Home · Work

Systems Work · No. 23

Front-Office Banking Copilot

Rajesh Ranjan Mahapatra

A familiar workspace surface over an enterprise-owned reasoning service, so front-line staff receive governed, cited answers drawn from context that a single-vendor assistant cannot reach.

Start reading
Systems WorkCase study

© 2026 Rajesh Ranjan Mahapatra. 6 min read

Front-Office Banking CopilotAll Systems Work
Home · Work

Front-Office Banking Copilot

A familiar workspace surface over an enterprise-owned reasoning service, so front-line staff receive governed, cited answers drawn from context that a single-vendor assistant cannot reach.

Challenge

Front-line staff at a large regulated enterprise prepare for dozens of client meetings a week, each taking the better part of an hour to research across six or more systems.

Architectural move

Six candidate options were scored against ten weighted dimensions covering capability, strategic fit and execution.

Control boundary

Every step carries an autonomy tier and a numeric evaluation gate, feeding the enterprise's certification process.

Outcome

The decision paper drove the build, and the recovered preparation time is modelled as hours per staff per week, which is material across a large coverage population.

FIG.00 // SYSTEM MAPREJECTED PATHSTORE / RECORDGATEAGENTSIGNAL / DATAL0 · SURFACECollaborationworkspaceStaff ask hereAnswers renderhereL1 · BOUNDARYIdentity-federatedcrossingEnterprise-ownedorchestratorPlan · act ·observe · replanL2 · AGENTSThin domain agentsAllow-listedtopologyGoverned toolcontractsL3 · KNOWLEDGEFour projectionsGraph: who isconnectedVector: what thetext saysMarts: how muchLive tier: whatjust changedL4 · GUARDBlocking citationguardrailAn unsourcednumber does notrenderREQUESTFAN-OUTRETRIEVALSYNTHESISSYS/FOBC · FIG.00 · SCHEMATIC · NOT TO SCALE
FIG.00 // SYSTEM MAPSIGNAL / DATAAGENTGATESTORE / RECORDREJECTED PATHL0 · SURFACECollaboration workspaceStaff ask hereAnswers render hereL1 · BOUNDARYIdentity-federated crossingEnterprise-owned orchestratorPlan · act · observe · replanL2 · AGENTSThin domain agentsAllow-listed topologyGoverned tool contractsL3 · KNOWLEDGEFour projectionsGraph: who is connectedVector: what the text saysMarts: how muchLive tier: what just changedL4 · GUARDBlocking citation guardrailAn unsourced number does not renderREQUESTFAN-OUTRETRIEVALSYNTHESISSYS/FOBC · FIG.00 · SCHEMATIC
FIG.00 System map. The workspace stays the surface; the reasoning runs on a system the enterprise owns and verifies on every turn.

The problem

FIG.01 // THE PROBLEMSTORE / RECORDMODELSIGNAL / DATARING 3RING 2RING 1REACHRING 1Workspace estatewhat a single-vendor chat agent reachesRING 2Internal cloud estateperformance reporting, credit-risk workflow, corelendingRING 3External market datalive market feedsMOST DECISION-WEIGHTED DATA SITS OUTSIDE RING 1SYS/FOBC · FIG.01 · SCHEMATIC · NOT TO SCALE
FIG.01 // THE PROBLEMSIGNAL / DATAMODELSTORE / RECORDRING 3RING 2RING 1REACHRING 1Workspace estatewhat a single-vendor chat agent reachesRING 2Internal cloud estateperformance reporting, credit-risk workflow,core lendingRING 3External market datalive market feedsMOST DECISION-WEIGHTED DATA SITSOUTSIDE RING 1SYS/FOBC · FIG.01 · SCHEMATIC
FIG.01 The problem. Meeting prep takes most of an hour across six or more systems, and most of that data lives outside the workspace boundary.

Front-line staff at a large regulated enterprise prepare for dozens of client meetings a week, each taking the better part of an hour to research across six or more systems. The obvious answer, an assistant inside the collaboration workspace they already use, has a quiet flaw: most of the decision-weighted data a coverage team needs lives outside that workspace boundary, in performance reporting, the credit-risk workflow, core lending systems and external market feeds. A single-vendor chat agent reaches one ring of that data and answers a fraction of what the role actually demands.

