An AI banking insight is not useful because it was generated, delivered, or opened. It is useful only when a customer can understand it, control it, and take a step that is plausibly better for them. That becomes harder to judge when the same notification surface carries a cash-flow warning, an educational tip, and an offer for a bank product.
On September 3, 2026, Truist said that its Insights experience had generated more than two billion personalized insights and advice since 2023 for consumer, small-business, and wealth clients. The bank describes placement in account details, planning dashboards, Truist Assist, and, for urgent items, push notifications. It also says it measures engagement by an opened insight followed by a next step, such as viewing detail or navigating to a recommended action.
Those are useful product signals, but they are not a complete customer-outcome measure. An open can mean relevance, curiosity, or accidental attention. A next step can be a sensible financial action, a product funnel, or a customer trying to make the message go away. This guide provides an intervention-outcome ledger for a bank or fintech team to test those distinctions before scaling proactive AI messages.
It is a governance and measurement workflow, not a recommendation to buy a financial product or take a financial action.
Start by separating three kinds of messages
Do not let one generic insight label hide materially different purposes. Classify a proposed message before it reaches a customer.
| Class | Customer-facing purpose | Example | Minimum control |
|---|---|---|---|
| Helpful insight | Help the customer notice or understand a change | “Your electricity bill is higher than your recent monthly range.” | Show source period, allow dismissal, and avoid an embedded sales claim |
| Product offer | Invite the customer to consider the institution’s product | “See whether this savings product fits your goal.” | Label commercial relationship, apply eligibility and marketing rules |
| Safety or servicing notice | Help prevent harm or complete required servicing | “A payment is due tomorrow” or “We detected a sign-in from a new device.” | Define urgency, delivery channel, and escalation or support route |
The classifications can share data and delivery infrastructure, but they should not share a success metric by default. A safety notice may be judged by timely resolution. A helpful insight should be judged by whether it improved a customer task without creating avoidable burden. An offer has a commercial purpose and needs that relationship made visible rather than implied.
This is an application of a broader control principle from AI financial analysis: input, interpretation, approval, and action should remain independently reviewable. Here, the equivalent separation is observation, message purpose, delivery decision, customer response, and any product journey.
Why sends and clicks are not enough
Delivery metrics answer operational questions: did the service run, did the message render, and did the customer open it? They do not establish that the message helped. Treat them as diagnostics, not proof of value.
| Metric | What it can show | What it cannot show alone |
|---|---|---|
| Delivered | The chosen channel accepted the message | The message was appropriate or seen |
| Opened | A customer viewed the message | The message was understood or wanted |
| Clicked | A customer navigated somewhere | The destination served the customer’s interest |
| Dismissed | The message was closed | Whether dismissal reflects irrelevance, completion, or overload |
| Opted out or suppressed | A preference or frequency boundary was expressed | Whether the system prevented future frustration |
| Customer-confirmed outcome | The customer reports a completed or useful task | Whether the result was caused by the message without a comparison group |
Truist’s published satisfaction and engagement statements are the bank’s own product reporting, not an independent causal study. That does not make them unhelpful; it means a team should retain the measurement definition, period, eligible population, and comparison method before treating a number as evidence of customer benefit.
For a personal-finance example, a customer who opens a subscription reminder and cancels an unwanted service may have experienced a useful intervention. A customer who opens the same reminder every month, never changes anything, and later turns off notifications signals a different outcome. The ledger below is designed to preserve that difference.
Build the intervention-outcome ledger
Create one immutable record per candidate, delivery, and observed response. Do not overwrite the original message when later data arrives; append events with timestamps and a responsible owner.
| Field | Record | Why it matters |
|---|---|---|
insight_id and version | Stable identifier, model/rule version, prompt or policy version | Supports replay after a model or rule changes |
| Observation | Source account data, source timestamp, calculation, and data-quality state | Lets reviewers test whether the message had a factual basis |
| Classification | Helpful insight, offer, or safety/servicing notice | Prevents commercial intent from being disguised as advice |
| Expected customer benefit | Plain-language task the message is intended to help with | Makes success testable before delivery |
| Commercial relationship | Product, referral, affiliate, or revenue connection; otherwise none | Gives compliance and reviewers a concrete disclosure check |
| Eligibility and exclusions | Customer segment rule, required consent, vulnerability or suitability constraints | Stops inappropriate targeting from looking like personalization |
| Delivery decision | Channel, time, urgency, frequency count, and reason a send was allowed | Makes contact pressure measurable |
| Customer controls | Dismiss, mute, opt-out, preference center, and support path shown | Treats control as a product requirement, not a footer |
| Outcome events | Open, detail view, correction, task completion, complaint, suppression, or no observed action | Separates engagement from a fuller outcome history |
| Review and retention | Owner, escalation result, source retention rule, and review date | Creates an evidence trail for testing and complaints |
The record should point to a source snapshot or durable identifier rather than copy a full transaction history into an analytics table. Data minimization helps keep the measurement system from becoming a second uncontrolled customer-data store. When an insight is uncertain, mark the confidence as uncertain and route it to a non-urgent channel or suppress it; do not present a guess as a fact.
