“Build a generative AI chatbot that reduces customer-service costs” is a common executive request. It is not yet a strategy, a problem statement or an investable use case. This article walks through Part I of the End-to-End AI Solution Engineering Playbook—strategy and discovery—using a realistic retail-banking customer-service scenario.

The worked example is MonGo Bank: a hypothetical retail and small-business bank with millions of customers, a large contact centre, mixed cloud and legacy platforms, and strict regulatory obligations. The goal is not to pick a model. It is to decide where AI should play, how the bank wins, what must be true and what evidence is required before further investment.

---

Part I: Scenario overview

Organisation​

MonGo Bank operates through mobile banking, online banking, branches and a contact centre. It has:

• 4.2 million retail customers
• 850,000 active mobile-banking users
• 2,400 customer-service employees
• Approximately 3.1 million service contacts each year
• A mixture of modern cloud services and legacy core-banking platforms
• Strict regulatory, security, privacy and conduct obligations

Current customer-service problem​

Customers contact the bank about card activation, lost or stolen cards, unrecognised transactions, payment status, direct debits, account details, password resets, address changes, product eligibility, fees and charges, complaints, financial difficulty and suspected fraud.

Measure	Current position
Average telephone wait	11 minutes
Average handling time	8.4 minutes
First-contact resolution	68%
Customer satisfaction	71%
Annual service interactions	3.1 million
Estimated repetitive interactions	62%
Annual contact-centre operating cost	$74 million
Knowledge search time per agent	95 seconds
Transfers between teams	18% of contacts
Complaint-related contacts	8%

Why the initial AI idea is inadequate​

The executive request assumes:

• A chatbot is the correct solution
• Cost reduction is the only important outcome
• Customers will accept automated support
• The bank has suitable data and knowledge
• Generative AI is appropriate for every service request
• Regulatory and conduct risks are manageable
• Existing operational problems are caused by the absence of AI

Strategy and discovery must test these assumptions.

Potential target solution​

The eventual solution may contain several components:

1. Customer self-service assistant — answers approved questions inside mobile and online banking
2. Agent-assist copilot — retrieves policies, suggests responses and summarises conversations
3. Intelligent routing — identifies intent and routes cases correctly
4. Conversation summarisation — structured notes for service records
5. Risk and vulnerability detection — fraud, financial difficulty or vulnerability signals
6. Human escalation — transfer to authorised employees when automation is unsuitable

The purpose of strategy and discovery is to determine which capabilities to pursue, in what order and under what controls.

---

Part II: Strategy frameworks

1. Strategy Choice Cascade​

The Strategy Choice Cascade translates ambition into five connected choices: winning aspiration, where to play, how to win, required capabilities and management systems. AI strategies often become technology lists; this framework forces explicit business choices.

Winning aspiration​

Weak:

Deploy generative AI in customer service.

Stronger:

Make MonGo Bank the easiest and most trusted regional bank for resolving everyday customer-service needs, while reducing avoidable service effort and maintaining human support for sensitive or regulated situations.

Where to play​

In scope initially: authenticated mobile and online customers; high-volume, low-complexity requests; agent assistance; English-language interactions; retail current accounts and credit cards; existing customers.

Explicitly excluded: autonomous lending; investment recommendations; final complaint adjudication; suspected-fraud decisions; financial-hardship resolution; high-value payment authorisation; account closure; unauthenticated disclosure of customer information.

How to win​

MonGo will not compete on maximum chatbot autonomy. It will compete through trusted answers grounded in approved information, seamless human transfer, controlled personalisation, faster routine resolution, clear AI boundaries, auditability and consistent service across digital and human channels.

Capabilities and management systems​

Required capabilities include intent recognition, authentication, knowledge retrieval, RAG, summarisation, escalation, vulnerability detection, prompt and model management, evaluation, monitoring, incident management, Responsible AI governance, data lineage, access control and service analytics.

