Business Fundamentals
Executive view
Technical view
Why this matters
As an AI Solution Engineer operating at big-four manager level, you sit between engineers who optimise tokens and executives who optimise EBITDA. Your job is not to recite accounting textbooks—it is to translate capability into economic language the client already uses in board packs, investor calls, and annual planning.
Three situations expose weak business fundamentals immediately:
-
The “AI chatbot” brief. A retail bank or insurer asks for conversational AI with no link to contact volume, average handle time (AHT), first-contact resolution (FCR), or cost-to-serve. Without fundamentals, you build a demo that wins applause and loses funding at business case.
-
The pilot that never scales. A team proves accuracy on a sandbox dataset but cannot explain marginal cost per transaction, payback period, or which OPEX line absorbs run cost. Finance blocks scale; the engagement stalls in “successful pilot, no production.”
-
The misaligned metric. You improve model latency while the COO is measured on SLA breach rate and the CMO on conversion and NPS. You delivered engineering success and commercial irrelevance.
Strong fundamentals let you: reframe vague asks into measurable problems; choose leading indicators you can move in 90 days while lagging indicators mature; size benefits conservatively; and articulate cost of delay when prioritisation committees defer AI investment. Partners trust engineers who can open a workshop with “here is how this institution earns a pound, and here is where we attach value”—not “here is our RAG stack.”
Learn
Business models: B2B (business-to-business)
Definition. B2B organisations sell products or services to other businesses. Revenue often comes from contracts, licences, professional services, or recurring platform fees. Sales cycles are longer; buyers evaluate ROI, risk, integration, and total cost of ownership (TCO) rather than impulse.
Why it matters. Enterprise AI engagements are almost always B2B internally (IT selling to the business) or B2B externally (vendor to client). Understanding B2B economics explains why proof of value, security review, and procurement dominate timelines—and why benefits must be expressed per seat, per site, or per million transactions.
How to use on engagement. Map the client’s revenue model: Are they selling to other firms (pure B2B) or using B2B infrastructure to reach consumers (see B2B2C)? Identify the economic buyer (often CFO or COO) vs the technical buyer (CIO/CISO). Size AI value per business unit or contract, not “for the whole bank.”
Pitfalls. Treating a B2B buyer like a B2C user (consumer-grade UX expectations without enterprise controls). Ignoring implementation cost in year one. Assuming one champion equals budget approval.
Mini example. A regtech vendor sells AML screening to tier-2 banks at £800K ACV (annual contract value) plus 20% maintenance. An AI false-positive reduction feature that cuts analyst review hours by 15% strengthens renewal and upsell—value is tied to analyst FTE cost and regulatory throughput, not “better ML.”
Business models: B2C (business-to-consumer)
Definition. B2C organisations sell directly to individuals: retail, telco, streaming, insurance direct, neobanks. Revenue drivers include transaction volume, subscription, advertising, and interchange. Competition focuses on acquisition, conversion, retention, and experience.
Why it matters. B2C clients measure AI success in conversion rate, churn, NPS/CSAT, and cost per contact—not enterprise licence counts. Personalisation and generative experiences must connect to funnel metrics or servicing cost.
How to use on engagement. Trace the customer journey: acquisition → onboarding → usage → support → renewal/churn. Place AI interventions where volume × value per step is largest. Watch privacy and consent—B2C amplifies reputational risk.
Pitfalls. Optimising engagement metrics that do not flow to margin (e.g. longer session time without purchase). Underestimating content moderation and brand risk for customer-facing GenAI.
Mini example. A UK retailer with 12M active app users sees 3.2% cart abandonment at payment. A checkout assistant that recovers 0.15% of abandoned carts at £45 AOV yields measurable revenue—size it before sizing GPU cost.
Business models: B2B2C and platforms
Definition. B2B2C (business-to-business-to-consumer) firms enable other businesses to reach end customers—payment networks, marketplaces white-labelling, embedded finance, insurer–broker–customer chains. The platform often earns fees, interchange, or revenue share.
Why it matters. Value attribution is split across parties. AI may benefit the platform (lower ops cost), the partner (faster onboarding), or the end customer (better experience)—who pays and who benefits may differ.
How to use on engagement. Draw a simple value flow: who holds the customer relationship, who carries regulatory liability, who owns data. Align the business case to the entity that holds budget.
