Home · Work

Systems Work · No. 22

Credit Underwriting with Auditable Reasoning

Rajesh Ranjan Mahapatra

An agent pipeline that compresses evidence gathering for referred cases, from documents through financial reasoning to a cited decision memo, while the decision stays with a reviewer.

Start reading
Systems WorkCase study

© 2026 Rajesh Ranjan Mahapatra. 6 min read

Credit Underwriting with Auditable ReasoningAll Systems Work
Home · Work

Credit Underwriting with Auditable Reasoning

An agent pipeline that compresses evidence gathering for referred cases, from documents through financial reasoning to a cited decision memo, while the decision stays with a reviewer.

Challenge

At a large regulated enterprise, most consumer and commercial credit is auto-decisioned by a deterministic decision engine.

Architectural move

The pipeline is an explicit state machine, not a free agent loop:

Control boundary

The system ran in shadow mode first, with recommendations logged but never shown, producing the accuracy, precision and recall evidence the model risk committee required before any assistant went live.

Outcome

The proof of concept collapsed several minutes of per-case gathering to well under a minute, and the business case behind it scored well over a hundred referral use cases across three lending domains.

FIG.00 // SYSTEM MAPGATE / HUMANAGENTMODELDETERMINISTICSIGNAL / DATAIN · 01Referred casethe residual thedecision engine couldnot auto-decisionAGT · 02Gatherer agentssix systems in parallelover defined toolcontractsDET · 03Groundingmailbox and contentstore converge on oneevidence schemaDET · 04Financial reasoningspreads figures, checksarithmetic invariantsMDL · 05Risk signals andpolicy checksproposed dispositionplus a cited decisionmemoAGT · 06Reviewer agentchecks the proposalagainst policyHUM · 07Human decision gatethe reviewer decidesevery caseNORMALISESTRUCTURE, NOT FORMATSRULE · 00Compress the gathering; never touch the judgmentgather · normalise · ground · propose · gate — an explicit state machine, replayable from the case recordSYS/CUAR · FIG.00 · SCHEMATIC · NOT TO SCALE
FIG.00 // SYSTEM MAPSIGNAL / DATADETERMINISTICAGENTMODELGATE / HUMANIN · 01Referred casethe residual the decision engine could notauto-decisionAGT · 02Gatherer agentssix systems in parallel over defined toolcontractsNORMALISEDET · 03Groundingmailbox and content store converge on oneevidence schemaDET · 04Financial reasoningspreads figures, checks arithmetic invariantsSTRUCTURE, NOT FORMATSMDL · 05Risk signals and policy checksproposed disposition plus a cited decisionmemoAGT · 06Reviewer agentchecks the proposal against policyHUM · 07Human decision gatethe reviewer decides every caseRULE · 00Compress the gathering; never touchthe judgmentgather · normalise · ground · propose · gate— an explicit state machine, replayable fromthe case recordSYS/CUAR · FIG.00 · SCHEMATIC
FIG.00 System map. Agents gather and ground; deterministic structure carries the reasoning; the reviewer keeps the judgment.

The problem

FIG.01 // THE PROBLEMDETERMINISTICPER REFERRED CASEMinutesSPENT GATHERING ACROSS SIX SYSTEMSTHE JUDGMENT ITSELFSecondsTHE PART ONLY THE REVIEWER CAN DOCOMMERCIAL CREDIT PACKWeeksELAPSED, ASSEMBLED BY HANDTHE ECONOMICS OF THE RESIDUAL ARE UPSIDE DOWNSYS/CUAR · FIG.01 · SCHEMATIC · NOT TO SCALE
FIG.01 // THE PROBLEMDETERMINISTICPER REFERRED CASEMinutesSPENT GATHERING ACROSS SIX SYSTEMSTHE JUDGMENT ITSELFSecondsTHE PART ONLY THE REVIEWER CAN DOCOMMERCIAL CREDIT PACKWeeksELAPSED, ASSEMBLED BY HANDTHE ECONOMICS OF THE RESIDUAL AREUPSIDE DOWNSYS/CUAR · FIG.01 · SCHEMATIC
FIG.01 The problem. Most credit is auto-decisioned; the referred residual spends minutes gathering for seconds of judgment, and commercial packs take weeks.

At a large regulated enterprise, most consumer and commercial credit is auto-decisioned by a deterministic decision engine. The residual cases it refers fall to reviewers, and the economics of that residual are upside down: a reviewer spends several minutes per case gathering data across six systems, while the judgment itself takes seconds. Meanwhile the memo-writing end of commercial lending had the same shape at larger scale, with credit packs assembled by hand from statements, registries and internal systems over weeks of elapsed time.

