top of page

Banking Knowledge Agent

A Tier-1 bank needed a dependable knowledge agent for frontline and back-office teams. The system had to answer policy, product, and compliance questions in EN/FR/DE while citing authoritative sources. Early pilots “sounded right” but occasionally missed policy nuance, produced inconsistent tone across languages, and slowed under peak branch and contact-center traffic. We introduced a reliability program focused on grounded answers, multilingual consistency, and predictable performance.

Objectives

  1. Lift grounded-answer rate on policy and product queries.

  2. Reduce policy variance across languages and channels.

  3. Meet a p95 latency budget aligned to agent handle time.

  4. Establish ongoing oversight for post-release changes and new policies.

Our Solution

We mapped the bank’s policy corpus and product docs into a structured RAG pipeline with freshness guarantees and locale tags. Retrieval was tuned for precision on policy clauses and exception rules; generation templates enforced citations, disclaimers, and tone. We created multilingual prompt packs and evaluation sets, aligning phrasing with brand voice and regulatory expectations.

Guardrails covered PII, suitability language, and jurisdiction-specific restrictions. Scorecards tracked groundedness, policy adherence, latency, and cost by route and language. Canary evaluations on anonymized conversation logs caught drift from new circulars before agents felt it.

Implementation Highlights

1. Policy-aware retrieval: Clause-level chunking with metadata for jurisdiction, effective dates, and exception handling.

2. Multilingual normalization: Locale-specific prompts and tone rules; code-switch handling for agent queries.

3. Source-first generation: Mandatory citations with inline anchors; uncertainty language when policy gaps were detected.

4. Performance envelope: Intent-based routing to smaller models for routine lookups and larger models for complex guidance.

5. Safety posture: PII redaction checks, high-risk topic escalation cues, and calibrated thresholds to minimize false blocks.

Before → After

Before: “Early repayment fee applies,” no clause cited; ambiguous exception for vulnerable customers.

​

After: “Early repayment fee does not apply for vulnerable-status clients per Policy 7.3(b) (source: Retail-Credit-Handbook, rev. 04-2025). If eligibility unclear, route to the Vulnerable Customer process. Confidence: high.”​

Collaboration

Compliance and product owners defined acceptance thresholds; support leads validated tone. We shipped in tight loops—one change per metric—so improvements were traceable and safe. Weekly digests summarized movement, regressions, and next actions.

Next Steps

Expand coverage to wealth and SME lending, add automatic policy-diff notifiers that generate new gold cases, and trial retrieval-time adapters for seasonal product changes.

bottom of page