Skip to main content
BAN-01 · Banking & Capital Markets

Relationship Manager Intelligence Copilot

Outcome​

Prepare a relationship manager for every client conversation by merging holdings, transaction behaviour, CRM history, and approved research into a cited briefing, with suitability review before any recommendation is shared.

Portfolio attributeValue
Primary personaRelationship manager
Primary aggregateClientRelationship
Cosmos DB containers2
Grounding dataCRM, holdings, transactions, approved research
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityClient preparation time10-18% reduction(baseline median preparation minutes - assisted median preparation minutes) / baseline median preparation minutes * 100
Quality or riskFirst-pass acceptance5-10 percentage-point improvementassisted percentage accepted without material rework - baseline percentage accepted without material rework

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: Relationship managers can prepare cited account narratives faster without attributing direct revenue or suitability decisions to the assistant.

Pilot guardrail: Keep advice within approved CRM, holdings, transaction, and research evidence; preserve suitability controls.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure client preparation time and first-pass acceptance with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Assemble relationship and holdings contextBriefing serviceClient group id and upcoming meeting → Consolidated relationship snapshotPolicy controlled
2. Detect portfolio drift and cash build-upAnalytics agentPositions versus mandate targets → Drift and concentration findingsPolicy controlled
3. Ground ideas in approved researchResearch retrieval agentFindings and client objectives → NextBestActionCandidate set with citationsPolicy controlled
4. Run suitability and conflict screeningSuitability engineCandidate actions and client risk profile → Screened candidates with blocking flagsPolicy controlled
5. Approve the client-facing briefingRelationship manager and supervisorScreened briefing pack → Released briefing and logged advice recordHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
clientRelationships/tenantId, /clientGroupIdClientRelationship, PortfolioPosition, InteractionNote, NextBestActionCandidate/summaryVectorEmbedding, /rationaleVectorEmbedding, /summaryVector, /rationaleVector
approvedResearchLibrary/tenantId, /researchIdResearchNote/bodyVectorEmbedding, /bodyVector

Query contract​

QueryContainerScopePurpose
Relationship briefing assemblyclientRelationshipssingle-partitionSingle partition read of relationship, positions, and recent interactions for the meeting pack.
Concentration exposure checkclientRelationshipssingle-partitionIdentify positions breaching single-issuer or asset-class concentration limits.
Approved research retrievalapprovedResearchLibrarycross-partition-boundedKnowledge-container query against approvedResearchLibrary (bounded fan-out across the /tenantId + /researchId partitions named by the shortlist): restrict cited commentary to research that passed compliance approval, as a bounded fan-out over the research ids returned by vector shortlisting.
Pending suitability review queueclientRelationshipscross-partition-boundedBounded cross-partition control-plane query: list proposed actions waiting on supervisor sign-off across the whole coverage book. It cannot carry the second partition key because a supervisor queue spans every client group the desk covers, so it runs as a tenant-scoped scheduled sweep off the interactive path, bounded by the pendingReview state and a 14-day createdUtc window, executed with MaxItemCount paging and continuation tokens.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/banking-capital-markets/relationship-manager-intelligence-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators