Commercial and Financial Modelling
Executive view
Technical view
Why this matters
AI business cases fail in two ways: they never get approved because benefits are vague, or they get approved on false economics and collapse at scale. Finance leaders have seen enough "productivity AI" slides to ask hard questions: Which P&L line moves? Who loses hours on the ground? What is year-three run rate including human review? Commercial modelling is how you earn and keep funding—by speaking the language of NPV, payback, TCO, and benefits ownership.
AI cost structures differ from traditional IT. Licence-plus-implementation underestimates variable inference, embedding refresh, eval compute, red-team cycles, and human-in-the-loop staffing. A RAG assistant at 50K requests/month may spend more on reviewers than APIs if quality gates require it. TCO without these lines produces surprise run-rate escalations that kill phase two and damage trust—exactly when FinOps and product adoption (topic 06) need sponsor air cover.
Pricing models (fixed, T&M, subscription, consumption, managed service, outcome-based) must align incentives with uncertainty. Fixed price on open-ended agent scope transfers risk to the delivery firm; consumption aligns with token economics but needs showback so BUs do not spray requests. Outcome-based pricing sounds attractive but demands baselines, counterfactuals, and audit most enterprises lack on day one. Modelling clarifies which commercial shape fits which maturity level.
Benefits realisation closes the loop. Approvals happen once; value accrues over quarters. Registers with owners, baselines, measurement methods, and review cadence prevent AI programmes from becoming permanent pilots. Pair with Business case and prioritisation, AI FinOps and Commercial Design, and topic 06 adoption metrics.
Learn
Business case structure and narrative
Definition. A business case documents strategic alignment, problem and options, recommended investment, costs, benefits, risks, and implementation path—so a decision body can approve, defer, or reject with explicit assumptions.
Engagement use. Standard sections: executive summary (decision ask), strategic context, problem and scope, options considered, recommended option, costs (TCO), benefits (quantified), financial summary (NPV/ROI/payback), sensitivity, risks and mitigations, implementation and benefits plan, assumptions and dependencies. One page for executives; appendices for model detail. Align numbers with PRD scope (topic 06).
Pitfalls.
- Benefits before problem baseline exists.
- Single-option case disguised as analysis.
- Missing implementation and change costs.
- No explicit assumptions register.
- Technical jargon in executive summary.
Worked example. Executive summary: "Recommend £1.4M over 18 months for claims summarisation MVP serving 800 adjusters; NPV £2.1M at 10% discount; payback 22 months; key sensitivity: adoption ≥45% and review rate ≤30% of outputs."
Cost-benefit analysis (CBA) and financial metrics
Definition. CBA compares monetised costs and benefits over a horizon. NPV discounts future cash flows; ROI = (gain − cost) / cost; payback = time to recover investment. IRR sometimes used for portfolio comparison.
Engagement use. Use client discount rate and planning horizon (often 3–5 years for AI). Separate capex vs opex if client accounting requires. Include realisation phasing—benefits rarely start month one. Document confidence level per benefit line (high/medium/low).
Pitfalls.
- Undiscounted multi-year sums presented as NPV.
- ROI on year-one only while costs run three years.
- Double-counting benefits across use cases.
- Ignoring opportunity cost of same team building platform elsewhere.
Worked example. 3-year horizon, 10% discount: costs £2.8M PV; benefits £4.9M PV; NPV £2.1M; ROI 75%; payback month 22 when cumulative net turns positive.
Total Cost of Ownership (TCO) for AI
Definition. TCO aggregates build, run, and change costs over the analysis period—including people, not only licences and tokens.
Engagement use. Line items:
| Category | Examples |
|---|---|
| Build | Discovery, design, engineering, integration, data prep, eval harness, security review |
| Run | Inference/API, embeddings, vector DB, logging, monitoring, support, human review, retraining/refresh |
| Change | Corpus updates, model upgrades, regulatory change, training, programme management |
Phase costs: pilot, rollout, steady state. Link token assumptions to architecture (topic 22).