Pitfalls. Building for the wrong sponsor in a multi-sided model. Ignoring contractual data restrictions between platform and partners.
Decision table.
| Pattern | Who pays for AI | Typical value metric |
|---|---|---|
| Embedded finance | Platform or partner (often split) | Activation rate, fraud loss, support cost |
| Insurer–broker–customer | Insurer or MGA | Loss ratio, quote turnaround, retention |
| Marketplace | Platform (+ seller optional co-fund) | GMV, take rate, dispute rate |
Business models: marketplace and network effects
Definition. Marketplaces match supply and demand (rides, listings, freelance talent, B2B parts). Revenue is often commission, listing fees, or subscription for sellers. Network effects mean value rises as more participants join—liquidity is strategic.
Why it matters. AI use cases include matching quality, fraud detection, dynamic pricing, and seller/buyer support. Benefits scale with transaction count; unit economics must include both sides of the market.
How to use on engagement. Quantify GMV (gross merchandise value), take rate, and cost of trust/safety. AI that reduces manual review per listing has a clear unit cost story.
Pitfalls. Optimising one side of the market and harming liquidity (e.g. aggressive auto-rejection of listings). Confusing marketplace GMV with revenue.
Mini example. A B2B parts marketplace processes 2.4M listings/year; manual compliance review costs £6 per listing. Semi-automated classification at 70% auto-approve with human QA on edge cases saves ~£8.4M OPEX annually—before counting faster time-to-live for sellers.
Revenue models: subscription
Definition. Customers pay recurring fees (monthly/annual) for access to product or service—SaaS, streaming, membership, insurance premiums structured as ongoing cover.
Why it matters. Subscription businesses live on retention, expansion revenue (upsell), and CAC payback. AI that reduces churn or increases ARPU (average revenue per user) has compounding value.
How to use on engagement. Model MRR/ARR, churn rate, and net revenue retention (NRR). Tie AI to reducing involuntary churn (failed payments), support-driven cancellation, or poor onboarding.
Pitfalls. Counting one-off efficiency as recurring benefit. Ignoring cohort behaviour—aggregate churn hides segment risk.
Revenue models: usage-based and consumption
Definition. Revenue scales with consumption: API calls, tokens, data volume, transactions processed, compute minutes, pay-as-you-go cloud.
Why it matters. AI inference is often usage-based internally and externally. FinOps and business case must align variable cost per unit with variable benefit per unit. Margin erodes if usage grows faster than price or value.
How to use on engagement. Build a unit economics sheet: cost per 1K queries, revenue or savings per 1K queries, gross margin per unit. Stress-test at 2× and 10× volume.
Pitfalls. Quoting pilot token cost as steady-state run cost. Forgetting embedding refresh, re-indexing, and eval runs in consumption forecasts.
Revenue models: outcome-based and performance contracts
Definition. The provider is paid (or bonused) based on results: claims reduced, fraud prevented, SLA attainment, collections recovered, energy saved.
Why it matters. Outcome models force clarity on baselines, measurement windows, and causal attribution—exactly what a credible AI business case needs anyway.
How to use on engagement. Define the outcome metric contractually (e.g. 8% reduction in escalations). Specify who owns process change; AI alone rarely delivers outcomes without workflow redesign.
Pitfalls. Selection bias and external factors (seasonality, macro) attributed to AI. Under-specified baselines that finance disputes at year-end.
Revenue models: professional and managed services
Definition. Revenue from people time: consulting, implementation, BPO, managed service desks, legal and audit. Margin depends on utilisation, rate cards, pyramid mix, and automation leverage.
Why it matters. Many AI engagements target productivity in professional services firms or captive ops centres—benefits are hours saved × loaded rate, minus tool and change cost.
How to use on engagement. Express value as capacity released (handle more volume without hiring) or margin improvement (same revenue, fewer hours). Be explicit about revenue cannibalisation if the firm bills by the hour.
Pitfalls. Assuming 100% of saved hours become cashable benefit—often only 30–50% is redeployed or billed. Ignoring quality and liability when junior work is automated.
Financial concepts: revenue, gross margin, and contribution margin
Definition. Revenue is income from sales of goods/services. Gross margin is revenue minus direct cost of delivering the product (COGS). Contribution margin extends to variable costs attributable to a product line or channel.
