Skip to main content

Responsible AI in Lending

Recourse Is Not an Apology — It Is Infrastructure

NomoCrit

Actionable recourse is a technical system requirement, not a communication gesture. Its design, validity, and lifecycle must be engineered and monitored with the same rigour as the model itself.

Recourse Is Not an Apology — It Is Infrastructure

Most rejection letters end with a sentence that sounds helpful and means nothing: "You may improve your eligibility by improving your credit profile." It is a gesture dressed as guidance. It tells the applicant that something is wrong without telling them what to do, and it shifts the burden of interpretation onto the person least equipped to carry it.

This essay argues that recourse — the set of actions a rejected applicant can take to be approved — is not a customer-service flourish appended to a decision. It is infrastructure. It has a formal definition, a measurable validity, a decay rate, and a lifecycle. Treat it as an apology and it rots silently. Treat it as a system and it becomes the mechanism through which underserved borrowers enter formal credit. The difference between those two outcomes is engineering discipline, not goodwill.

Section 1: Correlational vs Causal Recourse — A Technical Distinction

Correlational recourse

Most recourse in production today is correlational. It is read directly off a feature-importance ranking — global SHAP magnitudes, permutation importance, or coefficient weights — and converted into prose. The model says the credit_score feature carries large weight, so the letter says "improve your CIBIL score."

This is technically true and operationally useless. "Improve your CIBIL score" is not an action; it is the name of an outcome. It conflates a feature that is itself a downstream aggregate with a lever the borrower can pull. The applicant cannot directly set their bureau score the way they can set a thermostat. Correlational recourse identifies what correlates with approval in the population and presents it as what this individual should do — two entirely different claims.

Causal recourse, formally

Causal recourse asks a sharper question: what is the lowest-cost change to this individual's actionable features that flips the decision? This is a counterfactual optimisation problem:

$$ x' = \arg\min_{x'} ; d(x, x') \quad \text{subject to} \quad f(x') \geq \tau $$

where $f(\cdot)$ is the scoring model, $\tau$ is the approval threshold, and $d(x, x')$ is a cost-weighted distance between the applicant's current state $x$ and the counterfactual state $x'$. The distance is not Euclidean. It encodes how hard each change actually is:

$$ d(x, x') = \sum_{j} w_j \cdot \delta_j(x_j, x'_j) $$

The weights $w_j$ are the entire point. A unit change in overdraft utilisation and a unit change in years-of-credit-history are not equally available to a human being, and a distance metric that treats them as equal will produce recourse no one can follow.

Why mutability is the load-bearing concept

Features fall into mutability tiers, and recourse must respect them:

  • Immutable — age, geography, gender, caste, founding cohort. Cost $w_j \to \infty$. These must be excluded from recourse entirely, both because they cannot be changed and because instructing someone to change them is, at best, absurd and, at worst, discriminatory.
  • Slowly mutable — length of credit history, vintage of the business, age of the oldest tradeline. These improve only with the passage of time. They can inform when to reapply, never what to do this week.
  • Actionable — overdraft utilisation ratio, GST filing consistency, current-account cash-flow stability, DPD on existing lines. These are the only legitimate basis for recourse.

A recourse system that recommends an immutable or slowly-mutable feature is not merely unhelpful. It is misleading, because it implies a pathway that does not exist. The mutability constraint is therefore a hard constraint on the optimisation, not a preference: the feasible set of $x'$ is restricted to vary only along actionable dimensions.

Worked example: a self-employed merchant

Consider Priya, a self-employed merchant applying for a ₹15 lakh working-capital line. Her feature vector (abbreviated) and the model's output:

| Feature | Value | Mutability | |---|---|---| | overdraft_utilisation | 0.48 | actionable | | gst_filing_consistency | 0.61 | actionable | | cashflow_stability_6mo | 0.57 | actionable | | business_vintage_yrs | 3.2 | slowly mutable | | applicant_age | 41 | immutable | | pin_code_tier | 2 | immutable |

  • Model approval probability: $f(x) = 0.54$
  • Approval threshold: $\tau = 0.62$
  • Gap to clear: $0.08$

Local SHAP attribution (deviation from base value 0.50) reveals where the score is being suppressed:

| Feature | SHAP contribution | |---|---| | overdraft_utilisation | −0.061 | | gst_filing_consistency | −0.029 | | cashflow_stability_6mo | −0.011 | | business_vintage_yrs | +0.018 | | others (net) | +0.073 |

A correlational system stops here and tells Priya her utilisation is high. A causal system runs the counterfactual optimisation over the actionable subspace and returns a path with quantified score impact:

Causal recourse path R(x, t₀):
  1. reduce overdraft_utilisation: 0.48 → 0.28   →  Δf = +0.041
  2. increase gst_filing_consistency: 0.61 → 0.85 →  Δf = +0.038
  3. maintain cashflow_stability_6mo ≥ 0.57       →  Δf = +0.012
  ----------------------------------------------------------------
  Predicted f(x') = 0.54 + 0.091 = 0.631 ≥ τ = 0.62  ✓

Every instruction names an actionable feature, specifies a target value rather than a direction, and carries a predicted contribution to the decision. Priya does not have to interpret anything. She has a plan whose sufficiency the model itself has certified.