Pitfalls.
- API line only—omits review and knowledge ops.
- One-time build treated as no ongoing change.
- GPU capex omitted for private deploy.
- Support FTE at zero after go-live.
Worked example. Year-2 steady state for insurer assistant (800 users): API £380K, embeddings/index £45K, human QA sample review £620K, L2 support 2 FTE £180K, corpus ops £240K, eval/red-team £95K, platform share £110K → £1.67M/year run vs pilot year £420K—explains steering sticker shock without modelling.
AI-specific cost drivers
Definition. Variable and fixed drivers: tokens (prompt + completion), GPU (self-host), vector DB storage/query, data pipelines, eval compute, human review, change (corpus, prompts, models), observability, security/compliance tooling.
Engagement use. Build unit economics: cost per request, cost per successful task (topic 22), cost per user per month. Scenario low/medium/high volume (p50/p95). Include agent loop multiplier—3–8× single-shot for multi-step agents.
Pitfalls.
- Average tokens from demo prompts only.
- Ignoring retrieval embedding refresh on corpus churn.
- Human review modeled at 5% when policy requires 100% for tier-1 outputs.
- No cost of failed requests retried silently.
Worked example. 120K requests/month; 3,900 tokens/request blended; $2.50/1M blended rate → ~$1,170/month API—but review at 25% × 4 min × £35/hour loaded → ~£58K/month labour dominates.
Benefits types: revenue, cost, productivity, risk, CX/EX
Definition. Revenue (upsell, retention, faster sales cycle). Cost (avoided spend, vendor consolidation). Productivity (time saved × loaded rate). Risk reduction (fraud, compliance fines avoided—often non-cashable without history). CX/EX (NPS, engagement—usually non-cashable unless tied to churn/revenue).
Engagement use. Each benefit row: description, cashable?, baseline, target, calculation, owner, measurement source, realisation start month. Finance signs cashable lines; strategic lines tracked separately.
Pitfalls.
- Productivity = 100% of saved minutes × FTE count (ignores utilization and reinvestment).
- Risk benefits without incident history.
- Revenue uplift from AI with no attribution design.
- EX metrics presented as £ without linkage.
Worked example. Productivity: 800 adjusters × 12 min saved/day on eligible claims × 62% eligible mix × 45% adoption × 220 days × £40/hour loaded × 50% cashability factor (union agreement) = £1.05M/year cashable—not £3.2M gross story.
Cashable vs non-cashable benefits
Definition. Cashable benefits hit budget or P&L (headcount restraint, overtime reduction, avoided vendor, revenue). Non-cashable improve outcomes without automatic financial release (agent satisfaction, faster internal decisions, risk resilience).
Engagement use. Present both—finance approves on cashable core; programme tracks non-cashable for strategic narrative. Document realisation mechanism for cashable (e.g., "no backfill on attrition in pool X").
Pitfalls.
- Labeling all time saved as cashable—HR pushback at benefits review.
- Omitting non-cashable entirely—undervalues risk/CX cases.
- Cashable claims without workforce plan.
Worked example. Cashable: £1.05M productivity (with factor), £180K avoided translation vendor. Non-cashable: +4 NPS points on assisted queue; −12% regulatory finding risk on sampling (qualitative).
Sensitivity and scenario analysis
Definition. Sensitivity varies one assumption at a time (adoption, cost per request, benefit delay). Scenarios combine assumptions (base, upside, downside).
Engagement use. Tornado chart for NPV drivers. Identify break-even: minimum adoption, maximum cost per task, latest benefit start month. Present downside credibly—builds trust.
Pitfalls.
- Sensitivity only on token price—ignoring adoption.
- Upside scenario as "base case" in steering slides.
- No breakpoint narrative ("case fails below 38% WAU").
