End-to-End AI Solution Engineering Playbook: Commercial Case, Benefits and Investment for Banking Customer Service
Strategic fit and readiness do not fund a programme. MonGo Bank must still prove what the hybrid AI service will cost, which benefits are cash versus capacity, who owns them and when to continue, expand or stop.
This article is Part III of the Banking Customer-Service AI playbook: commercial case, benefits and investment. It follows Part I: Strategy and Discovery and Part II: Readiness, Maturity and Prioritisation.
1. Purpose of this phase
Prioritisation selected agent knowledge assistance, summarisation, intelligent routing, payment-status self-service, knowledge governance, evaluation and LLMOps. The commercial phase must still answer:
- What will it cost?
- What financial and non-financial benefits may appear—and when?
- Which assumptions drive the case?
- Which risks reduce value?
- Is it affordable?
- Which benefits are cash, capacity or cost avoidance?
- Who owns each benefit?
- When should the bank continue, expand, redesign or stop?
2. Commercial baseline
| Measure | Current |
|---|---|
| Annual service interactions | 3,100,000 |
| Agent-handled interactions | 2,200,000 |
| Digital self-service interactions | 900,000 |
| Average handling time | 8.4 minutes |
| Fully loaded agent cost | $35 / hour |
| Cost per agent interaction | $8.60 |
| First-contact resolution | 68% |
| Repeat contact within 7 days | 12% |
| Transfer rate | 18% |
| Knowledge-search time | 95 seconds |
| After-call work | 2 minutes |
| Annual contact-centre cost | $74 million |
Without a baseline, the bank cannot attribute AHT falls to AI versus demand, training, redesign, seasonality, policy change or staffing. Require historical data, seasonal adjustment, comparable teams, consistent definitions, control groups where possible and agreed calculation rules.
3. Business Case
A business case connects problem, strategy, options, value, costs, risks, delivery, governance and benefits ownership—not ROI alone. A profitable proposal can still fail on regulatory risk, customer harm, readiness, supplier concentration, change, ownership or misalignment.
MonGo structure
- Executive summary — phased hybrid programme: Wave 1 agent assist + evaluation/monitoring; Wave 2 controlled payment/card/fee self-service
- Problem — waits, recontacts, fragmented knowledge, inconsistent answers, manual notes, transfers, weak digital completion
- Strategic alignment — digital service, satisfaction, cost, resilience, productivity, responsible AI
- Options — do nothing; search-only; agent assist; customer chatbot only; phased hybrid; managed outsource
- Preferred — phased hybrid: near-term productivity, controlled customer value, lower initial risk, reusable capabilities, evidence before scale
- Financial case — implementation and operating cost, benefit timing and uncertainty, cash vs avoidance vs capacity, risk reduction
- Delivery — staged funding through foundation → employee pilot → customer pilot → controlled actions → scale
- Governance — evidence gates before each funding release
4. Five Case Model
4.1 Strategic Case — why intervene?
Demand is rising; customers expect faster digital resolution; knowledge is fragmented; costs remain high; competitors improve digital service; employees lose time to repetitive work. Without intervention: ~4% annual demand growth, longer waits, overtime and recruitment pressure, falling satisfaction, competitive attrition and channel inconsistency.
Objective: improve successful and safe resolution of routine needs while preserving human support for sensitive and complex situations.
Intervention is necessary—but the strategic case does not yet prove generative AI is the best intervention. That belongs to the economic case.
4.2 Economic Case — which option creates best value?
| Option | Cost | 5-yr gross benefit | Net (risk-adjusted where noted) |
|---|---|---|---|
| 0 Do nothing | — | — | −$18M (rising demand/cost) |
| 1 Process improvement only | $4M | $11M | $7M |
| 2 Rules-based chatbot | $6M | $15M | $9M |
| 3 Agent-assist AI only | $12M | $28M | $16M |
| 4 Customer-facing AI only | $14M | $32M | $18M / $10M risk-adjusted |
| 5 Phased hybrid AI | $18M | $47M | $29M / $24M risk-adjusted |
Preferred: Option 5 — best balance of value, feasibility, risk, alignment and flexibility.