Section 2: The Mathematics of Recourse Expiry

Here is the failure that no rejection letter discloses: recourse is valid only relative to the model and threshold that issued it. The moment either changes, the guarantee evaporates — silently, because nobody re-checks.

Validity, formally

A recourse recommendation $R(x, t_0)$ issued at time $t_0$ is valid at time $t$ if and only if applying its actions $A(R)$ to the input still clears the current decision boundary:

$$ \text{Valid}_t\big(R(x, t_0)\big) \iff f_t\big(x + A(R)\big) \geq \tau_t $$

where $f_t$ and $\tau_t$ are the model and threshold active at time $t$, not at issuance. Recourse is constructed to satisfy $f_{t_0}(x + A(R)) \geq \tau_{t_0}$. Nothing guarantees it survives a single deployment after that.

Failure modes

Validity breaks through four distinct mechanisms:

(a) Model retrained. A new training run shifts the learned function $f_{t_0} \to f_t$. The marginal effect of reducing overdraft utilisation may shrink, or its interaction with cash-flow may change sign. The same actions now land below the boundary.

(b) Threshold recalibrated. Risk appetite tightens; $\tau_{t_0} = 0.62$ becomes $\tau_t = 0.66$. Priya's certified $f(x') = 0.631$ was sufficient on Monday and insufficient on the following quarter's reset. The model never changed — only the line did.

(c) Feature schema updated. A new feature is added, a bureau field is re-binned, or gst_filing_consistency is redefined from a 12-month to a 24-month window. The input representation $x$ that the recourse was written against no longer maps cleanly onto $x + A(R)$.

(d) Policy amendment. An eligibility rule is layered on top of the model — say, a hard cutoff on utilisation regardless of score — and overrides the model output entirely. The recourse remains "model-valid" while being "policy-invalid."

Quantifying decay: the Recourse Validity Decay Rate

Across a portfolio of issued recourses, we measure survival with the Recourse Validity Decay Rate (RVDR):

$$ \text{RVDR}(t) = \frac{\big|{, R : f_t(x + A(R)) \geq \tau_t ,}\big|}{\big|{, R : f_{t_0}(x + A(R)) \geq \tau_{t_0} ,}\big|} $$

The denominator is the set of recourses that were valid at issuance — by construction, all of them. The numerator is the subset still valid under today's model and threshold. So $\text{RVDR}(t) \in [0, 1]$ is simply the fraction of outstanding recourse that still works.

$\text{RVDR}(t_0) = 1$ by definition. It decays monotonically (absent re-issuance) with every retrain, recalibration, and schema change. The governance question is: what decay is acceptable before intervention? A defensible policy treats RVDR like any other SLA:

  • $\text{RVDR} \geq 0.95$ — healthy; routine monitoring.
  • $0.85 \leq \text{RVDR} < 0.95$ — flag; queue affected cohorts for re-issuance.
  • $\text{RVDR} < 0.85$ — breach; suspend reliance on outstanding guidance and trigger mandatory re-issuance before any reapplication is judged against it.

A single model deployment can drop RVDR sharply. That is precisely why recourse expiry must be computed, not assumed.

Section 3: Recourse Governance Requirements

If recourse is infrastructure, it needs a lifecycle management system — the same way a model has versioning, monitoring, and rollback.

The recourse lifecycle

Issuance. Every recourse record is persisted with the metadata required to ever re-evaluate it: recourse_id, applicant_id, issued_at timestamp, model_version, threshold_value, the feature_schema_version, and the SHAP basis that justified the path. Without model_version and threshold_value stored at issuance, you cannot later compute whether the recourse expired — you have thrown away the denominator of RVDR.

Validity monitoring. On every model deployment, threshold change, or schema migration, the system re-evaluates outstanding recourse against the current $f_t, \tau_t$ and updates RVDR per cohort. This is a batch job, cheap to run, and it is the difference between knowing your guidance is stale and finding out from a complaint.

Expiry notification. When a borrower's specific recourse transitions to invalid, the borrower is told. Silence is the cardinal sin here: a customer following a plan that no longer works deserves to know the plan changed before they spend three months executing it.