Worked example. Base NPV £2.1M; if adoption 35% not 45%, NPV £0.4M; if review rate 40% not 25%, NPV −£0.3M; if API cost +50%, NPV £1.6M—adoption and review dominate, not API.
Pricing models for AI solutions
Definition. Common models: fixed price (defined scope/deliverables), time and materials, subscription (platform/seat), consumption (per token/request/outcome unit), managed service (monthly fee + SLA), outcome-based (payment tied to KPI achievement).
Engagement use. Match model to uncertainty and data ownership. Internal chargeback often consumption + platform fee. Managed service fits ongoing corpus/eval ops. Outcome-based needs baseline audit and dispute process—usually phase three commercial shape.
Pitfalls.
- Fixed price without scope cap on requests or change requests.
- Consumption with no budgets/alerts—bill shock.
- Outcome-based without counterfactual (macro effects confound).
- Subscription priced below TCO run rate—vendor loss leader then churn.
Worked example. Client internal AI platform: £0.002/request internal transfer price (covers API + platform amortization) + £15K/month fixed corpus team; BU budgets capped with FinOps alerts at 80%—links to AI FinOps and Commercial Design.
Commercial risk and contract assumptions
Definition. Commercial risk includes scope creep, data delays, regulatory change, model deprecation, liability caps, IP ownership, audit rights, and exit cost.
Engagement use. RAID log commercial items; mirror in SOW assumptions (topics 29–30). Price contingency (5–15%) for AI uncertainty. Model liability separately from delivery fee.
Pitfalls.
- Fixed price without client data readiness dependency.
- Silent assumption of frontier model price drops.
- No clause for eval failure rework.
- IP on fine-tunes unclear.
Worked example. SOW assumption: "Client provides golden eval set ≥200 cases by week 4; delay shifts timeline and T&M kicks in." Contingency £220K on £1.4M for model change and corpus gap.
Benefits realisation and tracking
Definition. Benefits realisation is the post-approval discipline: owners measure actuals vs plan, explain variance, trigger corrective actions (product, change, scope).
Engagement use. Benefits register live in PMO rhythm—monthly or quarterly. Tie to adoption metrics (topic 06). Benefits realisation plan: baseline collection pre-go-live, measurement window, governance forum, stop/continue rules.
Pitfalls.
- Register frozen at approval—never updated.
- No owner after SI leaves.
- Measuring proxy only (logins) vs cashable driver.
- Declaring success without finance sign-off on cashable lines.
Worked example. Month 6 review: planned cashable £525K run-rate; actual £310K—variance driven by adoption 39% vs 45% plan and review rate 33%. Actions: change sprint (topic 26), FinOps cache project (topic 22), defer phase 2 languages.
Portfolio prioritisation under capital constraints
Definition. Ranking multiple use cases by value, cost, risk, strategic fit, and readiness when budget cannot fund all.
Engagement use. Normalised scorecard; efficient frontier narrative—fund A and B now; C waits for shared platform. Show dependency: platform enables C at lower marginal cost.
Pitfalls.
- Ranking by executive loudest voice.
- Ignoring shared platform amortization.
- Identical ROI horizons for 3-month and 24-month payback items.
Worked example. £5M capex ceiling; use cases A–F scored; fund A (£1.2M NPV £3.1M), B, D; defer C until A hits adoption gate—portfolio NPV £8.4M vs £9.1M if all funded with 18-month delay and higher risk.
Working capital, payment milestones, and cash flow
Definition. Cash flow timing differs from P&L: milestone billing, prepaid API credits, GPU deposits, and benefits lag affect when money moves—not only NPV totals.
Engagement use. Model monthly cash for vendor and client: 30/60/90 payment terms; Azure/GCP commit discounts paid upfront; align milestone invoices to eval gates. Show CFO peak funding need month—not only year-one average.
Pitfalls.
- NPV without cash timing—approved project starves mid-build.
- Prepaid cloud commit assumed zero in TCO.