I wrote the architecture decision paper and designed the recommended hybrid: the workspace stays the surface; the reasoning runs on a system the enterprise owns and verifies on every turn.

Why it was hard

FIG.02 // WHY IT WAS HARDREJECTED PATHSTORE / RECORDGATETOO LITTLEDeclarative topic botreally a rules engine; breaksacross the question rangeDESIGNPlace agency preciselynot everywhere; grounding enforced,not encouragedTOO MUCHFully agentic everywherewould never clear governance→ FAILS THE QUESTION RANGEFAILS ASSURANCE ←TERRAIN EVERY OPTION MUST SURVIVEHZ · 01Three ringsdifferent owners, security regimes andfreshnessHZ · 02Policy fixes placementwhere a model runs is securitypolicy, not preferenceHZ · 03Numbers in meetingsa wrong figure is a conduct problemSYS/FOBC · FIG.02 · SCHEMATIC · NOT TO SCALE
FIG.02 // WHY IT WAS HARDGATESTORE / RECORDREJECTED PATHTOO LITTLEDeclarative topic botreally a rules engine; breaks across thequestion range↑ FAILS THE QUESTION RANGEDESIGNPlace agency preciselynot everywhere; grounding enforced, notencouraged↓ FAILS ASSURANCETOO MUCHFully agentic everywherewould never clear governanceTERRAIN EVERY OPTION MUST SURVIVEHZ · 01Three ringsdifferent owners, security regimes andfreshnessHZ · 02Policy fixes placementwhere a model runs is security policy, notpreferenceHZ · 03Numbers in meetingsa wrong figure is a conduct problemSYS/FOBC · FIG.02 · SCHEMATIC
FIG.02 Why it was hard. Agency had to be placed precisely between a rules engine and an ungovernable agent mesh.

The data falls into three rings with different owners, different security regimes and different freshness: the workspace estate, the enterprise's internal cloud estate, and external market data. No single platform spans all three with the entitlements intact, and pretending otherwise fails at the assurance gate, because where a model is allowed to run and what data may cross which boundary are fixed by security policy, not preference.

The question range was the second trap. A declarative topic bot is really a rules engine: fine for a handful of intents, broken across the spread of questions a front-line staff actually asks. But a fully agentic system across this surface would never clear governance. The design had to place agency precisely, not everywhere.

And the answers carry numbers into client conversations. A wrong figure in a meeting brief is not a quality problem, it is a conduct problem, so grounding had to be enforced rather than encouraged.

Key decisions

FIG.03 // KEY DECISIONSREJECTED PATHGATEAGENTMODELSIGNAL / DATAD · 01Scored, not arguedSix optionsTen weighted dimensionsHybrid held underre-weightingD · 02Only the root plansRoot planner is agenticEvery other step tieredand boundedD · 03Models against barsSmall encoder: intent,accuracy and latencySmall generative:faithfulnessFrontier: synthesis andcomplianceD · 04Failure modes firstBounded replan budgetPartial, cited answer,not open-ended spendSYS/FOBC · FIG.03 · SCHEMATIC · NOT TO SCALE
FIG.03 // KEY DECISIONSSIGNAL / DATAAGENTMODELGATEREJECTED PATHD · 01Scored, not arguedSix optionsTen weighted dimensionsHybrid held under re-weightingD · 02Only the root plansRoot planner is agenticEvery other step tiered and boundedD · 03Models against barsSmall encoder: intent, accuracy andlatencySmall generative: faithfulnessFrontier: synthesis and complianceD · 04Failure modes firstBounded replan budgetPartial, cited answer, notopen-ended spendSYS/FOBC · FIG.03 · SCHEMATIC
FIG.03 Key decisions. Each model choice clears an explicit acceptance bar; failure modes were engineered before features.

Six candidate options were scored against ten weighted dimensions covering capability, strategic fit and execution. The recommended hybrid led clearly, and the ranking held under any reasonable re-weighting. The split is clean: the collaboration workspace remains the workspace surface, the reasoning brain runs on the enterprise's own cloud estate, and vector and graph retrieval sit together as one layer.

