AI Product Design & Innovation · Individual project

BIDV-NBA

Explainable AI Next-Best-Action recommendations for retail banking sales and customer care.

Who first?Why now?What next? Dang Dam Linh · Business Analyst · Sep 2026
Photo: Saigon at night, Minh Khiem, Wikimedia Commons, CC0
The problem

RMs have lots of data but no single, trusted answer to three questions.

Who first?

Several signals compete for the same morning: maturities, balance drops, overdue items, card expiries, callbacks.

Why now?

Due dates don't compare urgency, value, service risk and recent contact.

What next?

The action must fit the evidence, current policy and what already happened.

Data abundance creates decision scarcity.

Photo: Santeri Viinamäki, Wikimedia Commons, CC BY-SA 4.0
Business context

The data already exists. What's scarce is the RM's attention.

Relationship Manager
Primary user. Reviews recommendations, decides, contacts the customer, records the outcome.
Team manager
Coaches usage and quality. Adoption is not a contact-volume quota.
Product / care owner
Approved action catalogue, eligibility, priority and suppression rules.
Data / AI team
Candidate generation, ranking, explanation, versions and monitoring.
Risk & Compliance
Data-use boundaries, controls for sensitive actions, review of inappropriate outcomes.

For this project BIDV is treated as not yet having an AI Next-Best-Action capability. AI features shown are proposals.

Current state (as-is)

The pain sits between “signals available” and “action chosen”.

SYSTEM OF RECORD — KEEP PAIN ZONE — MANUAL SYNTHESIS OF SIGNALS SYSTEM OF RECORD — KEEP 01Reviewwork queue 02Open customercontext 03Cross-checksignals 04Choose whoto contact 05Decidethe action 06Prepareconversation 07Contactcustomer 08Log result& follow-up Queues compete;no common priority Available, butnot summarised Recent contacteasily missed Differs by RMand by day Weak link toevidence & policy Repeatednavigation Wrong timing,duplicate contact Feedback not turnedinto learning PAIN ZONE 01Review work queueQueues compete; no common priority 02Open customer contextAvailable, but not summarised 03Cross-check signalsRecent contact easily missed 04Choose who to contactDiffers by RM and by day 05Decide the actionWeak link to evidence & policy 06Prepare conversationRepeated navigation 07Contact customerWrong timing, duplicate contact 08Log result & follow-upFeedback not turned into learning

Steps 1–2 and 7–8 are solid system-of-record capabilities, and they stay. We redesign only the decision stage.

Root cause

It's a gap in how decisions get made, not a lack of information.

Fragmented signals
→
Repeated navigation
→
Inconsistent prioritisation
→
Duplicate or mistimed contact
→
Weaker employee trust
→
Limited learning

Five whys, one answer

No shared decision layer applies common ranking and suppression; context isn't packaged around the action; feedback is logged as an outcome, not as a quality signal.

What must change

Turn available data into an explainable, governed, feedback-driven decision product inside the RM workflow. Keep human judgement for the decision itself instead of spending it on assembling information.

Design thinking

Start from how the RM decides.

RM decision journey 01Understand 02Observe 03Define 04Ideate 05Prototype 06Test
Primary persona · provisional
Nguyen Minh Anh

Retail Relationship Manager · 31 · 6 years in retail banking

Goal
Know which customers need attention today, and why.
Pain
Signals compete, and she repeats the same preparation for every customer.
Trust need
Evidence, timing and policy context. She pushes back on suggestions that ignore a recent call.
Empathy map · Relationship Manager

“I don't need ten more alerts.”

Says

“I need to know which few customers matter today.”

Thinks

“Can I trust the reason? Has someone already called? Is this policy still valid?”

Does

Checks tasks, profile, events and history before deciding. Reopens policy documents for the important cases.

Feels

Pressured and wary of black-box suggestions. Reassured by evidence, fresh data and the option to say no.

How might we

explain a recommendation in under 30 seconds using facts the RM can verify?

Ideation

Five ideas, one selected: an explainable ranked NBA inside the CRM.

HighMed-HighMedium LowMediumHigh PILOT FEASIBILITY → USER VALUE → Autonomous outreachEXCLUDED Explainable ranked NBASELECTED Rules-only priority queueBASELINE / GUARDRAIL General AI chat copilotNOT MVP Filters & dashboardsCOMPONENT ONLY
Filters & dashboards
Component
Rules-only priority queue
Guardrail
General AI chat copilot
Not MVP
Explainable ranked NBA
Selected
Autonomous outreach
Exclude

Rules handle hard eligibility, permission and suppression. AI is used where ranking and combining patterns adds value.