I led the architecture and design for the agentic adjudication work inside the enterprise's credit platform, from discovery through to a board-ready proof of concept. Compress the gathering; never touch the judgment. That rule survived every review.

Why it was hard

FIG.02 // WHY IT WAS HARDREJECTED PATHGATEDETERMINISTICSIGNAL / DATADEAD ENDRe-run the engine's logicthe engine is deterministic andalready had its chance; addsnothingONLY LEVERNew information the enginenever hada document it could not read, across-source reconciliation, a factthat arrived by emailDEAD ENDSecond-guess the decisionthe system must createinformation, not overrule anengine→ NO NEW SIGNALNO LICENCE ←CONSTRAINTS THE BUILD HAD TO LIVE INSIDEC · 01No external cloudmodelsacceptable-use policy:small models on theenterprise platformC · 02Two ingestion pathsshared mailbox as often asthe content storeC · 03Systems being retiredbuild on them and inherittheir death dateC · 04Statistical evidencefirstmodel risk needsaccuracy proof beforeany reviewer sees outputSYS/CUAR · FIG.02 · SCHEMATIC · NOT TO SCALE
FIG.02 // WHY IT WAS HARDSIGNAL / DATADETERMINISTICGATEREJECTED PATHDEAD ENDRe-run the engine's logicthe engine is deterministic and alreadyhad its chance; adds nothing↑ NO NEW SIGNALONLY LEVERNew information the engine never hada document it could not read, a cross-sourcereconciliation, a fact that arrived by email↓ NO LICENCEDEAD ENDSecond-guess the decisionthe system must create information, notoverrule an engineCONSTRAINTS THE BUILD HAD TO LIVE INSIDEC · 01No external cloud modelsacceptable-use policy: small models on theenterprise platformC · 02Two ingestion pathsshared mailbox as often as the content storeC · 03Systems being retiredbuild on them and inherit their death dateC · 04Statistical evidence firstmodel risk needs accuracy proof before anyreviewer sees outputSYS/CUAR · FIG.02 · SCHEMATIC
FIG.02 Why it was hard. The only lever is information the engine never had; everything else is second-guessing a deterministic decision.

Re-running the decision engine's logic adds nothing, because the engine is deterministic and already had its chance. The only lever is new information the engine never had: a document it could not read, a reconciliation across sources it does not query, a fact that arrived by email instead of through the workflow. So the system had to create information, not second-guess a decision.

The constraints were real. The enterprise's AI acceptable-use policy barred external cloud models from touching customer data, which forced a deterministic core with small models running on the enterprise's own platform. Documents arrive through a shared mailbox as often as through the content store, so a single tidy ingestion path was a fantasy. Parts of the surrounding estate were being decommissioned, so anything built against them would inherit a death date. And model risk requirements meant the system needed statistical evidence of its accuracy before any reviewer could see a recommendation.

Key decisions

FIG.03 // KEY DECISIONSSTORE / RECORDHUMAN AUTHORITYAGENTMODELDETERMINISTICSIGNAL / DATAS · 01Gathertool contracts,both ingestionpathsS · 02Normaliseone evidenceschema before anymodelS · 03Groundrules carry themajority at nomodel costS · 04Proposesmall fine-tunedmodel on theresidual onlyS · 05Gatethe reviewerdecidesGRADED BYEVAL · FIRSTGolden set builtbefore the pipelinehistorical referred caseswith known outcomes, gradedon agreement with thereviewerANCHORIntegration only on platforms that survive the re-engineeringnever on the systems being retiredSYS/CUAR · FIG.03 · SCHEMATIC · NOT TO SCALE
FIG.03 // KEY DECISIONSDETERMINISTICAGENTMODELHUMAN AUTHORITYSTORE / RECORDS · 01Gathertool contracts, both ingestion pathsS · 02Normaliseone evidence schema before any modelS · 03Groundrules carry the majority at no model costS · 04Proposesmall fine-tuned model on the residual onlyS · 05Gatethe reviewer decidesGRADED BYEVAL · FIRSTGolden set built before thepipelinehistorical referred cases with knownoutcomes, graded on agreement with thereviewerANCHORIntegration only on platforms thatsurvive the re-engineeringnever on the systems being retiredSYS/CUAR · FIG.03 · SCHEMATIC
FIG.03 Key decisions. An explicit state machine, not a free agent loop: each stage has a contract and a defined failure mode, and evaluation came first.

The pipeline is an explicit state machine, not a free agent loop: gather, normalise, ground, propose, gate. Each stage has a contract and a defined failure mode, which is what makes the whole thing replayable for a reviewer.

Both ingestion paths, mailbox and content store, converge on a single normalised evidence schema before any model sees a document. The model reasons over structure, never over formats.