Only the root planner is genuinely agentic. Every other step is tiered on the enterprise's autonomy taxonomy and behaves as a bounded component. Each model choice is made against an explicit acceptance bar rather than preference: a small encoder for intent classification against an accuracy and latency bar, small generative models for sub-agent summarisation against a faithfulness bar, a frontier model only for root synthesis and compliance review, where the small models fell short of the completeness bar.

Failure modes were engineered before features. The planner runs on a bounded replan budget, so a pathological question degrades into a partial, cited answer rather than an open-ended token spend.

Architecture

FIG.04 // ARCHITECTUREREJECTED PATHSTORE / RECORDGATEAGENTSIGNAL / DATAW · 01Staff ask in the workspaceORC · 02Root orchestratorplan, act, observe, replan; gatedon sufficiencyAGT · 03Domain agentsallow-listed topology, governed toolcontractsKL · 04Knowledge layerdense and sparse fused and rerankedGRD · 05Blocking citation guardrailW · 06Renders on the surfaceSERVED FROMVIEWSCurated row-level-secured viewsinternal data is replicated, neverqueried at source by agentsSYS/FOBC · FIG.04 · SCHEMATIC · NOT TO SCALE
FIG.04 // ARCHITECTURESIGNAL / DATAAGENTGATESTORE / RECORDREJECTED PATHW · 01Staff ask in the workspaceORC · 02Root orchestratorplan, act, observe, replan; gated onsufficiencyAGT · 03Domain agentsallow-listed topology, governed tool contractsKL · 04Knowledge layerdense and sparse fused and rerankedGRD · 05Blocking citation guardrailW · 06Renders on the surfaceSERVED FROMVIEWSCurated row-level-secured viewsinternal data is replicated, never queriedat source by agentsSYS/FOBC · FIG.04 · SCHEMATIC
FIG.04 Architecture. Agents never query internal systems at source; the knowledge layer serves curated, row-level-secured views.

The staff asks in the workspace. The request crosses an identity-federated boundary to the root orchestrator, which runs plan, act, observe, replan, gated on sufficiency. Thin domain agents fan out over an allow-listed topology and reach data through governed tool contracts. They stand on a four-projection knowledge layer: a graph for who is connected, a vector store for what the text says, structured marts for how much, and a live tier for what just changed. The synthesis passes a blocking citation guardrail before anything renders back on the surface.

Internal structured data is never queried at source by agents; it is replicated into curated, row-level-secured views that the knowledge layer serves. Retrieval is hybrid, dense and sparse fused and reranked, because front-line questions swing between the precise and the vague within a single session.

Trade-offs

FIG.05 // TRADE-OFFSREJECTED PATHGATEDETERMINISTIC▼ PAYSDESIGN CHOICE▲ BUYSCosts more to integrate thanall-in single vendorT-01HybridData reach and the assurancegateCalibration and escalationmachineryT-02Small-model-first routingInference cost held downsharplySome hard questions come backpartialT-03Bounded replan budgetA partial answer withcitationsA PARTIAL CITED ANSWER BEATS A COMPLETE UNCITED ONESYS/FOBC · FIG.05 · SCHEMATIC · NOT TO SCALE
FIG.05 // TRADE-OFFSDETERMINISTICGATEREJECTED PATHT-01Hybrid▼ PAYSCosts more tointegrate thanall-in singlevendor▲ BUYSData reach andthe assurancegateT-02Small-model-first routing▼ PAYSCalibration andescalationmachinery▲ BUYSInference costheld downsharplyT-03Bounded replan budget▼ PAYSSome hardquestions comeback partial▲ BUYSA partial answerwith citationsA PARTIAL CITED ANSWER BEATS ACOMPLETE UNCITED ONESYS/FOBC · FIG.05 · SCHEMATIC
FIG.05 Trade-offs. No integration saving compensates for answering a fifth of the question space.

The hybrid costs more to integrate than an all-in single-vendor design, and the decision paper says so plainly; it wins because the all-in options fail on data reach and on the assurance gate, and no integration saving compensates for answering a fifth of the question space. Small-model-first routing holds inference cost down sharply but adds calibration and escalation machinery. The bounded replan budget means some hard questions come back partial, by design: a partial answer with citations beats a complete answer without them.