4.3 Commercial Case — can it be acquired acceptably?
Build/own: banking knowledge structure, evaluation datasets, escalation rules, customer-protection logic, process integration, AI governance, benefits measurement.
Buy: foundation models, cloud, managed vector search, generic monitoring, contact-centre software.
Partner: initial platform build, independent validation, red teaming, accessibility testing, RAI assurance.
Commercial risks: lock-in, token pricing, model retirement, residency, weak audit rights, supplier concentration, service credits, subprocessors, incident notification, IP terms.
Contract protections: SLAs; data-processing and residency; security; incident notification; audit; model-change notice; subprocessor transparency; usage/cost reporting; exit and portability; prompt/output ownership; ban on training on bank data.
4.4 Financial Case — is it affordable?
| Year | Implementation | Operating | Total |
|---|---|---|---|
| 0 | $5.5M | $0.5M | $6.0M |
| 1 | $3.0M | $2.0M | $5.0M |
| 2 | $1.5M | $2.5M | $4.0M |
| 3 | $0.5M | $2.7M | $3.2M |
| 4 | $0.3M | $2.9M | $3.2M |
| Five-year total | $21.4M |
Staged funding: foundation (knowledge, evaluation, security, agent pilot) → pilot (thresholds, security, acceptance, initial value) → scale (measured benefits, stability, risk appetite, customer outcomes, updated assumptions).
4.5 Management Case — can it be delivered and governed?
Requires programme governance, product ownership, architecture authority, risk/compliance oversight, benefits and change management, supplier management, roadmap, quality gates and operational transition.
Manage as business transformation, product development, knowledge improvement, AI engineering, risk/governance change and operational service change—not a standalone chatbot project.
5. Total Cost of Ownership
Credible TCO includes more than licences and API calls.
| Cost category | Five-year cost |
|---|---|
| Discovery and design | $1.2M |
| Data and knowledge | $2.3M |
| Platform and infrastructure | $4.4M |
| Model consumption | $2.6M |
| Integration | $3.2M |
| Engineering and delivery | $3.8M |
| Governance and assurance | $1.4M |
| Change and adoption | $1.0M |
| Operations and support | $1.2M |
| Contingency (10%) | $2.0M |
| Total | $23.1M |
6. Return on Investment
ROI = (Total benefit − Total cost) ÷ Total cost × 100
$47M − $23.1M = $23.9M net → 103.5% ROI (~$1.04 net per $1 invested).
Does not mean value is guaranteed, all benefits are cash, benefits are immediate, risk is fully priced or organisational action is optional.
7. Benefit classification
| Type | Meaning | Illustrative 5-yr |
|---|---|---|
| Cash-releasing | Reduces expenditure (overtime, contractors, outsourcing, licences, recruitment) | $9M |
| Cost avoidance | Prevents future spend (hiring growth, infra, complaints, rework) | $14M |
| Capacity release | Frees time without automatic cost cut | $15M |
| Non-financial | Satisfaction, consistency, accessibility, traceability, resilience (monetised equivalent) | $9M |
Capacity becomes valuable only when converted into more demand handled, less overtime, avoided hiring, better service, reallocation or backlog reduction.
8. Net Present Value
Discount rate: 8%.
| Year | Benefits | Costs | Net | Discount factor | PV |
|---|---|---|---|---|---|
| 0 | $0.0M | $6.0M | −$6.0M | 1.000 | −$6.00M |
| 1 | $4.0M | $5.0M | −$1.0M | 0.926 | −$0.93M |
| 2 | $9.0M | $4.0M | $5.0M | 0.857 | $4.29M |
| 3 | $13.0M | $3.2M | $9.8M | 0.794 | $7.78M |
| 4 | $15.0M | $3.2M | $11.8M | 0.735 | $8.67M |
| 5 | $16.0M | $3.5M | $12.5M | 0.681 | $8.51M |
NPV ≈ $22.32M. More reliable than simple ROI for multi-year programmes.