Required management systems include an AI governance committee, named system owners, use-case approval, evaluation thresholds, model and prompt change controls, human-oversight procedures, monthly performance reviews, customer-feedback monitoring, incident escalation, risk assurance, benefits tracking and vendor-performance monitoring.

Strategic output​

MonGo Bank will initially use AI to improve routine authenticated customer service and employee productivity. It will differentiate through trusted, explainable and well-controlled assistance rather than maximum automation. Sensitive, regulated and high-impact interactions will remain under human control.

---

2. Playing to Win​

Playing to Win emphasises choices that create advantage and prevents pursuing every AI opportunity at once.

Option	Position	Verdict
A	Maximum automation	Large theoretical cost reduction; high regulatory, reputational and vulnerability risk
B	Employee productivity first	Lower customer-facing risk; customers still wait; demand remains high
C	Trusted hybrid service	Controlled self-service plus agent assist; progressive introduction; human escalation retained

MonGo chooses Option C. Differentiation:

Digital convenience without removing access to human support.

The bank already has strong mobile adoption, a substantial knowledge base, authentication capabilities, mature risk and compliance functions and large volumes of repetitive demand.

---

3. Three Horizons​

Three Horizons separates near-term improvement from medium-term transformation and longer-term innovation.

Horizon 1 (0–12 months): improve the current model​

Initiatives: AI knowledge search for agents, summarisation, suggested responses, better intent recognition, automated classification, improved routing, basic authenticated self-service, evaluation and monitoring.

Target: reduce average handling time from 8.4 to 7.2 minutes.

Input	Value
Agent-handled interactions	2.2 million / year
Time saved per interaction	1.2 minutes
Hours saved	44,000
Loaded cost	$35 / hour
Annual capacity benefit	$1.54 million

This may appear as reduced overtime, avoided recruitment, increased capacity or better quality—not automatically as cash savings. Benefits ownership must be explicit.

Horizon 2 (12–30 months): transform journeys​

Authenticated conversational self-service, proactive notifications, personalised guidance, cross-channel continuity, vulnerability identification, intelligent case orchestration and controlled automated post-interaction actions.

Horizon 3 (30–60 months): reinvent banking service​

Highly personalised assistants, multimodal support, proactive financial-health guidance, multi-agent orchestration and ecosystem-wide real-time service—kept exploratory as regulation, technology and expectations evolve.

Portfolio allocation​

Horizon	Investment
Horizon 1	60%
Horizon 2	30%
Horizon 3	10%

---

4. SWOT and TOWS​

SWOT is useful only when it produces action.

Strengths: authenticated digital base, brand, knowledge assets, compliance experience, repeatable volume, digital platforms, historical service data.

Weaknesses: inconsistent knowledge, legacy integrations, fragmented records, limited AI engineering, manual QA, slow policy updates, incomplete lineage, separate channel technologies.

Opportunities: lower customer effort, higher employee productivity, digital self-service, earlier vulnerability detection, reusable AI platform capabilities, better analytics.

Threats: hallucinations, prompt injection, data exposure, regulatory non-compliance, customer exclusion, vendor dependency, model-price increases, reputational damage, competitive digital service.

TOWS actions​

• Strength–Opportunity: use authenticated mobile adoption for controlled routine self-service
• Weakness–Opportunity: govern knowledge before scaling generative AI
• Strength–Threat: use the mature risk function for AI assurance and approval controls
• Weakness–Threat: withhold autonomous AI access to legacy transactional systems until access controls and integration monitoring improve

---

5. PESTLE Analysis​

Force	Implication
Political	Frame the programme as service improvement, resilience and responsible access—not pure headcount reduction
Economic	Model labour capacity and ongoing AI operating costs
Social	Preserve accessible human escalation; do not force all customers into automation
Technological	Support model substitution; avoid single-provider lock-in
Legal	Require traceability, retention controls, supplier due diligence and human oversight
Environmental	Measure inference usage; select models proportionate to the task

---

6. Porter’s Five Forces​

Competitive rivalry and buyer power make service quality strategic. Digital banks raise the threat of new entrants. Model and cloud vendors concentrate supplier power—so multi-model design, contractual protections and exit planning matter. Substitutes (fintech apps, wallets, comparison platforms) raise the bar for integrated, immediate, trustworthy service.