Deterministic rules carry the majority of cases at no model cost. A small fine-tuned model handles only the residual that genuinely needs judgment, sized to the triage task rather than a frontier model on every case.

Evaluation was built before the pipeline: a golden set of historical referred cases with known outcomes, graded on agreement with the reviewer's actual decision. That decision converted weeks of shadow running into committee evidence rather than anecdote.

Integration was anchored on the platforms that will survive the current re-engineering, never on the systems being retired.

Architecture

FIG.04 // ARCHITECTURESTORE / RECORDHUMAN AUTHORITYAGENTMODELDETERMINISTICSIGNAL / DATAL1 · GATHERGatherer agentsSix systems inparallelDefined toolcontractsMailbox andcontent storeL2 · GROUNDGrounding and financeEvidence schemaFigures spreadArithmeticinvariants:unbalanced sheetsrejectedL3 · PROPOSERisk and policyRisk signalsProposeddispositionMemo: everyfigure cites itssourceL4 · GATEReview and decideReviewer agentchecks policyHuman decisiongateTOOL RESULTSSTRUCTURED EVIDENCEPROPOSAL + MEMOSAME GRAMMARSpreading,policysearch, memopacks,obligations,pre-eligibilitySYS/CUAR · FIG.04 · SCHEMATIC · NOT TO SCALE
FIG.04 // ARCHITECTURESIGNAL / DATADETERMINISTICAGENTMODELHUMAN AUTHORITYSTORE / RECORDL1 · GATHERGatherer agentsSix systems in parallelDefined tool contractsMailbox and content storeL2 · GROUNDGrounding and financeEvidence schemaFigures spreadArithmetic invariants: unbalancedsheets rejectedL3 · PROPOSERisk and policyRisk signalsProposed dispositionMemo: every figure cites its sourceL4 · GATEReview and decideReviewer agent checks policyHuman decision gateTOOL RESULTSSTRUCTURED EVIDENCEPROPOSAL + MEMOSAME GRAMMARSpreading, policy search, memo packs,obligations, pre-eligibilitySYS/CUAR · FIG.04 · SCHEMATIC
FIG.04 Architecture. The same gather, ground, propose, gate grammar extends across the credit lifecycle; the refer-back build is the entry point.

Gatherer agents pull from the six systems in parallel over defined tool contracts. A grounding agent normalises what they return into the evidence schema. Financial reasoning runs over that structure, spreading figures and checking them against arithmetic invariants, so a balance sheet that does not balance is rejected rather than narrated. Risk signals and policy checks feed a proposed disposition and a drafted decision memo in which every material figure cites its source. A reviewer agent checks the proposal against policy, and only then does the case reach the human gate.

The same gather, ground, propose, gate grammar extends across the credit lifecycle: financial spreading from client statements, credit-policy search for analysts, automated memo packs, obligations monitoring and pre-eligibility checks that understand a borrower before a decision is asked for. The refer-back build was designed as the entry point into that wider estate, not as a one-off.

Trade-offs

FIG.05 // TRADE-OFFSREJECTED PATHGATEDETERMINISTIC▼ PAYSDESIGN CHOICE▲ BUYSGives up the flexibility of afree agent loopT-01Explicit state machineA run a reviewer can replayLower peak capabilityT-02Small model on the residualOwnership, cost and policycomplianceWeeks of delayed visiblevalueT-03Shadow mode firstEvidence a risk committeecould signA LOOP THAT CAN IMPROVISE IS A LOOP NOBODY CAN REPLAYSYS/CUAR · FIG.05 · SCHEMATIC · NOT TO SCALE
FIG.05 // TRADE-OFFSDETERMINISTICGATEREJECTED PATHT-01Explicit state machine▼ PAYSGives up theflexibility of afree agent loop▲ BUYSA run a reviewercan replayT-02Small model on the residual▼ PAYSLower peakcapability▲ BUYSOwnership, costand policycomplianceT-03Shadow mode first▼ PAYSWeeks of delayedvisible value▲ BUYSEvidence a riskcommittee couldsignA LOOP THAT CAN IMPROVISE IS A LOOPNOBODY CAN REPLAYSYS/CUAR · FIG.05 · SCHEMATIC
FIG.05 Trade-offs. Each trade gives up something a demo would show for something a committee can sign.

An explicit state machine gives up the flexibility of a free agent loop, and that was the point: a loop that can improvise is a loop a reviewer cannot replay. The small-model choice trades peak capability for ownership, cost and policy compliance, and it holds because the hard reasoning was pushed into deterministic structure before the model ever runs. Shadow mode delayed visible value by weeks, which is uncomfortable in a programme hungry for demos, but it is what turned the system from a promising prototype into something a risk committee could sign.