9. Internal Rate of Return
Using the same cash flows, illustrative IRR ≈ 71%—attractive vs an 8% hurdle, but misleading alone when cash flows are uncertain, benefits are non-cash, project sizes differ or value depends on organisational action. Use IRR with NPV.
10. Payback Period
| Year | Net | Cumulative |
|---|---|---|
| 0 | −$6.0M | −$6.0M |
| 1 | −$1.0M | −$7.0M |
| 2 | $5.0M | −$2.0M |
| 3 | $9.8M | $7.8M |
Unrecovered at end of Year 2: $2.0M ÷ $9.8M ≈ 0.204 → payback ≈ 2.2 years.
11. Break-Even Analysis
Annual operating cost after scale: $3.2M. Net benefit per successful automated interaction: $4.50.
Required successful interactions: $3.2M ÷ $4.50 ≈ 711,111 (~712,000).
Against 1.4M eligible interactions → 50.9% successful resolution rate needed to cover annual operating economics.
12. Cost-Benefit Analysis
Beyond money: waiting, FCR, admin load, policy consistency, availability, audit evidence and digital adoption—versus technology, remediation, integration, evaluation, security, change, oversight, operations and supplier cost—plus negative impacts (confusion, exclusion, wrong answers, employee anxiety, cyber threats, supplier dependency, cloud use, reputation).
Proceed because the hybrid design limits high-impact automation, preserves escalation, requires evaluation, prioritises lower-risk cases and builds reusable capabilities. High ROI does not justify uncontrolled customer harm.
13. Unit Economics
Variable cost per AI customer interaction
| Component | Cost |
|---|---|
| Model I/O | $0.07 |
| Embeddings and retrieval | $0.01 |
| Cloud processing | $0.03 |
| Monitoring and logging | $0.02 |
| Security services | $0.01 |
| Variable total | $0.14 |
Allocated fixed: $2.1M ÷ 2M interactions = $1.05 → total $1.19 per interaction.
At 70% successful resolution: $1.70 per success vs $8.60 agent cost → potential unit saving $6.90—only if accurate, no recontact, no harm, escalation cost not hidden and quality met.
14. Cost per Successful Safe Resolution
A successful safe resolution must resolve the eligible need, use accurate information, avoid inappropriate action, comply with policy, avoid 7-day recontact and provide human access where required.
Example month: 100,000 conversations → 62,000 safe successes; $130,000 operating cost → $2.10 per successful safe resolution. Stronger than cost per conversation.
15. Sensitivity Analysis
Base: 1.4M eligible; 60% adoption; 70% safe resolution; $6.90 net benefit; $3.2M annual platform; $9M implementation.
| Adoption | Annual gross benefit |
|---|---|
| 30% | $2.03M |
| 45% | $3.04M |
| 60% | $4.06M |
| 75% | $5.07M |
| Resolution rate | Annual gross benefit |
|---|---|
| 45% | $2.61M |
| 55% | $3.19M |
| 70% | $4.06M |
| 80% | $4.64M |
| Benefit / resolution | Annual gross benefit |
|---|---|
| $4.00 | $2.35M |
| $5.50 | $3.23M |
| $6.90 | $4.06M |
| $8.00 | $4.70M |
Most sensitive to: (1) successful safe resolution, (2) adoption, (3) converting employee time to value, (4) integration cost. Token cost matters but is not the largest driver.
16. Scenario Analysis
| Scenario | Assumptions | Benefits | Costs | Net | Decision |
|---|---|---|---|---|---|
| Conservative | 35% adoption; 50% safe; 50% conversion; +25% integration; +20% opex | $26M | $28M | −$2M | Do not scale without redesign |
| Base | 60% / 70% / 70%; costs near plan | $47M | $23.1M | $23.9M | Proceed staged |
| Optimistic | 75% / 80%; strong uptake; lower model cost; reuse | $66M | $24M | $42M | Accelerate controlled scale |
17. Monte Carlo Simulation
Variables: adoption, resolution, benefit per interaction, implementation cost, model cost, integration delay, employee uptake, recontact reduction.