Conclusion: the AI programme is not justified only by cost reduction. It also defends customer experience and reduces dependence on external service ecosystems.

---

7. Value Chain Analysis​

Customer-service value chain stages: need → channel → authenticate → understand → collect → retrieve knowledge → guide → act → document → follow up → feedback → learn.

Activity	Potential AI capability
Request understanding	Intent classification
Information collection	Conversational data capture
Knowledge retrieval	Retrieval-augmented generation
Guidance	Controlled response generation
Routing	Intelligent routing
Documentation	Automated summarisation
Follow-up	Proactive notifications
Feedback	Sentiment and theme analysis
Learning	Conversation analytics

Value leakage found: 95 seconds of knowledge search per contact; 18% transfers; 12% recontact within seven days; 7% notes failing quality checks; multi-day policy propagation. Improving internal knowledge and routing may create more value than launching a fully autonomous chatbot first.

---

8. Business Model Canvas​

Adapted to an internal AI service:

• Segments: retail and digitally confident customers, accessibility-support customers, agents, branch staff, service managers, risk and compliance
• Value: faster answers, 24-hour support, less repetition, seamless escalation; for employees—faster knowledge, less admin, better summaries; for the bank—lower effort, consistency, reusable capabilities, insight
• Channels: mobile, online, contact-centre desktop, telephone, secure messaging, branch
• Key activities: knowledge management, evaluation, model and prompt management, operations, security monitoring, incident management, regulatory assurance
• Cost structure: model usage, cloud, integration, data preparation, knowledge, engineering, evaluation, monitoring, oversight, training, vendor management
• Value measures: contact volume, handling time, first-contact resolution, satisfaction, complaints, employee productivity

---

9. Operating Model Canvas​

• Processes: use-case approval, knowledge publication, prompt change control, model release, evaluation, incidents, escalation, benefits reporting, supplier review
• Organisation: Customer Service owns outcomes; Digital Product owns experience; AI Engineering builds the platform; Data Office governs quality; Cybersecurity approves controls; Risk and Compliance challenge independently; Operations monitors performance; Internal Audit assures periodically
• Information assets: product terms, policies, customer records, transaction metadata, complaints, fraud warnings, vulnerability guidance, conversation records, evaluation results
• Management systems: AI inventory, risk register, model and prompt registries, evaluation dashboard, incident register, benefits dashboard, vendor scorecard

---

10. Capability-Based Planning​

Existing strengths include authentication, contact-centre operations, digital banking, complaints, fraud, data governance, cybersecurity and vendor management.

Largest gaps are evaluation, monitoring and governed knowledge—not model availability.

Capability	Current	Required	Gap
Customer authentication	4	4	0
Knowledge management	2	4	2
AI engineering	2	4	2
AI evaluation	1	4	3
Human oversight	2	4	2
AI monitoring	1	4	3
AI governance	2	4	2
Customer-service analytics	3	4	1

This changes the investment roadmap toward enabling capabilities before aggressive automation.

---

11. Wardley Mapping​

User need: resolve a banking service issue quickly and safely.

Component	Evolution	Treatment
Cloud infrastructure	Commodity	Buy
Foundation model	Product / commodity	Buy with portability
Vector database	Product	Buy or managed service
Customer authentication	Mature internal	Reuse
Generic chatbot interface	Product	Buy or configure
Banking knowledge structure	Custom	Build
Evaluation dataset	Custom	Build
Human-escalation rules	Custom	Build
Vulnerability controls	Custom	Build
AI governance workflow	Custom	Build and adapt

Insight: do not build a foundation model. Invest in banking-specific knowledge, evaluation, integration, human oversight, customer protection and operational controls—where differentiation and risk ownership live.