Why it matters. AI benefits often appear as COGS reduction (cheaper servicing), SG&A reduction (fewer support staff), or revenue uplift (conversion). Executives want to know which line moves.
How to use on engagement. Ask finance for a P&L waterfall for the target process. Tag each benefit line as revenue, gross margin, or OPEX.
Pitfalls. Confusing bookings with revenue. Claiming revenue benefit from faster internal reports with no customer impact.
Mini example. Contact centre cost sits in OPEX, not COGS—but reducing AHT by 90 seconds on 4M calls/year at £4.20 loaded cost per minute saves ~£25M OPEX annually if adoption holds.
Financial concepts: EBITDA and operating profit
Definition. EBITDA (earnings before interest, taxes, depreciation, and amortisation) is a common proxy for operating cash generation. Operating profit accounts for depreciation and amortisation. Both exclude financing and tax effects.
Why it matters. Board and PE sponsors often frame targets as EBITDA margin % or absolute EBITDA improvement. AI programmes pitched only as “digital transformation” fail; those pitched as +50 bps EBITDA margin in servicing get funded.
How to use on engagement. Translate operational metrics to EBITDA impact using finance-provided conversion factors (do not invent corporate overhead allocations). Show timing: year-one investment vs year-two run-rate benefit.
Pitfalls. Double-counting benefits across workstreams. Ignoring depreciation of new AI infrastructure when the client cares about operating profit, not EBITDA.
CAPEX vs OPEX
Definition. CAPEX (capital expenditure) buys long-lived assets (servers, software capitalised, implementation build). OPEX (operating expenditure) is ongoing run cost (cloud consumption, licences, support, staff).
Why it matters. AI solutions blur the line: model API fees are OPEX; capitalised platform build is CAPEX. Budget owners differ—IT capex committees vs business OPEX holders. Approval paths and hurdle rates differ (often higher for capex).
How to use on engagement. Map funding source for build vs run. Prefer OPEX-friendly phasing for pilots if capex gates are slow—but show TCO over 3–5 years so finance sees no hidden capex shift.
Pitfalls. Treating cloud as always OPEX (long commitments and reserved capacity behave like capex). Forgetting implementation labour classification (capitalise vs expense).
Typical AI spend split (year 1 vs steady state)
Note: Human review often dominates API cost in regulated B2B.
Unit economics: CAC (customer acquisition cost)
Definition. CAC is total sales and marketing spend divided by new customers acquired in a period.
Why it matters. In B2C and SaaS B2B, AI that improves conversion or targeting affects CAC efficiency. Size: incremental customers × contribution margin vs incremental AI cost.
How to use on engagement. Only claim CAC impact if you touch acquisition funnel data and can A/B test. Otherwise focus on retention or servicing.
Pitfalls. Allocating all marketing spend to one digital assistant. Ignoring brand risk from poorly governed GenAI in acquisition.
Unit economics: LTV (lifetime value)
Definition. LTV is the present value of gross profit expected from a customer over the relationship—often modelled as ARPU × gross margin % / churn rate (simplified).
Why it matters. Retention AI (proactive outreach, better servicing) increases LTV. Express benefit as reduced churn % on a cohort with known ARPU.
How to use on engagement. Work with marketing/finance on cohort LTV, not blended averages. Model 12–24 month payback on retention interventions.
Pitfalls. Using industry benchmark LTV without client-specific churn. Confusing LTV with revenue (margin matters).
Unit economics: CAC/LTV ratio and payback period
Definition. Healthy subscription businesses often target LTV:CAC ≥ 3:1 and CAC payback under 12–18 months (varies by sector). Payback period for an AI investment is time until cumulative benefits exceed cumulative costs.
Why it matters. CFOs compare AI against other investments with explicit payback. If payback exceeds the planning horizon, you need phased value or risk-sharing.
How to use on engagement. Produce a simple payback chart: months on X-axis, cumulative £ on Y-axis, mark breakeven. Include sensitivity (adoption 50% vs 80%).
Pitfalls. Using gross benefits without adoption ramp. Omitting run-rate cost after go-live.