- Benefits start month optimistic—cash gap unspoken.
Worked example. £900K build over 6 months; client pays 30% start, 40% eval gate, 30% go-live; vendor GPU prepay £110K month 2—peak client outlay month 4 £520K before benefits begin month 7; treasury needs visibility.
Capitalisation vs expense (AI assets)
Definition. Accounting rules may capitalize certain software development costs vs expense R&D—varies by jurisdiction and client policy. Affects ROI optics and balance sheet.
Engagement use. Early alignment with client finance: which workstreams are capex-eligible (platform, integration) vs opex (pilot learning, training). Document in business case appendix—do not pretend to be accountants; flag for finance decision.
Pitfalls.
- Assuming all build is capex—audit rejection.
- Expensing everything—ROI looks worse than peer projects.
- Mixing vendor licence (opex) with custom build (capex) without split.
Worked example. Client capitalizes internal labour on qualifying platform epics only (£420K of £780K build); remainder opex—NPV unchanged but year-one ROI presentation differs for steering narrative.
Multi-currency and inflation in long-horizon cases
Definition. Global programmes face FX exposure (USD API, local labour) and inflation on FTE and vendor renewals.
Engagement use. State currency basis; sensitivity on ±10% FX on API-heavy cases; index FTE costs 2–4% in years 2–3 for 5-year horizons. Hedge narrative: reserved throughput contracts lock partial API rate.
Pitfalls.
- USD API in case, local currency benefits without conversion note.
- Ignoring renewal uplift on 3-year enterprise agreements.
Worked example. EU case in EUR; API priced USD 60% of run—10% USD appreciation increases NPV cost €180K; mitigation: regional open-weight route for 40% traffic planned year 2.
Warranty, liability, and insurance cost lines
Definition. Commercial risk has price: professional indemnity uplift, liability caps, error & omission coverage, warranty periods on deliverables.
Engagement use. For outcome-touched AI (financial advice adjacent, medical adjacent), model insurance or contingency line 3–8% of fee. Link to contract caps (topic 30).
Pitfalls.
- Zero contingency on high-liability use cases.
- Unlimited warranty implied in fixed price.
Worked example. Legal research assistant: £85K contingency (7% of build) for remediation sprint if eval fails regulatory review—explicit in case, reduces partner margin surprise.
Discount rates, hurdle rates, and approval thresholds
Enterprises use hurdle rates (minimum acceptable IRR/NPV) that may exceed nominal WACC for risky tech. AI programmes often face +200–400 bps risk premium internally. Ask finance explicitly: "What discount rate and hurdle apply to this class of investment?"
Engagement use. If hurdle is 15% IRR and base case shows 12%, document mitigations (phasing, faster payback option, risk reduction milestones) or recommend deferral. Do not silently use generic 10% when client policy differs.
Worked example. Client hurdle 14% IRR; base AI case 13.2%—restructure to phase 1 lower capex achieving 16.8% IRR standalone; phase 2 option valued separately—approval unlocked without dishonest bundling.
Frameworks and methods
TCO three-horizon model
Pilot (0–6 months): prove value, higher unit cost acceptable. Rollout (6–18): integration + change dominate. Steady state (18+): run + change ops dominate. Model each separately—do not extrapolate pilot API bill linearly.
Benefits register template
| ID | Benefit | Cashable? | Baseline | Target | £/year | Owner | Measure | Start month | Confidence |
|---|
Sensitivity tornado (typical AI drivers)
- Adoption / active usage rate
- Human review percentage and minutes
- Benefit realisation delay
- Volume (requests/month)
- Token cost / model routing
- Integration delay (benefits shift right)
Pricing model selection matrix
| Model | Best when | Watch out |
|---|---|---|
| Fixed price | Scope fixed, goldens agreed, low unknowns | Agent scope creep |
| T&M | Discovery-heavy, client retains risk | Budget overrun perception |
| Subscription | Platform reuse, seat-based value | Below run-rate TCO |
| Consumption | Variable volume, mature FinOps | Bill shock without caps |
| Managed service | Ongoing corpus/eval ops | SLAs vs cost trade-off |
| Outcome-based | Auditable baseline, single KPI | Attribution disputes |
Real options thinking (defer / expand)
Treat phase 2 funding as option—pay small for pilot learning; expand only if gates hit. Models better decision quality than all-or-nothing NPV on full scale day one.