Governance

FIG.06 // GOVERNANCEREJECTED PATHSTORE / RECORDGATEMODELCONTROLMECHANISMSTATEEvery step is tieredAutonomy tier and numeric evaluation gate, intocertificationCERTIFIEDCitation guardrailAn answer with an unsourced number does not renderBLOCKINGBehavioural driftTool-selection distribution monitored per agentMONITOREDAssurance referenceDecision, autonomy artefact and model evaluationsacceptedREFERENCESYS/FOBC · FIG.06 · SCHEMATIC · NOT TO SCALE
FIG.06 // GOVERNANCEMODELGATESTORE / RECORDREJECTED PATHCERTIFIEDEvery step is tieredAutonomy tier and numeric evaluation gate,into certificationBLOCKINGCitation guardrailAn answer with an unsourced number does notrenderMONITOREDBehavioural driftTool-selection distribution monitored peragentREFERENCEAssurance referenceDecision, autonomy artefact and modelevaluations acceptedSYS/FOBC · FIG.06 · SCHEMATIC
FIG.06 Governance. Behavioural drift shows up before verbal drift.

Every step carries an autonomy tier and a numeric evaluation gate, feeding the enterprise's certification process. The citation guardrail is blocking, not advisory: an answer with an unsourced number does not render. The platform monitors each agent's tool-selection distribution over time, because behavioural drift shows up before verbal drift. The architecture decision, the autonomy artefact and the model evaluations were accepted as the enterprise's assurance reference for work of this class.

Outcome

FIG.07 // OUTCOMEDETERMINISTICSIGNAL / DATARECOVERED PREPARATION TIMEHoursPER STAFF PER WEEK, MODELLEDSURFACE STRATEGYNo conflictVENDOR ASSISTANT AND OWNED BRAIN ADVANCE TOGETHEREXTENDING TO ADJACENT FRONT-LINE WORKFLOWSSchedule automation in commercial lendingQuote preparation inside pricing thresholdsSYS/FOBC · FIG.07 · SCHEMATIC · NOT TO SCALE
FIG.07 // OUTCOMESIGNAL / DATADETERMINISTICRECOVERED PREPARATION TIMEHoursPER STAFF PER WEEK, MODELLEDSURFACE STRATEGYNo conflictVENDOR ASSISTANT AND OWNED BRAIN ADVANCETOGETHEREXTENDING TO ADJACENT FRONT-LINEWORKFLOWSSchedule automation in commercial lendingQuote preparation inside pricing thresholdsSYS/FOBC · FIG.07 · SCHEMATIC
FIG.07 Outcome. Material capacity across a large coverage population, and a surface strategy that avoids vendor conflict.

The decision paper drove the build, and the recovered preparation time is modelled as hours per staff per week, which is material across a large coverage population. More durable than the capacity case: the surface strategy lets the vendor assistant and the enterprise-owned brain advance without conflict, and the same pattern is extending to adjacent front-line workflows such as schedule automation in commercial lending and quote preparation inside pricing thresholds.

What I learned

FIG.08 // WHAT I LEARNEDREJECTED PATHMODELSIGNAL / DATAAX-01The surface is not the systemConflating the two is the most commonerror in enterprise assistant design.AX-02Bars end model argumentsWith a numeric bar, model selection isengineering, not a taste war.AX-03Design the refusals firstReplan budget, topology, blockingguardrail: everyone knew what it wouldrefuse.SYS/FOBC · FIG.08 · SCHEMATIC · NOT TO SCALE
FIG.08 // WHAT I LEARNEDSIGNAL / DATAMODELREJECTED PATHAX-01The surface is not the systemConflating the two is the most common error inenterprise assistant design.AX-02Bars end model argumentsWith a numeric bar, model selection isengineering, not a taste war.AX-03Design the refusals firstReplan budget, topology, blocking guardrail:everyone knew what it would refuse.SYS/FOBC · FIG.08 · SCHEMATIC
FIG.08 What I learned. Three lessons from the engagement.

The surface is not the system. Conflating the two is the single most common error I see in enterprise assistant design. Acceptance bars ended the model arguments: once every step had a numeric bar, model selection became an engineering decision instead of a taste war. And designing the failure modes first, the replan budget, the topology, the blocking guardrail, made the feature conversations easier, because everyone knew what the system would refuse to do.