---

12. Value-Driver Tree​

Top objective: increase customer-service value through four drivers—reduce cost, improve experience, reduce operational risk and increase capacity.

Worked financial branch (Horizon 1): 2.2 million interactions × 1.2 minutes = 44,000 hours × $35 = $1.54 million annual capacity value. Benefits ownership must distinguish cash savings, capacity and cost avoidance.

---

13. Scenario Planning​

Scenario	Response
Controlled evolution	Proceed with hybrid roadmap
Strong regulatory restrictions	Prioritise agent assist and internal productivity
Rapid agentic breakthrough	Accelerate controlled orchestration within approval boundaries
Major AI trust incident	Emphasise transparency, human escalation and independently assured evaluation

No-regret investments across all scenarios: knowledge management, authentication, evaluation, data governance, human escalation, monitoring, model portability and employee AI skills.

---

14. AI North Star​

Every MonGo customer should be able to resolve a routine service need quickly, safely and with the option of human help.

Principles: human support remains available; AI uses approved information; sensitive decisions need human authority; customers know when AI is used; material answers are traceable; performance is measured continuously; automation must not disadvantage vulnerable customers.

North Star metric — successful safe resolution rate: percentage of eligible requests resolved accurately, without repeat contact, policy breach or inappropriate denial of human assistance. Stronger than chatbot containment, which can reward poor outcomes.

---

Part III: Discovery frameworks

15. Design Thinking​

Empathise: 25 customer interviews, 15 agent interviews, 5 vulnerable-customer specialist interviews, contact-centre observation, 1,000 transcript reviews and accessibility testing.

Findings: unclear banking terminology; anxiety about missing payments; dislike of repeating information after transfer; agents struggle to find latest policy; vulnerable customers need reassurance more than speed; customers want confirmation that actions completed.

Define (refined): customers with routine but time-sensitive questions cannot obtain a trusted answer quickly, while agents spend significant time searching fragmented information and recording repetitive details.

Ideate / Prototype / Test: agent search, authenticated self-service, proactive payment notifications, routing, summarisation, shared context, callback scheduling. Customers prefer a clearly bounded assistant that states what it can do and offers human support over one that claims to handle everything.

---

16. Double Diamond​

Discover → Define: narrow first release to authenticated card and payment-status questions via better self-service and agent assistance—not whole contact-centre transformation.

Develop → Deliver: selected concept is a customer-facing RAG assistant for approved routine questions, an agent copilot for complex interactions, human escalation for sensitive situations and proactive status information where possible.

---

17. Stakeholder Mapping​

• High power, high interest: executive sponsor, Head of Customer Service, CDO, CRO, CISO, product owner, regulatory compliance—involve in key decisions and risk appetite
• High power, lower day-to-day interest: board risk committee, finance, procurement, internal audit—formal papers on risk, value and assurance
• Lower power, high interest: agents, team leaders, customer advocates, accessibility groups, data scientists, knowledge managers—deep involvement in discovery and testing
• External: customers, regulators, cloud and model providers, contact-centre vendor, consumer groups

Surface tensions: leadership may want high automation; risk prefers limits; agents may fear job loss; customers may want immediate humans.

---

18. Voice of the Customer​

From 10,000 comments, frequent themes translate into needs: status visibility, context continuity, accessible escalation, consistency, plain language, reassurance and urgency.

Critical-to-quality: plain language; current payment status; human escalation; no forced repetition; immediate identification of high-risk situations; responses consistent with approved policy.

---

19. Jobs to Be Done​

Functional: when I see a card transaction I do not recognise, I want to understand what it is and secure my account, so I can avoid losing money.

Emotional / social: feel the bank takes it seriously; feel competent and in control.