8D and VALUE alignment
Business case feeds Fund/Commit gate in 8D Framework. Benefits register feeds Operate reviews. Cross-check VALUE gate assumptions.
Real-world scenarios
Scenario A: UK retail bank contact centre assistant
Context. 4,200 agents; pilot 620; approved expansion to 2,000 agents if case holds.
TCO (3 years, £000s). Build/integration £2,100; run steady-state year 3 £1,850/year (API £320, review £780, ops £410, platform £340); change £620/year avg.
Benefits. Cashable productivity £2.4M/year at target adoption (45% WAU, 50% cashability factor); non-cashable NPS +3 on assisted queue.
Sensitivity. NPV base £4.2M; fails if WAU <36% or review >38% of sessions.
Pricing. Internal consumption charge £0.0018/request + shared platform levy; BU showback monthly.
Benefits realisation (month 9). Actual WAU 41%—on track; cashable run-rate £1.7M vs plan £1.9M—variance explained by slower branch rollout; no scope expansion until month 12 review.
Scenario B: Manufacturing spare-parts search (EU)
Context. €180M aftermarket division; 320 field engineers; parts lookup drives 18% of downtime tickets.
Business case. Build €940K; run €290K/year (lower review—lookup not customer-facing); benefits €1.1M/year avoided downtime (cashable via SLA penalties avoided—historical €2.3M/year penalties, conservative 40% attribution).
Pricing (SI to client). Phase 1 fixed €680K; phase 2 T&M capped €120K; managed corpus €18K/month optional.
Sensitivity. Attribution 30% not 40% still NPV positive; penalty reduction must be measured per region—baseline year agreed in contract.
Outcome. Steering approves with downside scenario NPV still €0.6M—credibility from honest attribution band.
Scenario C: Health payer prior-auth summarisation (US — stretch)
Context. 12M auth requests/year; pilot 50K; strict HIPAA; nurse review 100% on medical necessity summaries in MVP.
TCO insight. API $180K/year at scale trivial vs $4.2M/year nurse review time reallocated—not eliminated; cashable benefit = throughput enabling 18% more auths without headcount ($3.1M), not raw labour deletion.
Commercial. Outcome-based vendor offer rejected—baseline auth throughput contested. Client chooses subscription + consumption cap with FinOps dashboard.
Lesson. Benefits realisation must match workflow reality—100% review makes "FTE removal" case dishonest; throughput case wins.
Practice exercises
Primary exercise: Full TCO + benefits register + sensitivity (3–4 hours)
Prompt: RAG assistant for internal HR policy Q&A; 25,000 employees; expected 8 queries/employee/month at steady state; 15% review rate at 3 min/review.
Deliver:
-
TCO table — build, run (year 2), change (annual); minimum 12 line items.
-
Benefits register — 5 rows minimum; mark cashable; show calculation.
-
Financial summary — 3-year NPV, ROI, payback (state discount rate).
-
Sensitivity — tornado narrative with two break-even thresholds.
-
Pricing recommendation — one internal chargeback model with rationale.
Acceptance criteria: Human review cost exceeds API in run state; at least one benefit is non-cashable; downside scenario documented.
Stretch exercise: Fixed-price bid decision (2 hours)
You are bidding fixed £950K for 12-month agent deployment. Client scope vague; volume unknown. Write go/no-go memo with commercial risks, contingency %, assumptions for SOW, and alternative T&M structure.