Decision table: payback acceptability (illustrative—confirm with client finance)
| Payback (months) | Typical reaction | Your response |
|---|---|---|
| < 12 | Strong yes if benefits credible | Stress-test assumptions |
| 12–24 | Normal for enterprise AI | Show phased milestones |
| 24–36 | Scrutiny; needs strategic alignment | Link to cost of delay / risk |
| > 36 | Usually deprioritised | Rescope MVP or outcome-based pilot |
ROI (return on investment)
Definition. ROI = (Net benefits − Cost) / Cost, expressed as a percentage over a defined period (often 3–5 years).
Why it matters. Steering committees rank initiatives by ROI and strategic fit. AI projects compete with core IT and M&A—not just other AI pilots.
How to use on engagement. Use conservative benefits (P50, not best case). Separate one-off vs recurring costs. Document assumptions in a benefits register.
Pitfalls. ROI on a pilot budget that excludes scale cost. Cherry-picked numerators (labour savings only, ignoring quality rework).
NPV (net present value) and discount rate
Definition. NPV discounts future cash flows to today using a discount rate (WACC or hurdle rate). Positive NPV means the investment beats the hurdle.
Why it matters. Large capex AI platforms (custom LLM platform, enterprise vector search) need NPV over 5 years—especially if benefits ramp slowly.
How to use on engagement. Ask treasury or finance for the approved discount rate. Model benefits ramp (e.g. 40% year 1, 80% year 2, 100% year 3).
Pitfalls. Using 0% discount “because AI is strategic.” Inconsistent treatment of terminal value or salvage value of assets.
IRR (internal rate of return)
Definition. IRR is the discount rate at which NPV equals zero. Compared against hurdle rate for go/no-go.
Why it matters. Some investment committees prefer IRR to ROI for multi-year flows. Useful when benefits are back-loaded (platform build year 1, scale benefits years 2–5).
How to use on engagement. Pair IRR with NPV—IRR alone can mis-rank projects with different scales. Show benefit sensitivity that collapses IRR below hurdle.
Pitfalls. Multiple IRRs with non-standard cash flows—use NPV as primary check.
TCO (total cost of ownership)
Definition. TCO includes build, run, change, and retire costs over the ownership period: licences, cloud, data engineering, MLOps, security, human review, training, incident response, and decommission.
Why it matters. AI pilots look cheap; production TCO often 3–10× pilot spend. Understated TCO destroys trust at scale gates.
How to use on engagement. Use a TCO line-item template (see topic 07). Explicitly include eval, red-team, content refresh, and on-call.
Pitfalls. Counting only API list price. Ignoring integration debt with core systems.
Cost of delay
Definition. Cost of delay quantifies the economic impact of not delivering capability now—lost revenue, excess OPEX, regulatory fine exposure, or competitive share loss per month/quarter deferred.
Why it matters. When committees say “ revisit next year,” cost of delay reframes deferral as a choice to absorb ongoing loss. It prioritises AI alongside non-AI investments.
How to use on engagement. Formula sketch: (monthly benefit run-rate) × (months delayed) plus optional risk cost (e.g. complaint volume trending up). Use ranges, not false precision.
Pitfalls. Double-counting benefits already in base budget. Using cost of delay without credible monthly run-rate.
Mini example. Insurer manual FNOL (first notice of loss) processing costs £1.2M/month in ops FTE and delay penalties. Deferring a document-AI MVP by 6 months costs ~£7M+ in addressable OPEX—not zero.
Performance: KPIs (key performance indicators)
Definition. KPIs are quantifiable measures tracked against targets to judge business health—often tied to balanced scorecard perspectives (financial, customer, process, learning).
Why it matters. Every AI story ends with “which KPI moves, by how much, by when.” If you cannot name KPIs, you are not ready for discovery exit.
How to use on engagement. Request the official KPI dictionary (definitions, owners, reporting cadence). Align solution metrics to one primary KPI and two supporting metrics.
Pitfalls. Inventing KPIs the client does not report. Optimising proxy metrics (model accuracy) that weakly correlate to business KPIs.
Performance: OKRs (objectives and key results)
Definition. OKRs pair qualitative objectives with measurable key results (often quarterly). They cascade from corporate to team level.
Why it matters. AI initiatives funded from transformation or product OKRs must cite the parent OKR explicitly—“KR2: reduce cost-to-serve by 12%” not “enable AI.”
How to use on engagement. Map the use case to O and KR with baseline and target. Define how AI contribution is isolated in quarterly reviews.