Future state (to-be)

AI proposes. The RM decides. The system learns.

AI DECISION LAYER — PROPOSES RM + EXISTING CRM — DECIDES & ACTS HUMAN CONTROL BOUNDARY 01PermittedsignalsProfile, events,history 02Guardrails &suppressionRules run first 03RankingengineEligible pairs only 04Why-now &groundingEvidence, freshness,policy 05RM reviewAccept · reject ·snooze · alt. 06Create CRMactivityOnly after RMconfirms 07ContactcustomerApproved channel;RM owns words 08 Feedback & audit Reason codes · outcome · evidence timestamp · model/rule version AI DECISION LAYER — PROPOSES 01 Permitted signalsProfile, events, history 02 Guardrails & suppressionRules run first 03 Ranking engineEligible pairs only 04 Why-now & groundingEvidence, freshness, policy HUMAN CONTROL BOUNDARY RM + EXISTING CRM — DECIDES & ACTS 05 RM reviewAccept · reject · snooze · alternative 06 Create CRM activityOnly after RM confirms 07 Contact customerApproved channel; RM owns words ↓ 08 Feedback & audit ↺Reason codes · outcome · evidencetimestamp · model/rule version

Rules run before ranking. Every decision is logged with its reason code, evidence timestamp and model/rule version.

Solution

Nine components. Rules sit before the AI.

Data
Permitted inputs
1

Signal layer

Profile, events, products, activity

2

Customer insight package

Persona/need, NBO, churn — hypotheses

Control
Deterministic rules
3

Candidate-action library

3–4 approved action types for the pilot

4

Guardrail & suppression

Permission, eligibility, recency, expiry

Intelligence
AI ranks & explains
5

Ranking engine

Urgency, relevance, outcomes, workload

6

Explanation & policy grounding

Why-now from traceable facts

Workflow
Human-confirmed
7

CRM execution integration

Create next activity only after the RM confirms — drafts are never auto-sent

Learning
Governed improvement
8

Feedback & audit

Review action, reason, outcome, versions

9

Monitoring

Adoption, rejection, drift; pause by type

Avoiding irrelevant recommendations

Fewer, but better: 3–5 a day.

All permitted customer signals
Approved pilot action types only
Permission · eligibility · consent
Recent contact · dedupe · freshness · policy expiry
Minimum relevance threshold
Daily cap — top 3–5 per RM

Illustrative, not to scale. RM snooze and “not relevant” feedback feed back into monitoring and tuning.

Prototype · four low-fidelity screens

Daily queue → why now → customer context → learning.

A · MY PRIORITIES TODAY
My priorities today · 3 of 3
1Nguyen A.
Term deposit matures in 5 days
→ Maturity follow-up
HIGH
expires 2 days
2Tran B.
Balance declined materially
→ Retention / service check
MEDIUM
expires 3 days
3Le C.
Card expires next month
→ Card-renewal support
MEDIUM
expires 7 days
B · RECOMMENDATION DETAIL
Recommended next action
Discuss deposit-maturity options
Why now?
• Term deposit matures in 5 days
• Premier customer, managed by you
• No related contact in the last 14 days
Evidence
Source: customer event · CRM history
Data freshness: today 07:00 · Rule v1.2
AcceptRejectSnoozeAlternative…
C · CUSTOMER 360 — NBA WIDGET
Customer 360
CIF ****1234 · Segment: Premier · RM: you
Next best action
Retention check after significant outflow
Monthly balance declined; no recent follow-up
Signal: churn-like · Action horizon: 3 days
Recent activity
12 Sep · Service inquiry — completed
04 Sep · Loan maturity reminder — completed
Create activity
D · TEAM LEARNING DASHBOARD
Pilot quality & adoption · week 8
—Adoption
—Rejected
—Median prep
0Severe incidents
Top rejection reasons
Already contacted
Not relevant
Wrong timing
Missing context
illustrative — baseline pending

BIDV-NBA · fictional placeholder data

Agile thinking · MVP backlog

Keep version one small enough to test.

Must
  • Ranked “My Priorities Today”
  • “Why now?” evidence and freshness
  • Accept · reject · snooze · alternative
  • Rules run before ranking
  • Next activity after RM confirms
Should
  • Persona, product fit and churn signal
  • Approved policy with expiry and reference
Could
  • Optional talking points for RM review
  • Notification when async analysis is ready
Won't (MVP)
  • Autonomous outreach or automatic product application

Every 2-week cycle: review rejection reasons, false positives, stale-data cases, interviews, prep time and incidents, then re-prioritise by evidence.

