← Agent catalogue
AI Agent Profile · LendingIQ · Bengaluru
Credit Decision Agent AI
Invoked via: loan origination pipeline APIRuntime: AWS Bedrock · ap-south-1Model: Claude Sonnet 4Context window: 200K tokens
DivisionRisk division
Resume
What this agent does
The Credit Decision Agent AI executes the policy rules layer of LendingIQ's credit decisioning — applying the scorecard thresholds, hard policy filters, and authority matrix to the underwriting assessment, producing a structured decision record with a documented rationale for every outcome, and routing the decision to the correct tier for either autonomous approval or human review. It is the rules engine between the Underwriting Agent AI's assessment and the final credit verdict.
Primary functions
Scorecard Execution
Every application — synchronousInvoked when: underwriting assessment and all input data are available for policy rules application
- Retrieves the current credit policy scorecard via RAG at the moment of invocation — always the live version, never a cached copy. Applies every scorecard dimension: bureau score against the segment cut-off, FOIR against the product limit using the bank-statement-verified income figure, LTV against the product maximum, leverage ratio for MSME borrowers, and sector concentration check against the current portfolio limits. Each dimension is evaluated independently before the composite decision is formed.
- Applies hard filters first — bureau score below cut-off, FOIR above hard limit, sector exposure exceeding concentration limit — as absolute stops regardless of how strong the borrower's profile is on other dimensions. A hard filter is non-negotiable: the agent does not construct arguments for why a hard filter should be waived, and it does not present the case as borderline when it is not. Hard filter declines are final at L1.
- For applications that clear all hard filters, computes the composite scorecard result and maps it to the decision authority matrix: which tier is authorised to approve an application with this profile, this ticket size, and this exception count.
Output: Scorecard results — each dimension with borrower figure, policy limit, pass/fail, and margin above/below limit. Hard filter status. Composite result and authority tier determination.
Policy Rules Enforcement
Every application — all active policy rules appliedInvoked simultaneously with scorecard execution — policy rules are a separate layer from the scorecard
- Applies the non-scorecard policy rules: product eligibility rules (minimum business vintage for MSME products, minimum employment tenure for salaried products), geographic restrictions where applicable, borrower category restrictions (negative list sectors, excluded borrower types), and any current moratorium or temporary policy restrictions active in the portfolio.
- Checks for policy exceptions: where any application attribute falls outside the policy norm but not into a hard filter (FOIR 52% where policy limit is 50%, business vintage 11 months where minimum is 12 months), identifies the exception, classifies it as single or multiple, and routes appropriately. Single exceptions within the L2 officer's mandate are flagged to L2. Multiple exceptions always go to L3.
- Does not apply judgment about whether a policy exception is reasonable. Judgment on exceptions is the human credit officer's function. The agent identifies and routes; the human decides.
Output: Policy rules check — each applicable rule with application value, policy requirement, pass/fail. Exception list with exception type and routing tier. Combined rules-and-scorecard decision recommendation: Approve (L1) / Refer L2 / Refer L3 / Decline.
Decision Record + Rationale
Every application — mandatory outputInvoked after scorecard and policy rules execution — decision record is the final output of this agent
- Produces a structured decision record for every application outcome — Approve, Refer, or Decline — with the specific policy basis for the outcome: which rule or filter drove the decision, what the borrower's figure was, what the policy limit is, and what the margin or deviation is. A decision without a specific, cited basis is not produced — every output is traceable to a policy clause.
- Produces a plain-language rationale for the borrower where the decision is an Adverse Action (decline or approval with conditions that materially limit the borrower's access to credit): the specific reason, stated in non-technical language, citing the financial parameter that led to the decision. This is a regulatory obligation under RBI's fair practices guidelines for algorithmic lending — not optional, not a summary.
- The full decision record — including all scorecard inputs, policy checks, exception flags, and rationale — is passed to the Audit Trail Agent AI for logging. Every credit decision made by this agent is permanently retrievable with its complete policy basis, for regulatory inspection, internal audit, or borrower grievance response.