Pitfalls. Writing OKRs that are outputs (“launch chatbot”) not outcomes (“reduce email volume 20%”). Misalignment between corporate OKRs and BU OKRs.
Performance: SLAs (service level agreements)
Definition. SLAs are contractual or internal commitments on service performance: response time, resolution time, availability, accuracy thresholds—with penalties or reputational consequences for breach.
Why it matters. AI in servicing often targets SLA attainment (e.g. 80% of chats answered in 60 seconds). Benefits may be penalty avoidance or capacity to meet growth without SLA slip.
How to use on engagement. Baseline current SLA performance and breach cost. Model AI impact on P95 latency and queue depth, not average only.
Pitfalls. Improving average response while P99 breaches persist. Ignoring handoff SLAs between bot and human.
Performance: CSAT and NPS
Definition. CSAT (customer satisfaction) measures satisfaction with a specific interaction or product, often 1–5 scale. NPS (Net Promoter Score) measures likelihood to recommend, segmented promoters minus detractors.
Why it matters. Experience-led strategies fund AI that improves CSAT/NPS if linked to retention or revenue. Pure CSAT uplift without financial linkage is weak in CFO forums.
How to use on engagement. Segment CSAT/NPS by journey (claims, onboarding). Tie +5 NPS points in high-churn cohort to retention £ using finance models.
Pitfalls. Survey bias after AI-only journeys. Gaming scores by deflecting hard cases to opaque channels.
Performance: retention, conversion, and funnel metrics
Definition. Retention is continued customer relationship (renewal, repeat purchase). Conversion is progression between funnel stages (visit → quote → bind → pay).
Why it matters. B2C and SaaS AI investments often target conversion lift (0.1% can be material at scale) or churn reduction.
How to use on engagement. Size: volume × baseline rate × lift × value per conversion. Run holdout tests where possible.
Pitfalls. Attribution to AI when pricing or campaign changed concurrently. Ignoring ** cannibalisation** between channels.
Performance: leading vs lagging indicators
Definition. Leading indicators predict future performance (training completion, deflection rate, draft acceptance rate, time-to-first-response). Lagging indicators confirm outcomes after the fact (churn, EBITDA, loss ratio, NPS).
Why it matters. AI programmes need leading metrics for 90-day steering and lagging metrics for benefits realisation. Relying only on lagging metrics means flying blind for quarters.
How to use on engagement. Define a measurement ladder: leading (weekly) → operational (monthly) → financial (quarterly). Set explicit thresholds for pivot or scale.
Pitfalls. Declaring success on leading indicators when lagging ones flatline (high deflection, rising complaints). Goodhart’s law—when a measure becomes a target, it ceases to be a good measure.
Frameworks and methods
| Framework | Use when | Avoid when |
|---|---|---|
| Unit economics canvas | Sizing per-transaction AI in payments, servicing, listings | Benefits are purely strategic brand play with no volume |
| Benefits register (cashable / non-cashable) | Business case and year-end benefits tracking | You have no finance partner to validate classifications |
| P&L bridge | Explaining how AHT → OPEX → EBITDA | Line items are politically sensitive—get finance to co-sign |
| Cost of delay estimate | Prioritisation vs deferral | Baseline monthly loss is unknown—do discovery first |
| KPI tree | Linking model metrics to executive scorecard | Client has no stable KPI definitions (fix governance first) |
| Sensitivity / tornado chart | CFO review prep | Used without naming which assumptions drive 80% of value |
Method: value hypothesis one-pager (engagement standard)
- Problem in business language (not technology).
- Primary KPI + baseline + target + measurement owner.
- Benefit mechanism (cost out, revenue up, risk down).
- Order-of-magnitude annual value (range).
- Payback breakeven month (range).
- Key assumptions and kill criteria.
Architecture and operating model notes
Business fundamentals shape where AI sits in the operating model, not only the financial case.
Chargeback and cost allocation. In large enterprises, AI run cost must map to a cost centre or internal SKU. If the business unit does not see AI on its P&L, adoption stalls. Define whether AI is central platform OPEX with allocation keys (headcount, transaction volume) or BU-funded.
Benefits ownership. Name a benefits owner (often ops or finance BP) who signs baseline and realisation—not only the product owner.