Design implication: do not only return a generic card-security article. Confirm concern, explain status, show merchant information, offer freeze controls, explain next steps, escalate suspected fraud and record the interaction.

---

20. Customer Journey Mapping​

Worked journey for an unrecognised transaction: notice → investigate → contact → verify → resolve → follow-up. Pain points include unclear merchant names, fragmented information, long waits, repeated authentication and poor status visibility.

Moments that matter: first acknowledgement, status explanation, freeze decision, fraud transfer, next-step confirmation and follow-up communication. Optimise the whole journey, not only the chatbot turn.

---

21. Service Blueprinting​

For “Why has my debit-card payment not gone through?”:

• Frontstage: authenticate, identify payment, explain status, next steps, offer human support
• Backstage: query status, retrieve policy, check restrictions, apply response rules, record interaction, score confidence, escalate if needed
• Supporting systems: mobile banking, identity, payments, CRM, knowledge, foundation model, monitoring, case management
• Failure points: payment data unavailable, outdated knowledge, unsupported model explanation, auth failure, agent unavailability, CRM write failure

---

22. SIPOC — payment-status enquiry​

Suppliers: customer, payment network, core banking, merchant data, knowledge team, identity provider.

Process: receive → authenticate → identify transaction → retrieve status → apply policy → explain → offer next action → escalate → record.

Insight: the foundation model is only one component. Accurate payment status depends primarily on reliable system integration and policy logic.

---

23. Value Stream Mapping​

Activity	Processing	Waiting
Navigate help pages	4 min	0
Wait for agent	0	11 min
Authentication	2 min	0
Identify issue	2 min	0
Search knowledge	1.6 min	0
Check system	2 min	0
Transfer to specialist	1 min	6 min
Resolution	4 min	0
After-call documentation	2 min	0

Total processing 16.6 minutes; waiting 17 minutes; elapsed 33.6 minutes. Future state for eligible requests: authenticated mobile start, automatic identification and status retrieval, approved explanation, specialist transfer only when required, summary handoff—target elapsed 3–5 minutes. Most delay is waiting, handoffs and information retrieval—not the final decision.

---

24. BPMN — lost-card process​

AI-assisted future flow: authenticated entry → lost / stolen / misplaced choice → card confirmation → eligibility check → temporary freeze or freeze-and-replace → escalate to fraud if stolen or suspicious → audit record → confirmation.

Gateways expose where AI recommends, where deterministic rules decide and where humans must intervene: authentication, digital eligibility, suspicious transactions, vulnerability, human approval and action success.

---

25. Process Mining​

Event logs for payment enquiries show only 41% follow the expected route; 24% need knowledge search; 18% transfer; 9% create a second contact; 5% need supervisor support; 3% produce complaints. Monday mornings wait longest; new agents transfer 32% more; one payment type drives disproportionate recontact; policy updates temporarily raise handling time; some chatbot intents increase subsequent calls.

Prioritise: the payment type causing repeat contacts, knowledge retrieval for new agents, routing problems and poor chatbot intents—evidence-based prioritisation, not general automation.

---

26. Five Whys, Fishbone and Problem Tree​

Five Whys root cause for inconsistent pending-payment answers: knowledge management treated as local documentation rather than an enterprise capability—no single accountable owner. Fix ownership, authoritative sources, publication workflows and versioning before scaling AI retrieval.

Fishbone on 68% first-contact resolution shows people, process, technology, data, policy, measurement and environment causes. An AI assistant may address knowledge and search issues; it will not independently fix policy ambiguity, staffing peaks or unclear ownership.

Problem Tree core problem: customers cannot resolve routine needs quickly and consistently. Consequences include waits, recontacts, low satisfaction, cost, employee frustration, complaints, inconsistent outcomes and regulatory exposure. The portfolio must combine AI with process redesign, knowledge governance, integration, training and measurement change.

---

27. Opportunity Solution Tree​

Desired outcome: raise successful safe resolution of routine requests from 68% to 82%.