After 10,000 simulations (illustrative): 78% probability of positive NPV; median NPV $18M; 10th percentile −$4M; 90th percentile $41M; 64% probability of payback within three years.
Implication: attractive but not risk-free—use stage gates, limit initial commitment, measure pilots, preserve stop rights, update the model each phase.
18. Risk-Adjusted NPV
| Benefit | Undiscounted | Probability | Risk-adjusted |
|---|---|---|---|
| Handling-time reduction | $12M | 80% | $9.6M |
| Avoided recruitment | $10M | 65% | $6.5M |
| Reduced repeat contacts | $8M | 70% | $5.6M |
| Improved digital resolution | $9M | 60% | $5.4M |
| Risk and quality improvement | $8M | 55% | $4.4M |
| Total | $47M | $31.5M |
Risk-adjusted net before discounting: $31.5M − $23.1M = $8.4M. Still positive; much thinner than the headline case—supports cautious staging.
19. Real Options Analysis
| Option | Investment | Condition |
|---|---|---|
| 1 Pilot | $3M | Test agent assist and evaluation |
| 2 Expand | $5M | Productivity and quality proven |
| 3 Customer self-service | $4M | Knowledge, security, escalation thresholds met |
| 4 Controlled actions | $3M | Self-service safe and stable |
| 5 Abandon | — | Safe-resolution, adoption, cost, regulation or benefit conversion fails |
Staged options beat a single $15M upfront commitment: buy evidence before exercising the next investment.
20. Benefits Dependency Network
Example — reduce AHT by 1.2 minutes:
- Objective: improve productivity
- Business changes: agents use knowledge assist; review summaries; managers use quality dashboards; knowledge owners maintain content
- Enablers: training, workflow redesign, procedures, performance measures, ownership, coaching
- Technology: RAG, citation, summarisation, CRM integration, evaluation, monitoring
Logic: assistant → faster retrieval → less search → lower AHT → capacity → overtime/hiring reduction. Without adoption, current content, workflow change and capacity reallocation, financial benefit does not materialise.
21. Benefits Map
Agent knowledge: faster approved information → lower search time and consistency → faster reliable service → higher FCR / lower effort → trusted efficient digital service.
Summarisation: structured notes → less after-call work and better handoff → less repetition → lower handling cost / better records → service quality and compliance evidence.
22. Value-Driver Tree
Top: increase customer-service value via unit cost, capacity, customer value and risk reduction.
| Branch | Calculation | Annual value |
|---|---|---|
| Handling time | 2.2M × 1.2 min ÷ 60 × $35 | $1.54M |
| After-call work | 1.5M × 45 sec ÷ 60 × $35 | $656K |
| Repeat contact | 372K × 20% × $8.60 | $640K |
23. Benefits Realisation Plan
| Benefit | Baseline | Target | Owner | Timing |
|---|---|---|---|---|
| Knowledge-search time | 95 sec | 35 sec | Head of Service Operations | 6 months |
| Average handling time | 8.4 min | 7.2 min | Contact Centre Director | 12 months |
| First-contact resolution | 68% | 78% | CS Product Owner | 18 months |
| Repeat contact | 12% | 8% | Journey Owner | 18 months |
| After-call work | 2.0 min | 1.1 min | Operations Director | 9 months |
| Employee satisfaction | 62% | 75% | HR and Service Leadership | 12 months |
| Grounded answer rate | Not measured | 95%+ | AI Product Owner | Before production |
| Policy-related errors | 7% | 2% | Knowledge Management Lead | 12 months |
Benefit owners agree baselines and methods, drive business change, review performance and confirm realised value. AI engineering cannot own all business benefits.