Real-world scenarios
Scenario A: UK retail bank — servicing and contact deflection
Context. A tier-1 UK bank receives 38M serviced contacts/year across mobile, web, and phone. Cost-to-serve averages £3.80 per contact; 42% are “policy and balance lookup” suitable for authenticated self-service. AHT for phone is 11m 20s; digital chat 8m 45s. Steering committee metric: reduce servicing OPEX by £40M over 3 years without CSAT drop > 2 points.
Reframe. “AI chatbot” becomes authenticated knowledge and transaction assistant with measurable targets: 18% contact deflection in target intents, 90s AHT reduction on remaining assisted contacts via agent copilot, FCR +4 pts on complaints.
Sizing (illustrative).
| Lever | Volume | Unit saving | Annual value |
|---|---|---|---|
| Deflection | 38M × 42% × 18% = 2.87M contacts | £3.80/contact | ~£10.9M |
| Agent copilot | 25M assisted × 90s × £0.07/min | loaded agent cost | ~£26.3M potential; adopt 50% → ~£13.1M |
| Total run-rate (range) | £18–24M before TCO |
TCO guardrail. Year-2 steady-state TCO (API, platform, 25 FTE knowledge ops/QA, change) estimated £9–12M → payback < 18 months if adoption holds.
Leading indicators. Weekly: deflection rate by intent, copilot suggestion acceptance %, escalation rate, P95 response latency. Monthly: CSAT by channel, complaint uphold rate.
Outcome. Programme funded as OPEX transformation tied to servicing KPI tree—not “GenAI innovation fund.” Kill criterion: if month-6 deflection < 8% with stable CSAT, rescope before scale spend.
Scenario B: European P&C insurer — claims FNOL and loss adjustment expense
Context. Mid-size insurer, £2.1B GWP. Loss ratio 64.2%; combined ratio 98.5%. Strategy pillar: reduce loss adjustment expense (LAE) and speed FNOL. 1.1M FNOL/year; 35% require manual document chase; average cycle time 4.2 days; LAE £118 per claim.
Reframe. Document intake, extraction, and triage assistant for FNOL—human-in-the-loop for coverage disputes—not “GPT for claims.”
Sizing (illustrative).
| Lever | Assumption | Annual value |
|---|---|---|
| Manual chase reduction | 35% → 15% of 1.1M claims × £45 manual cost/claim | ~£7.7M |
| Cycle time | 0.8 day faster on 40% eligible claims → fraud/supplier savings + customer retention | £2–4M (range; finance validation) |
| LAE per claim | £118 → target £105 on automated path | Supports combined ratio narrative |
Cost of delay. Each quarter deferred at current LAE burn ≈ £32M × 0.35 manual share × partial addressability → £2–3M/quarter addressable OPEX stuck in manual processing.
Leading indicators. Auto-classification accuracy, STP (straight-through processing) rate, rework rate, auditor override rate. Lagging: LAE/claim, loss ratio (12-month), retention on claims NPS cohort.
Outcome. Business case approved with outcome tracking on LAE/claim; phased rollout by peril type (motor first) to contain model risk.
Scenario C (stretch): UK grocery retail — demand forecasting and waste
Context. National grocer, thin gross margin (~3.5%). Fresh waste 4.1% of category sales; spoilage cost £180M/year. Promo-driven conversion spikes cause stockouts on 8% of SKUs during campaigns.
AI angle. Demand forecasting and markdown optimisation—not GenAI chat. Value tied to margin % and waste bps, not token cost.
Sizing sketch. 10% waste reduction on addressable £120M fresh waste → £12M gross margin benefit; payback depends on data platform sunk cost.
Lesson. Business fundamentals dictate forecasting ML may beat customer chatbot for P&L impact—follow the money, not the trend.
Practice exercises
Primary exercise: value hypothesis for a public financial services firm
Task. Pick one listed bank or insurer (annual report + investor presentation). Produce:
- Business model summary (200 words): revenue mix, B2B/B2C, subscription/transaction elements.
- Three KPIs an AI assistant could influence—with baseline from public sources and proposed target direction.
- One-line payback hypothesis: “If we move [KPI] by [X], order-of-magnitude annual value is [£ range] because [mechanism].”
- Leading vs lagging ladder (table, minimum 3 rows each).
