Use casesFinancial institutions & fintech

Next best offer

Every customer gets the same campaign, and the relationship manager does not know what changed for this one.

Kinds of AI
ModelsSLMs
Who decides
Relationship manager
FIG. 1 · Next best offerIllustrative
RetailSalaried, long-standing customer

Model · ranked offers

OfferValueIncremental P
1 Card limit increase (band C)eligible, pre-approved · service message OKhighmedium-high
2 Auto loan refinanceneeds an income document · İYS: SMS not permitted, call OKmediummedium
3 Savings producteligibility n/alowlow

SLM · RM talking points, draft

Your salary has been coming in steadily…

  • Model
  • SLM

Why a person decidesNothing is sent automatically; the relationship manager decides.

DecidesRelationship manager

1. The desk

Today

  1. CheckOffers ignore credit eligibility until after the customer says yes.
  2. RankCampaigns go to whole segments; response is low and opt-outs rise.
  3. BriefThe relationship manager (RM) does not see that the customer's salary rose or that they hold two cards elsewhere.
  4. DecideThe RM decides whether to present an offer chosen for a segment, not for this customer.

With the system

  1. CheckThe bank's own credit model and policy leave only the products and limits the customer is eligible for.Your rule
  2. RankPropensity models estimate the incremental acceptance per product, and ranking keeps one offer per customer inside the contact caps and the İYS permission record.Model
  3. BriefThe SLM writes talking points for the RM from the supplied facts only, with what to avoid saying.SLM
  4. DecideThe RM decides whether and how to present the offer, as before; nothing is sent automatically, and the customer decides.Person decides

What lands on the desk

The RM sees one offer the customer is already eligible for, why it ranks first, and whether the customer has given permission to be contacted. The talking points use only the facts supplied, and list what not to say. Nothing is sent until the RM decides.

The product owner sets offer rules and contact caps; credit policy owns eligibility; compliance owns permission rules; the RM decides whether to present; the customer decides. No message is sent automatically.

FIG. 2 · Next best offerIllustrative
RetailSalaried, long-standing customer

Model · ranked offers

OfferValueIncremental P
1 Card limit increase (band C)eligible, pre-approved · service message OKhighmedium-high
2 Auto loan refinanceneeds an income document · İYS: SMS not permitted, call OKmediummedium
3 Savings producteligibility n/alowlow

Model · why #1

  • up: Salary inflow up for 3 months
  • positive: High card use
  • Open-banking data, consented

SLM · RM talking points, draft

Your salary has been coming in steadily…

Do not

  • flaggedMention other-bank details
  • flaggedOffer if a limit offer was declined in the last 90 days
  • Model
  • SLM

Why a person decidesNothing is sent automatically; the relationship manager decides.

DecidesRelationship manager

2. The system

For your technical team — skip it if you like.

How it works

In

  • Own products, balances, transactions
  • Past campaign responses, complaints
  • Consented open-banking data

  1. Model · estimates

    Eligibility gate

    The bank's own credit model and policy

    Reads

    • Everything under In

    Out

    • Eligible products and limits only

  2. Model · estimates

    Propensity models per product

    Calibrated GBMs; uplift where campaigns had random holdouts

    Reads

    • Eligible customers

    Out

    • Incremental acceptance per product

  3. Code · calculates

    Arbitration and ranking

    Reads

    • Propensity × value − risk cost
    • Contact-pressure caps
    • The İYS permission record

    Out

    • One offer per customer
    • Reason codes

  4. SLM · understands

    Talking-points writer

    Grounded only in the supplied facts

    Reads

    • The offer and its reason codes
    • Customer facts

    Out

    • Talking points for the RM
    • What to avoid saying

  5. Decision

    Relationship manager

    Reads

    • The offer and its talking points

    Out

    • The relationship manager decides whether and how to present it
    • The customer decides

No agent is needed: offers land in the RM's queue and nothing is sent automatically, so there is no step for an agent to take.

How the models are trained and tested

Your expert writes the test first. A step the test does not need is skipped, and we say why.

Show the training steps

Propensity and uplift models

  1. Target

    Acceptance within the campaign window.

  2. Method and challenger

    A gradient-boosted model against a logistic model, per product; uplift only where campaigns had randomized holdouts.

  3. Validation evidence

    A persistent holdout measures incrementality.

  4. Monitoring

    Acceptance, and early default of cross-sold credit.

Small language model (SLM)

  1. Test set

    Talking points graded by senior RMs: only supplied facts, the correct reason, nothing forbidden mentioned.

  2. Dataset

    Customer-fact bundles with RM-approved talking points.

  3. Continued pre-training (CPT)Skipped

    Skipped: talking points use everyday banking language.

  4. Supervised fine-tuning (SFT)

    On format and tone.

  5. Reinforcement learning (RL)Skipped

    Skipped: there is no reviewable reward beyond the rule check that every claim maps to a supplied fact.

3. Value and conditions

Where the time goes, and what your institution will measure.

Where the value comes from

  1. Incremental product sales and more products per customer.
  2. Lower campaign cost.
  3. Fewer irrelevant contacts.

What your institution measures

  • Incremental conversion vs. holdout
  • Value per contact
  • Products per customer
  • Opt-out and complaint rate
  • Early default of cross-sold credit
Published elsewhere
Bain describes a North American bank whose credit card offers, reworked with a test-and-learn approach, produced 2.5 times the incremental net revenue relative to previous campaigns.

Bain & Company, "Smarter Bank Pricing to Balance Profits and Risk" (2018)

Measured in the Decision Blueprint

Baseline conversion is measured against a holdout; your institution sets the target.

Conditions

The rules, the data, and the known error, with its guard.

  • Using KKB or Risk Merkezi bureau data to target marketing is legally risky, because bureau sharing under the Banking Law rests on credit-evaluation purposes: use it only inside the eligibility gate, and consented open-banking data for other-bank context (the bank's legal owners decide).
  • Marketing messages need an İYS permission record (Law 6563); profiling for marketing needs KVKK explicit consent, and in the EU the GDPR Art. 21 right to object applies.
  • Limit increases need a creditworthiness check.

Start here

We take a recent campaign with its holdout, measure the incremental response it really had, and check which data your consent and permission records let each step use.

Request a Decision BlueprintWhat the Decision Blueprint covers
  • ModelsSLMsAgentic AI

    SME credit referrals

    TodayReferred files wait while analysts rebuild the firm from four systems and last year's committee notes.

    With the systemOne file with the rating, its reasons, and the authority level.

    Credit manager decides

  • ModelsSLMsAgentic AI

    Early warning on the loan book

    TodayThe borrower is current; other banks are already cutting its limits, and you see its quarter only at renewal.

    With the systemA file, not a spreadsheet: signals checked in parallel and gathered in one place.

    Watch-list committee decides

Equalize the gap.

Bring one decision. Leave knowing the first step.

A short description of the decision is enough; no confidential records.

Start with a Decision BlueprintTalk to the engineer

https://eqlx.ai/use-cases/next-best-offer · Printed copy; the website is the current version.