Re-issuance. Invalidated recourse is regenerated under the current model — a fresh counterfactual optimisation producing updated targets, with a new timestamp and version stamp. The borrower receives a revised, currently-valid path rather than a quiet failure.

Regulatory and legal exposure

Now the sharp edge. Suppose Priya follows her issued recourse in good faith — she pays down her overdraft, files three quarters of GST on time, holds her cash flow steady — reapplies, and is rejected, because between issuance and reapplication the threshold was recalibrated and nobody told her.

What is the institution's exposure?

Under the RBI's FREE-AI framework and broader fairness expectations, a lender is expected to provide explanations and a fair, non-arbitrary path to redress. Recourse that the institution itself issued, that the borrower demonstrably followed, and that the institution silently invalidated is not a neutral event. It is a representation the borrower relied on to her detriment. The defensible position — "the model changed" — is exactly the admission that sinks it: the institution had the information to know the guidance had expired (it controls $f_t$ and $\tau_t$) and chose not to act on it. The absence of RVDR monitoring is not a mitigating factor; it is evidence that the failure was foreseeable and unmonitored.

Engineered recourse with expiry tracking converts this from an open-ended liability into a documented, auditable process: we issued, we monitored, we detected expiry, we notified, we re-issued. That trail is the difference between a governable system and an indefensible one.

Section 4: Recourse as Financial Infrastructure

For thick-file prime borrowers, recourse is a convenience. For thin-file and underserved applicants — the self-employed, the first-time formal borrower, the merchant with strong cash flows and no bureau depth — causal recourse is the entry ramp into the formal credit system itself.

A thin-file applicant who is rejected with correlational boilerplate has nowhere to go. "Improve your credit history" is, for someone with no credit history, a closed loop: you need credit to build the history that gets you credit. Causal recourse over actionable, non-bureau features — utilisation discipline, GST consistency, demonstrable cash-flow stability — breaks that loop. It gives the rejected applicant a concrete, model-certified path that does not depend on a bureau footprint they have not yet earned. Without it, "rejected" means "rejected forever." With it, "rejected" means "not yet."

There is a second, quieter benefit that compounds. Properly governed recourse with RVDR monitoring closes a feedback loop that improves the model itself. Borrowers who follow valid causal recourse, reapply, and are approved generate exactly the data the model is starved of: positive outcomes in the underserved segment, labelled with the actions that produced them. Each successful reapplicant is a new, well-attributed positive training signal in a region of feature space the model previously saw only as rejections. Over successive versions, the model learns the genuine relationship between actionable improvement and creditworthiness for thin-file segments — rather than perpetually extrapolating from prime-borrower patterns. Recourse stops being a cost centre and becomes a data-acquisition strategy for financial inclusion.

This is why broken recourse is both a regulatory failure and an economic waste. The regulatory failure is obvious: arbitrary, expired, unmonitored guidance that the framework was written to prevent. The economic waste is subtler and larger — every thin-file borrower turned away with a meaningless instruction is a creditworthy customer the lender failed to convert and a positive training example the model never received. The institution loses the loan, the borrower loses the pathway, and the model loses the chance to get better at exactly the segment that represents its growth.

Recourse is not the polite closing line on a rejection. It is the load-bearing structure connecting a model's decision to a borrower's future — and like any structure, it must be engineered, inspected, and maintained, or it falls down on the people standing under it.

FAQ

What is the difference between correlational and causal recourse? Correlational recourse is read off feature-importance rankings and names outcomes ("improve your credit score") rather than actions. Causal recourse solves a counterfactual optimisation — the lowest-cost change to actionable features that flips the decision — and returns specific target values the applicant can act on.

Why does feature mutability matter for recourse? Recourse must reference only actionable features (utilisation, GST consistency, cash flow). Immutable features (age, geography) and slowly-mutable ones (credit-history length) cannot be the basis of guidance, because instructing someone to change them implies a pathway that does not exist.

What is RVDR (Recourse Validity Decay Rate)? RVDR(t) is the fraction of previously-valid recourse recommendations that still clear the current model and threshold. It starts at 1 at issuance and decays with every retrain, threshold change, or schema update. Falling below a defined level (e.g. 0.85) should trigger mandatory re-issuance.

What happens if a borrower follows recourse that silently expired? If the institution issued the guidance, the borrower followed it, and it was invalidated without notice, the lender faces real exposure under RBI fairness expectations — the failure was foreseeable because the institution controlled the model and threshold. RVDR monitoring with expiry notification and re-issuance converts this into a defensible, auditable process.

How does recourse support financial inclusion? For thin-file borrowers, causal recourse over non-bureau actionable features is the only entry ramp into formal credit. Successful reapplicants who followed valid recourse also generate positive training signal in underserved segments, improving future model versions.