Artefact criteria (acceptance).
| Criterion | Pass |
|---|---|
| Names specific P&L line (OPEX/ revenue/ LAE/etc.) | Yes |
| Uses client or sector numbers, not only generics | Yes |
| Payback stated as range with explicit assumption | Yes |
| No vendor/product named in value story | Yes |
| Finance-style caveat on assumptions included | Yes |
Stretch exercise: cost of delay and sensitivity
Task. Using the same firm, build a 12-month cash flow sketch (spreadsheet): costs (build, run, change) vs benefits (ramped). Compute ROI year 3, breakeven month, and cost of delay if start slips 6 months. Add tornado sensitivity on adoption rate, unit benefit, and run-rate TCO.
Artefact criteria. Breakeven identified; two assumptions flagged as value-critical; one paragraph CFO-ready summary without jargon.
Questions you should be able to answer
- How does this organisation make money—and which segments matter most this planning cycle?
- Which P&L line (revenue, COGS, OPEX, LAE, etc.) will the AI solution affect?
- What is the primary KPI executives are judged on for this problem space?
- What is the baseline and target for that KPI, and who owns measurement?
- What is order-of-magnitude annual benefit (conservative range)?
- What is year-1 vs steady-state TCO, including human review and knowledge ops?
- When does cumulative benefit exceed cumulative cost (payback month)?
- What is cost of delay per quarter if the programme slips?
- Which benefits are cashable vs non-cashable per finance definitions?
- What leading indicators will we review weekly in the first 90 days?
- What lagging indicators confirm benefits at 6–12 months?
- Is funding CAPEX or OPEX, and who holds the budget?
- What adoption rate is required for the case to hold—and is that credible?
- What unit economics apply (cost and benefit per transaction/contact/claim)?
- What would cause you to stop the initiative (kill criteria)?
Negative cases
| Failure mode | What goes wrong | Early warning signal |
|---|---|---|
| Vendor-first business case | ROI built on licence discount, not client KPIs | Deck leads with product screenshots |
| Pilot economics only | Scale TCO 5× higher; programme halted at gate 2 | No line item for eval, QA FTE, re-index |
| Proxy metric success | F1 score up; AHT flat | Model metrics green, ops KPIs amber |
| Wrong P&L line | “Revenue uplift” claimed from internal report automation | No customer-facing metric movement |
| Ignored adoption | 100% benefit at 30% usage | No change plan; agents bypass copilot |
| Double-counted savings | Same FTE reduction in three workstreams | Finance finds one pot of heads |
| Cost of delay abuse | Fantasy monthly loss; credibility lost | No baseline; round numbers only |
| B2C privacy backlash | NPS drops; regulatory complaint | GenAI in regulated advice without HITL |
| CAPEX/OPEX mismatch | Built but no run budget | Ops refuses chargeback |
| Lag-only steering | No correction until year-end | Quarterly review has nothing to say |
Case study: the £0 benefit chatbot. A wealth manager launched public GenAI FAQ. Marketing tracked sessions; finance tracked advised client revenue and complaint rate. Sessions rose 40%; revenue unchanged; complaints up 12% on misinterpreted tax guidance. Root cause: no link to KPI tree, no HITL for regulated topics, no lagging monitoring. Programme paused; remediation cost exceeded pilot.
Related playbook content
- Business case and prioritisation — structure benefits, scenarios, and approval narrative
- Discovery — baseline metrics and problem framing before sizing
- VALUE gate — quality bar for value evidence before scale
- FinOps and commercial — run-rate AI cost and chargeback (pairs with topic 07)
- Commercial value frameworks — ROI, NPV, and sensitivity patterns
- Business Learning overview — domain fluency packs and vocabulary depth
- Business Keywords — fast reference for exec meetings
- Domain Knowledge workbook — downloadable practice workbook
- Commercial and Financial Modelling — next-level TCO and pricing models
- How to use this Learning Map — study method and artefact standards
- 8D Framework — Define and Diagnose stages require solid value hypothesis
Practice checklist
- I can explain how the client makes money without opening a vendor deck
- I named a primary KPI with baseline and target direction
- I completed the primary practice exercise with pass criteria met
- I produced payback sketch with at least one sensitivity assumption
- I identified leading indicators for the first 90 days
- I know which stakeholder (CFO, COO, CMO, etc.) cares most about this KPI
- I documented at least two negative cases relevant to this client type
- I filed artefacts in my personal pattern library with date and sector
Discussion
Comments
Share feedback or questions about this page. No account required.
Loading comments…