Apply a suppression matrix before delivery
The most important decision may be not to send. A frequency cap alone is weak because it treats a low-value sales prompt and a potentially urgent servicing notice as the same event. Use a small, explainable matrix.
| Condition | Helpful insight | Product offer | Safety or servicing notice | | --- | --- | --- | | Customer opted out of this category | Suppress | Suppress | Use only required or authorized channels | | Same insight unresolved recently | Suppress or wait for a material change | Suppress | Escalate only if deadline or risk changed | | Source data is stale or incomplete | Suppress | Suppress | Fall back to verified servicing flow | | Model cannot explain observation | Suppress | Suppress | Route to rules-based or human review | | Customer recently dismissed similar message | Reduce frequency and require a new trigger | Suppress | Reassess urgency and channel | | Message contains product relationship | Keep out of helpful-insight surface unless clearly relabeled | Require offer disclosures and eligibility check | Not applicable |
Make the reason observable in the ledger: preference, recent-dismissal, stale-data, unsupported-explanation, or frequency-cap. A reviewer should be able to answer why a message was sent and why a similar message was withheld without asking a model to reconstruct its own past reasoning.
Use an outcome scorecard that cannot be gamed by a click
Define outcome measures by message class, then pair them with burden and control measures.
INTERVENTION-OUTCOME SCORECARD
Population and measurement period:
Message class and intended customer task:
Comparison group or baseline:
Customer benefit evidence
- Verified task completion or customer-confirmed usefulness:
- Correction rate: customer says the observation was wrong or misleading:
- Repeat exposure before a meaningful change:
Customer burden and control
- Dismissal rate and time-to-dismiss:
- Mute/opt-out rate after exposure:
- Complaint or support-contact rate:
- Suppression rule triggered and honored:
Commercial separation
- Number of helpful-insight messages that contained a commercial relationship:
- Offer disclosure and eligibility check completion:
- Product-navigation rate reported separately from helpful-outcome rate:
Decision
- Continue, revise, suppress, or escalate:
- Evidence owner and next review date:Avoid a single composite score that lets high clicks compensate for high opt-outs or corrections. A helpful-insight program can be smaller and more successful when it sends fewer messages, respects a dismissal, and produces a clearer customer task outcome.
The NIST AI Risk Management Framework and its Playbook are useful references for turning this into a governed practice: map the context, measure outcomes and failure modes, manage the response, and assign governance ownership. They do not replace consumer-protection, privacy, marketing, or institution-specific obligations.
A copyable prompt for message design and review
Use a model to draft a controlled record, not to decide that a customer should receive a product pitch.
You are drafting a proposed proactive banking message for internal review.
You may not make a financial recommendation, infer a sensitive trait, or decide to send.
Inputs
- Observation and source timestamp: [provided data]
- Data-quality state: [complete / stale / uncertain]
- Proposed purpose: [helpful insight / product offer / safety-servicing]
- Customer preferences and recent-message history: [provided]
- Applicable eligibility, consent, and disclosure rules: [provided]
Return exactly these sections:
1. Plain-language observation, including what is directly observed versus inferred.
2. Classification and expected customer benefit.
3. Commercial relationship: none, or name the product/referral relationship and required disclosure.
4. Delivery recommendation: send only if all preconditions are met; otherwise return SUPPRESS with a reason.
5. Customer-control text: dismiss, mute, preference, and support route.
6. Ledger fields: source timestamp, policy version, proposed frequency count, confidence, and review owner.
7. Failure checks: stale data, duplicate message, unexplained inference, missing eligibility evidence, and misleading wording.
If any required input is absent or uncertain, return SUPPRESS or HUMAN REVIEW.Test this prompt with at least five cases before use: a stale balance, a repeated dismissed notification, a message with an undisclosed product link, a customer who has opted out, and a genuine urgent servicing event. The expected result is not always a polished message. In several cases, the correct result is silence, a rule-based notice, or a human review queue.
Keep advice, explanation, and execution apart
An insight can explain an observed change without telling a customer what to do. “Your recurring charge increased” is different from “move money from savings” or “apply for this credit product.” The latter can introduce suitability, approval, and product-specific obligations that are not solved by better prompt wording.
For connected-account experiences, retain the same boundary used in AI and Banking: PicPay vs ChatGPT Finances: permission to read data does not establish that an interpretation is correct, nor does it grant transaction authority. The intervention-outcome ledger adds a related question: did the institution earn the right to interrupt the customer at all?
Review a sample of sent, suppressed, dismissed, and complained-about messages together. If an insight is factually accurate but repeatedly dismissed, the product should learn from that evidence. If a product offer is useful, say that it is an offer. Clear purpose, real controls, and a durable outcome record make proactive banking more accountable than a high send count ever can.
About Enis
AI Engineer specializing in Machine Learning and LLMs. Combining Computer Engineering and Economics to build data-driven financial tools.