> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kaireonai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Banking & Financial Services
> Cross-sell and retain across a regulated book of business — with scoring you can explain to a regulator and guardrails that keep every offer suitable.
## The problem in this industry
Banks sit on the richest customer data of any industry — balances, transactions, tenure, credit history — and yet most cross-sell still feels like spam. The reason is trust. A customer who gets pitched a high-fee product they obviously don't qualify for, or a credit card the week after a fraud scare, doesn't just ignore it; they lose a little faith in the institution. In financial services the cost of a bad recommendation is measured in relationship erosion, not just a wasted impression.
Layered on top of trust is regulation. Suitability rules, fair-lending obligations, and disclosure requirements mean you can't simply let an opaque model decide who sees a loan offer. When a customer — or an auditor — asks "why was I shown this?", "the algorithm decided" is not an acceptable answer. You need a decision you can reconstruct and defend, feature by feature.
KaireonAI is built for exactly this tension. Hard eligibility gates enforce the non-negotiables (credit thresholds, KYC status, product ownership) before anything is scored. Scoring can run on a transparent, per-feature-explainable model where the regulation demands it, and a richer model where it's allowed. And every decision leaves a trace that shows which gate a customer passed, which policy suppressed an offer, and why the winner ranked first.
## What you build in KaireonAI
You express your product catalog, compliance rules, and channel mix as a small set of platform building blocks. The table below shows how a typical retail-bank setup maps on:
| Platform concept | How you use it in banking |
| ------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Offers** | Premium credit card, balance-transfer offer, personal loan, mortgage refinance, high-yield savings / CD, wealth-advisory intro, overdraft protection, card-limit increase |
| **Channels** | Mobile banking app, online banking, email, branch tablet, call-center script, SMS |
| **Decisioning gates** (Eligibility → Fit → Match) | *Eligibility*: `credit_score >= 650`, `kyc_verified == true`, and suppress a product the customer already holds (`offer_attribute`). *Fit*: income and debt-to-income thresholds for a loan (`attribute_condition` on `annual_income`, `dti_ratio`). *Match*: boost offers aligned to a customer's inferred financial goal |
| **Contact policies** | `frequency_cap` (a customer sees at most a couple of marketing offers per week); `cooldown` between loan pitches; `outcome_based` 90-day quiet period after a complaint; `offer_category_cap` so acquisition messaging doesn't crowd out service messaging; `do_not_contact` for regulatory opt-outs |
| **Scoring approach** | Use a [`scorecard`](/ai-ml/algorithms/scorecard) or [`logistic_regression`](/ai-ml/algorithms/logistic-regression) where explainability is mandatory — both produce per-feature contributions you can put in writing. Where richer prediction is permitted (e.g. deposit-product propensity), a [`gradient_boosted`](/ai-ml/algorithms/gradient-boosted) model captures interactions linear models miss, with [SHAP](/ai-ml/shap) explanations on top |
To avoid over-selling the "sure things," add an [uplift model](/ai-ml/uplift-modeling): it estimates the *incremental* effect of an offer per customer, so the platform can down-rank a card a customer would have opened anyway (a "sure thing") and prioritize the truly persuadable.
## A worked example
**David Chen** is a 7-year customer with a checking and savings account, a `credit_score` of 742, `annual_income` of 95,000, and no credit card with the bank. He recently filed a complaint about a fee, resolved two weeks ago.
Premium card, balance transfer, personal loan, mortgage refi, high-yield CD, wealth-advisory intro, overdraft protection, limit increase.
Overdraft protection drops (he rarely overdrafts — a fit rule). The card-limit-increase offer drops (he has no card to increase). Six survive the hard gates; his credit score clears the 650 floor and KYC is current.
An `outcome_based` policy enforces a 90-day quiet period after his complaint for *acquisition* categories, so the balance-transfer and personal-loan pitches are held back. Service-oriented and low-pressure offers remain eligible.
A logistic-regression propensity model scores the survivors. Ranking blends [PRIE](/decisioning/prie-formula-design), and the uplift term suppresses the wealth-advisory intro (he'd engage regardless) in favor of the offer where the bank actually moves the needle: the premium credit card.
The [decision trace](/api-reference/decision-traces) shows he passed the credit gate at 742, why the loan offers were suppressed (complaint quiet period), and the per-feature contributions behind the card's score — exactly what a suitability review needs.
**Change one attribute and the decision changes.** Run the same flow for a customer with a `credit_score` of 610 and the premium card and loan offers never survive the eligibility gate — they'd instead see a secured-card or savings offer. Give David a clean complaint history and the balance-transfer and personal-loan pitches come back into contention. The rules are fixed and auditable; the outcome adapts to each customer's real situation.
## Measuring success
* **[Business Dashboard](/operations-reporting/dashboards#business-dashboard)** — product-level acceptance, conversion, and revenue per offer.
* **[Model Health Dashboard](/operations-reporting/dashboards#model-health-dashboard)** — AUC trend and feature importance, plus the built-in **Fairness** panel (demographic-parity gap, disparate-impact ratio, four-fifths check) so you can evidence fair treatment across groups. See [Fairness & Drift](/ai-ml/fairness-drift).
* **[Executive Dashboard](/operations-reporting/executive-dashboard)** — a narrated weekly summary of the book-wide numbers for leadership.
* **Uplift and holdout** — an [experiment](/ai-ml/uplift-modeling) with a control group proves incremental balances and revenue rather than borrowed conversions.
* **Suppression rate** — confirms your compliance quiet-periods and frequency caps are actually firing.
## Where the agentic layer helps
[Decisioning Autopilot](/ai-ml/autopilot) turns experiment results and model drift into concrete, reviewable proposals — "promote this challenger propensity model, it beats the champion with significance" — that a risk owner approves through four-eyes review before anything changes. In banking you almost always want the `suggest` or `auto_gated` mode so a human signs off, and the audit log captures every application.
[Decision Sentinel](/ai-ml/sentinel) guards against the silent failure that regulated teams fear most: a mis-scoped policy or an expired offer schedule that quietly stops decisions from going out. It watches the decision stream for suppression spikes and empty-decision rates and alerts before a channel goes dark. Combined with the [governed AI assistant](/ai-ml/ai-assistant) — which lets a product manager describe a change in plain language and routes it through approvals rather than editing production — you get faster iteration without loosening controls.
## Try it
Build an uplift-aware cross-sell flow that avoids pitching products a customer already owns.
Retention scoring, qualification gates, and contact frequency end to end.
Apply the BFSI starter kit: 15 offers, credit-score gates, and a ready decision flow.
Register and build a banking decision flow end to end.