Output: Structured JSON decision record — outcome (Approve/Refer L2/Refer L3/Decline), all scorecard and policy check results with borrower figures and policy limits, exception list, routing instruction, plain-language borrower rationale for Adverse Action decisions, and audit trail submission confirmation.
Referral Routing
Every Refer outcome — routed to correct tierInvoked when: the decision outcome is a Refer at any tier
- Routes Refer decisions to the appropriate human credit officer tier based on the decision authority matrix — L2 for single-exception within-policy cases, L3 for multi-exception or high-ticket cases — and packages the referral brief for the receiving officer: the decision record, the specific exception or flag that triggered the referral, the recommendation from the agent (approve within exception authority, decline, or request additional information), and the expiry window for the referral before the application is auto-expired.
- For L3 referrals above a configured exposure threshold, additionally routes the case to the CRO AI for portfolio-level context: does approving this exception create concentration risk, and is the borrower's segment currently in the portfolio's watch list? The CRO AI's portfolio context is an input to the L3 officer's decision, not a veto.
Output: Referral package — tier routed to, exception or flag driving referral, agent recommendation, referral brief for the receiving officer, expiry window, CRO AI portfolio context note for L3 cases.
Hard guardrails
Will notOverride a hard policy filter under any circumstances. Hard filters are non-negotiable architectural constraints. Any application requesting a hard filter override is routed to L3 human credit committee with an explicit notation that the request falls outside the agent's authority — it is not approved at any automated tier.
Will notProduce a decision without a documented policy basis. Every outcome — approve, refer, or decline — cites the specific rule, scorecard dimension, or filter that drove it. A decision the agent cannot trace to a policy basis is routed to human review, not produced autonomously.
Will notApply judgment on policy exceptions. Exceptions are identified, quantified, and routed. The credit officer at the appropriate tier makes the judgment call on whether to approve the exception.
Known limitations
The agent is only as current as the policy corpus retrieved at invocation. A policy change that has been approved but not yet reflected in the RAG corpus will result in the agent applying the prior policy. Any policy amendment must update the corpus before the next application processing cycle.Configure the credit policy corpus with a version control system that alerts when the live policy document and the corpus version diverge. A mismatch of more than 24 hours on any active policy document should trigger an escalation to the operations team.
The authority matrix thresholds are configured by the credit team and applied by the agent. If the thresholds are misconfigured — for example, the L1 autonomy threshold is set too high, causing cases that should require human review to be approved autonomously — the agent will apply the misconfigured thresholds correctly. Configuration governance is a human responsibility.Audit the decision authority matrix configuration quarterly against the credit policy document. Any deviation between the configured thresholds and the approved policy limits is a governance gap that requires immediate correction.
Important Reads
Learn more about how to deploy Credit Decision Agent AI to your lending workflow.
- Use case #0001Scorecard execution: how Credit Decision AI runs 40 policy rules in parallelA credit policy is a document. A credit scorecard is a decision engine. The Credit Decision Agent AI converts the institution's written credit policy — every rule, every threshold, every exception pathway — into an executable scorecard that runs 40 checks in parallel, produces a decision in under 90 seconds, and generates a documented rationale for every outcome. The underwriter's job begins after the scorecard, not before it.Read article →
- Use case #0002Decision rationale: how Credit AI explains every outcome in plain languageA credit decision without a rationale is not a credit decision — it is a verdict. The RBI's Fair Practices Code requires that every loan decline provides the borrower with the specific basis for the decision. The Credit Decision Agent AI generates that rationale automatically for every outcome: an approval explains what drove the positive decision and what the borrower's account will look like; a decline names the specific rules that failed and, where possible, what the borrower can do to address them; a referral brief tells the human underwriter exactly what the AI found and what requires their judgment.Read article →
- Use case #0003Referral routing: when Credit Decision AI sends applications to a human underwriterThe Credit Decision Agent AI auto-approves 68% of applications and auto-declines 14%. The remaining 18% are referred to a human underwriter — not because the system failed to process them, but because they contain a specific, identified ambiguity that requires human judgment to resolve. The referral is not a failure state. It is the AI's most precise output: a complete case brief, a specific question, and a defined set of options for the underwriter to choose from.Read article →