Opportunity	Example solutions	Experiments
Payment status unclear	Plain-language explanation, proactive notifications, timeline, conversational guidance	Comprehension tests; notification vs chatbot
Agents cannot find policy	Search, RAG, knowledge restructuring, context-sensitive guidance	Accuracy and retrieval-time comparison
Customers repeat after transfer	Shared summary, cross-channel context, structured handoff	Pilot with one specialist team
Sensitive cases identified late	Vulnerability detection, fraud indicators, escalation rules	Retrospective labelled evaluation; FPR/FNR review

---

28. Assumption Mapping and Hypothesis Tree​

Test assumptions in this order: knowledge quality → safety and answer accuracy → API availability → human escalation → customer acceptance → financial benefit.

Test the assumptions that could invalidate the solution before optimising the experience.

Main hypothesis: a hybrid AI service model can improve customer resolution while controlling banking risk—supported by sub-hypotheses on accuracy, adoption, productivity, risk appetite and benefits exceeding costs, each with concrete tests (golden datasets, pilots, threat models, TCO, benefits dependency).

---

29. Current-state and future-state mapping​

Current: static help, generic chatbot links, late authentication, multi-repository search, manual notes, incomplete specialist context, small-sample QA, multi-channel policy distribution.

Future: authenticated start, intent and account context, approved knowledge responses, immediate routine resolution, human specialists for sensitive cases, automatic context transfer, continuous quality monitoring, governed knowledge propagation.

Transformation requirements: unified knowledge ownership, API integration, evaluation platform, shared context, agent workflow redesign, escalation controls, monitoring, training and updated performance measures.

---

Part IV: Discovery synthesis

Original request​

Build a generative AI chatbot to reduce customer-service cost.

Refined opportunity​

Create a trusted hybrid AI service that helps customers and employees resolve routine banking requests quickly, while preserving human control for sensitive, regulated and high-risk situations.

Recommended first release​

Customer-facing: authenticated payment-status, card-status and basic fee/account-information questions; clear human escalation; no autonomous regulated decisions.

Employee-facing: knowledge retrieval, suggested responses, conversation summarisation, routing support and policy citations.

Enabling work: knowledge consolidation, AI evaluation, human-escalation integration, security testing, data and system integration, AI governance, employee training and benefits baseline.

Success measures: successful safe resolution, first-contact resolution, average handling time, repeat contact, customer satisfaction, human-escalation accuracy, grounded answer rate, policy-compliance rate, cost per successful resolution, employee adoption and complaints.

---

Part V: End-to-end strategy and discovery sequence

Use the frameworks as a sequence, not unrelated workshops:

1. AI North Star
2. Strategy Choice Cascade
3. Three Horizons
4. PESTLE and Porter’s Five Forces
5. SWOT and TOWS
6. Value Chain Analysis
7. Business Model Canvas
8. Operating Model Canvas
9. Capability-Based Planning
10. Wardley Mapping
11. Value-Driver Tree
12. Scenario Planning
13. Stakeholder Mapping
14. Voice of the Customer
15. Design Thinking and Double Diamond
16. Jobs to Be Done
17. Customer Journey Mapping
18. Service Blueprinting
19. SIPOC
20. Value Stream Mapping
21. BPMN
22. Process Mining
23. Five Whys, Fishbone and Problem Tree
24. Opportunity Solution Tree
25. Assumption Mapping
26. Hypothesis Tree
27. Current-State and Future-State Mapping

---

Final principle

Each framework should produce one of five outcomes: a clearer problem, a better strategic choice, a testable hypothesis, a controlled design decision or evidence for investment, redesign or termination.

The strongest AI solution engineer does not ask:

Which AI model should we deploy?

They ask:

What outcome matters, what is preventing it today, where can AI create defensible value, what must be true for the solution to succeed, and what evidence is required before the organisation invests further?