Acceptance criteria: Lists ≥6 commercial risks; recommends structure with numbers; includes eval failure rework clause recommendation.
Reflection exercise: Benefits realisation variance (30 minutes)
Pick a fictional month-6 miss (adoption −20%). Write variance commentary and three corrective actions mapped to topics 06, 22, 26—not "upgrade model" as first lever.
Questions you should be able to answer
- What is three-year TCO split across build, run, and change?
- Which cost line dominates steady-state—API, review, or integration ops?
- What benefits are cashable vs non-cashable, and who signed cashability?
- How was productivity translated from minutes saved to pounds?
- What discount rate and horizon did you use—and why?
- What is NPV, ROI, and payback in the base case?
- What two assumptions break the case in sensitivity analysis?
- What is cost per successful task—and how does it feed the model (topic 22)?
- Which pricing model fits client maturity—and what are the incentives?
- What contingencies are included, and what risks do they cover?
- What baselines were captured pre-go-live for benefits tracking?
- Who owns each benefit line after the SI rolls off?
- What is the downside scenario NPV—and is it still fundable?
- How do portfolio constraints affect phasing of this use case?
- What triggers a benefits realisation stop/continue review?
Negative cases
Pilot extrapolation. £30K pilot → £30K/year production assumption. Fix: steady-state TCO with volume percentiles.
API-only TCO. Fix: review, eval, corpus, support lines mandatory.
Cashable fantasy. 100% FTE removal from time studies. Fix: cashability factors and workforce plans.
Benefits without baselines. Fix: measure pre-go-live; control cohort where possible.
Single-option case. Fix: real alternatives including process-only.
Sensitivity theatre. Token price only. Fix: adoption and review drivers first.
Outcome pricing day one. Fix: prove measurement; phase commercial model.
Frozen register. Fix: monthly/quarterly benefits forum.
Double-counting portfolio. Fix: shared benefit ownership map.
Hidden commercial risk. Fixed price + unlimited change requests. Fix: assumptions and caps in SOW.
Model deprecation surprise. Fix: model-agnostic architecture cost and contract clauses.
Ignoring change cost. Corpus drift treated as zero. Fix: change FTE or managed service line.
Operating model: commercial governance through delivery
Financial models are not static PDFs—they are living instruments governed through delivery.
Governance forum (recommended). Monthly value and economics session: finance, product owner, delivery lead, FinOps (topic 22), change lead. Agenda: actual vs plan spend, cost per successful task trend, benefits register variance, sensitivity re-check if drivers moved >15%, decision on phase gates.
Roles.
| Role | Commercial accountability |
|---|---|
| AI Solution Engineer | Model integrity, assumption log, scenario updates |
| Client finance partner | Cashability sign-off, discount rate, benefit recognition |
| Product owner | Adoption drivers linked to benefits lines |
| FinOps lead | Run-rate actuals, chargeback reports |
| PMO | Benefits register hygiene, meeting cadence |
Version control. Business case v1.0 at fund gate; v1.x when scope or volume shifts >20%; full refresh at pilot exit before scale funding.
Step-by-step NPV construction (worked mini-example)
Assumptions: 10% discount rate; 36-month horizon; costs front-loaded; benefits start month 7 ramping to steady month 12.
| Month | Cost (£K) | Benefit (£K) | Net (£K) | PV factor | PV net (£K) |
|---|---|---|---|---|---|
| 0–6 (build) | 900 | 0 | −900 | 1.0–0.87 | −780 |
| 7–12 (ramp) | 200 | 450 | 250 | 0.75 avg | 188 |
| 13–36 (steady) | 140/mo | 520/mo | 380/mo | 0.65 avg | ~7,400 cumulative PV benefit minus costs |
Simplified narrative: cumulative PV benefits £8.2M, PV costs £6.1M, NPV £2.1M—always show formula and inputs in appendix for finance audit.
