations (common in finance)

Pilot acceptance: numbers, not “better experience”

• Wrong-promise rate (guarantee/return/fee): sampled near zero
• Missing-disclosure rate: required risk notice not appended
• Missing-citation rate: product answers without doc_id
• Review latency: P50/P90 on pending_review
• Missed-handoff count: should-be-human but auto_allowed (QA finds)
• Repeat-contact rate and FCR on process-only intents

Sensible pilot slices (verifiable in 8–12 weeks)

Do not start with “full-book robo-advisor.” Safer cuts: (1) one retail product line—prospectus Q&A + checklist; (2) card progress/loss guidance (read-only + official links); (3) branch hours / basic appointment copy. Then expand to calculated drafts under review. Delivery: /blog/enterprise-ai-agent-delivery-4-steps. Pricing factors: /blog/custom-ai-agent-pricing-factors.

What to bring to GeonAI

Bring: priority product lines and channels, current compliant scripts and don’t-say list, prospectus samples (with versions), review org and SLA, private/VPC needs, read-only system list, target acceptance metrics. Email [email protected], /pricing, or Live chat. We design the state machine and KB ACL to your compliance outcome—not a “fully automated finance CS” slogan.