Critical assumptions

Six things that need to be true, and how we'll test them.

Ranked list
cuts prep time
→ baseline vs. pilot, median by action type
Explanations
build appropriate trust
→ adoption, rejection reasons, trust survey
A small daily list
beats many alerts
→ test caps, thresholds and expiry
Suppression
removes duplicate prompts
→ historical replay and shadow review
RM feedback
produces usable labels
→ 1-tap reasons, completion and consistency
Policy references
stay authoritative
→ approved links with expiry; block the action type if they fail
Controlled pilot

One Premier team, ten weeks, evidence before exposure.

8–12

Relationship managers

250–400

Premier customers

3–4

Maturity · card expiry · retention · overdue support

10 wks

Then scale, redesign, narrow or stop, per action type

W1
W2
W3
W4
W5
W6
W7
W8
W9
W10
1 Discovery & baseline
2 weeks
2 Prototype test
1 week
3 Historical replay
2 weeks
◆ go / no-go for shadow
4 Shadow mode
2 weeks
5 Controlled live pilot
3 weeks
6 Retrospective / scale gate
Scale · redesign · narrow · stop
No AI-driven customer contact
Live, RM-reviewed
Measurement

Five KPIs. Targets come after the baseline.

↓

Median prep time

without quality incidents

↑

Adoption rate

a healthy rise, not a quota

↑=

Outcome / conversion

improves, or holds while saving time

↓

Wrong-context rejections

not relevant, already contacted, wrong timing, missing context

0

Severe inappropriate outcomes

the primary safety constraint

Trust is tracked with a short recurring survey and interviews, never inferred from acceptance alone.

Risks

The biggest risks are about people and process, not just the model.

High
R9
R3R4R5R6
R2
R1
Medium
R8
R7
Low–Med
Medium
Med–High
High
Impact ↑ · Likelihood →
R1
Alert fatigue → caps, thresholds, expiry, dedupe
R2
Stale or missing data → freshness badge, hard suppression
R3
False confidence from explanations → cite evidence only
R4
Automation bias → mandatory review, audits
R5
Outdated policy → expiry and link monitoring
R6
Bias across groups → segmented fairness review
R7–9
Label noise, misread adoption, untraceable changes → taxonomy, interviews, versioning
Governance · human in the loop

AI can propose. It can't make a binding customer decision.

The AI may
  • Prioritise eligible customer–action pairs
  • Summarise context and recent activity
  • Explain why now, with evidence and sources
  • Retrieve approved policy references
Only the RM and authorised processes
  • Decide whether, how and what to say to the customer
  • Approve credit, pricing and exceptions
  • Override permission, consent or suppression
  • Interpret eligibility and handle low-evidence cases
Business value

Data that's ready for decisions, not just available.

Experience

One ranked queue with evidence instead of several lists

Process

Guardrails + ranking + explicit human review

People

RM judgement becomes visible and learnable

Data

Signal, freshness, policy and outcome around each pair

Operations

Monitoring by action type; pause what doesn't work

No ROI claim yet: a cost-aware plan and a baseline, not a made-up payback period.

Photo: Diego Delso, Wikimedia Commons, CC BY-SA 3.0
Conclusion

Scale evidence,
not enthusiasm.

“BIDV-NBA doesn't replace the Relationship Manager. It helps the RM focus their attention, understand why, and choose the next action with better context.”

Thank you · Q&A
Photo: Hanoi skyline with Ba Vi Mountain, Quangnlnhe182394, Wikimedia Commons, CC0
Appendix · AI use disclosure

AI supported the work. The author owns the decisions.

ChatGPT (OpenAI)
Helped structure the analysis, English wording, tables, the Design Thinking artefacts, backlog, KPI and risk layouts, and low-fidelity visuals.
Claude (Anthropic)
Designed the layout, SVG diagrams and slides for this edition, working from the author's draft.
Images
Saigon at night: Minh Khiem (CC0) · ATM PIN buttons: Santeri Viinamäki (CC BY-SA 4.0) · Ho Chi Minh City from Bitexco Tower: Diego Delso (CC BY-SA 3.0) · Hanoi skyline with Ba Vi Mountain: Quangnlnhe182394 (CC0), via Wikimedia Commons.
Author
Owns the framing, banking context, design choices, risk boundaries and final review. No production customer data was used, and none of the KPIs are claimed as achieved.
← → · space · F fullscreen