24. Benefits Register (example)
BEN-04 — reduce repeat contacts (cost avoidance + CX). Baseline 12% → target 8%; ~$640K annually; owner Head of Customer Journeys. Dependencies: accurate answers, context, proactive follow-up, escalation, measurement. Risk: distrust drives recontact. Mitigation: sources, confirmation, follow-up summaries, reason coding.
25. Benefit Leakage Analysis
Theoretical AHT capacity: $1.54M. With 80% adoption × 85% availability × 75% content coverage × 70% capacity conversion ≈ $550K realised. Theoretical-to-realised leakage is material.
26. Benefits Tracking Dashboard
- Technical: groundedness, retrieval accuracy, latency, availability, tool failure, model cost
- Operational: AHT, after-call work, transfers, recontact, escalation
- Customer: satisfaction, effort, safe resolution, human-access rate, complaints
- Employee: adoption, override, trust, time saved, training completion
- Financial: cost per successful resolution, cash, avoidance, capacity, realisation, TCO variance
27. VALUE Gate
Before each major investment:
- Valuable — material problem; measurable value; named owner
- Actionable — clear next actions; business changes; thresholds
- Logical — evidenced assumptions; traceable maths; risks included
- Understandable — exec-readable logic; cash vs capacity; explicit uncertainty
- Executable — deliverable changes; affordable funding; funded dependencies; operationally realisable benefits
28. Investment stage gates
| Gate | Approve when | Investment |
|---|---|---|
| 1 Foundation | Strategic case clear; use cases prioritised; gaps known; foundation TCO OK; baselines exist | $3M |
| 2 Employee pilot | Knowledge enough; golden datasets; security defined; users trained; benefits measurable | $3M |
| 3 Customer pilot | Agent value shown; groundedness threshold; handoff tested; DPIA approved; operating model ready | $4M |
| 4 Controlled actions | Customer pilot stable; restricted tools; reversible; audit logging; incident response tested | $3M |
| 5 Scale | Risk-adjusted NPV positive; benefits realised; CX improved; costs controlled; risk in appetite | — |
29. Stop criteria
Pause or stop expansion if:
- Successful safe resolution stays below 55%
- Complaints rise materially
- Human escalation fails
- Knowledge quality stays below threshold
- Unit cost exceeds agent-service cost
- Adoption stays below 30%
- Integration cost exceeds plan by 40%
- Risk-adjusted NPV turns negative
- Critical security weaknesses remain
- Benefit owners cannot convert capacity into value
30. Final commercial conclusion
| Measure | Result |
|---|---|
| Total benefit | $47.0M |
| Total cost (TCO) | $23.1M |
| Net benefit | $23.9M |
| ROI | 103.5% |
| NPV | $22.3M |
| Payback | 2.2 years |
| P(positive NPV) | 78% |
Value depends on adoption, knowledge quality, safe resolution, integration, operational change, benefit ownership and cost control—not model deployment alone.
31. Phase outputs
Business Case; Five Case Model; option appraisal; five-year TCO; ROI; NPV and IRR; payback; break-even; unit economics; sensitivity and scenario analysis; risk-adjusted value; real-options model; Benefits Dependency Network; Value-Driver Tree; Benefits Realisation Plan and Register; named owners; investment gates; stop criteria.
32. Core principle
Not: “How much money could AI theoretically save?”
Rather:
Which measurable outcomes will change, what organisational and technical conditions are required, who owns the change, how uncertain are the assumptions, and will the risk-adjusted value justify the full lifecycle cost?
Next phase: Architecture, Operating Model and Engineering Design—TOM, AI operating model, build–buy–partner, TOGAF, ArchiMate, C4, DDD, Event Storming, API-first and event-driven architecture, CAF, Well-Architected, Zero Trust, Data Mesh/Fabric, lakehouse/medallion, platform engineering, MLOps, LLMOps, AgentOps, DevSecOps, SRE and FinOps.
Discussion
Comments
Share feedback or questions about this page. No account required.
Loading comments…