Governance

FIG.06 // GOVERNANCESTORE / RECORDGATE / HUMANDETERMINISTICSIGNAL / DATAPH · 01Shadow moderecommendations logged,never shownPH · 02Accuracy evidenceaccuracy, precision andrecall against reviewerdecisionsPH · 03Model riskcommitteesigns before anyassistant goes livePH · 04Live, reviewerdecidesevery gated decisionreplayable from the caserecordALWAYSEvery memo claim maps to a source; customer data never leaves the platformunsupported figures are stripped or force regenerationSYS/CUAR · FIG.06 · SCHEMATIC · NOT TO SCALE
FIG.06 // GOVERNANCEDETERMINISTICGATE / HUMANSTORE / RECORDPH · 01Shadow moderecommendations logged, never shownPH · 02Accuracy evidenceaccuracy, precision and recall againstreviewer decisionsPH · 03Model risk committeesigns before any assistant goes livePH · 04Live, reviewer decidesevery gated decision replayable from the caserecordALWAYSEvery memo claim maps to a source;customer data never leaves theplatformunsupported figures are stripped or forceregenerationSYS/CUAR · FIG.06 · SCHEMATIC
FIG.06 Governance. Shadow running produced the committee evidence; citations and data residency hold in every phase.

The system ran in shadow mode first, with recommendations logged but never shown, producing the accuracy, precision and recall evidence the model risk committee required before any assistant went live. Every claim in a memo maps to a source; unsupported figures are stripped or force regeneration. Customer data never leaves the enterprise's own platform. The reviewer decides every case, and each gated decision is auditable end to end, replayable from the case record alone.

Outcome

FIG.07 // OUTCOMEDETERMINISTICPER-CASE GATHERINGseveral minutes< 1 minPROOF OF CONCEPTCOMMERCIAL MEMO PREPweeks elapsedHoursAGENTS RECOMMEND, REVIEWERS DECIDEBUSINESS CASE100+REFERRAL USE CASES SCORED ACROSSTHREE LENDING DOMAINSACCEPTED AS THE TEMPLATE FOR RE-ENGINEERING LENDING ORIGINATIONSYS/CUAR · FIG.07 · SCHEMATIC · NOT TO SCALE
FIG.07 // OUTCOMEDETERMINISTICPER-CASE GATHERINGseveral minutes< 1 minPROOF OF CONCEPTCOMMERCIAL MEMO PREPweeks elapsedHoursAGENTS RECOMMEND, REVIEWERS DECIDEBUSINESS CASE100+REFERRAL USE CASES SCORED ACROSS THREELENDING DOMAINSACCEPTED AS THE TEMPLATE FORRE-ENGINEERING LENDING ORIGINATIONSYS/CUAR · FIG.07 · SCHEMATIC
FIG.07 Outcome. Gathering collapsed from minutes to under a minute and memo preparation from weeks to hours.

The proof of concept collapsed several minutes of per-case gathering to well under a minute, and the business case behind it scored well over a hundred referral use cases across three lending domains, a large subset modelled down to activity level and cost per case. On the commercial side, the same pattern took memo preparation from weeks of elapsed time to hours, with agents recommending and reviewers deciding. The design was accepted as the template for re-engineering lending origination as the legacy platform is replaced.

What I learned

FIG.08 // WHAT I LEARNEDSTORE / RECORDGATEDETERMINISTICAX-01Find the only leverInformation the engine never hadkilled a dozen seductive ideas andfocused the build.AX-02Build the evaluation firstA golden set with known outcomes beatsany architecture diagram in front of arisk committee.AX-03Boring defends betterBoring pipelines defend better thanclever ones.SYS/CUAR · FIG.08 · SCHEMATIC · NOT TO SCALE
FIG.08 // WHAT I LEARNEDDETERMINISTICGATESTORE / RECORDAX-01Find the only leverInformation the engine never had killed adozen seductive ideas and focused the build.AX-02Build the evaluation firstA golden set with known outcomes beats anyarchitecture diagram in front of a riskcommittee.AX-03Boring defends betterBoring pipelines defend better than cleverones.SYS/CUAR · FIG.08 · SCHEMATIC
FIG.08 What I learned. Three lessons from the engagement.

The most useful sentence in the whole engagement was the constraint: the only lever is information the engine never had. It killed a dozen seductive ideas about second-guessing the decision engine and focused the build on evidence. I also learned to build the evaluation before the pipeline; a golden set of historical cases with known outcomes is worth more than any architecture diagram when the audience is a risk committee. Boring pipelines defend better than clever ones.