Pricing deep dive: consumption + platform hybrid (internal chargeback)
Most enterprises land on hybrid pricing for internal AI platforms:
-
Platform fee — amortizes shared ingestion, eval, security, FinOps tooling (£ fixed/month per BU or enterprise).
-
Variable consumption — £ or $ per 1K tokens or per successful task at marginal cost + small margin.
-
Capacity reservation — optional prepay for peak TPM/GPU guaranteeing SLA.
Example rate card (illustrative):
| Line | Rate | Notes |
|---|---|---|
| Platform base | £22K/month enterprise | Shared corpus ops, eval, logging |
| Inference marginal | £0.0016 / 1K tokens blended | Pass-through + 12% |
| Successful task surcharge | £0.04 / success | Aligns BUs to outcomes not spam |
| Human review (if shared QA) | £0.18 / reviewed output | Transparent labour pass-through |
| Premium frontier route | +35% multiplier | Requires product approval flag |
Governance: BU budgets set quarterly; 80/100% alerts; frontier multiplier requires product + finance approval for new use cases.
Pair implementation details with AI FinOps and Commercial Design.
Benefits realisation operating rhythm
| Week | Activity |
|---|---|
| −4 to 0 pre-go-live | Capture baselines (AHT, volume, error rate, cost) |
| 0 | Launch with adoption war room (topic 26) |
| 2 | First telemetry read—leading indicators only |
| 6 | First variance report vs plan—no cashable claims yet |
| 12 | Cashable benefits preliminary—finance review |
| 26 | Full benefits realisation review—continue/stop/expand |
Variance template: planned vs actual vs driver decomposition (adoption effect vs unit cost effect vs volume effect)—prevents blaming "AI" generically.
Scenario D: Retail media network — ad copy generation (stretch)
Context. Retail media arm; 120 brand partners; gen-AI for ad copy variants from product feeds; 8M SKU attributes/month batch + 15K interactive edits/month.
TCO twist. Batch embed + generate dominates; human brand review 100% for premium tier, 20% sample for long-tail tier.
Benefits. Revenue uplift from +11% campaign CTR on AI-assisted variants (A/B platform data)—£2.8M/year incremental media revenue at 35% margin → £980K cashable gross margin; cost save secondary.
Sensitivity. CTR uplift 7% not 11% → NPV still positive; 4% → case fails—CTR measurement pre-registered.
Pricing (SI). T&M build + revenue share 2% on attributed uplift band—client prefers aligned incentive over fixed £1.1M.
Shows revenue-linked benefits and hybrid commercial structures—not all AI cases are cost-out.
Scenario E: Logistics — document extraction for customs (EU)
Context. Freight forwarder; 2.3M customs docs/year; 18 fields extracted; current manual keying €4.2M/year; error rate 3.8% causing delays/fines €600K/year historical.
TCO (3 years). Build €1.1M (integration-heavy); run €380K/year (API €95K, human exception handling €210K, ops €75K); change €120K/year.
Benefits (cashable). Labour reduction €2.1M/year at 85% straight-through processing (STP) target with 50% cashability factor (redeploy not layoff) = €892K/year; fine reduction €240K/year at 60% attribution = €144K/year; total cashable €1.04M/year steady.
Sensitivity. STP 70% not 85% → NPV still €0.9M positive; STP 55% → NPV negative—pilot must prove STP on representative doc mix not cherry-picked PDFs.
Pricing. Managed service €45K/month includes change corpus retraining quarterly—client prefers opex predictability over capex spike.
Benefits realisation twist. Month 8 STP 62%—below plan; variance driver = low-quality scans from 3 origin countries; product fix (topic 06) adds capture guidance epic, not model swap—commercial case updated with 3-month benefit delay in downside scenario.
Payback period and break-even analysis (step-by-step)
Payback = first period cumulative net cash flow turns positive. Break-even volume = minimum successful tasks/month where benefits = run-rate costs.
Worked example (continuing logistics case).
- Steady run cost €380K/year (€31.7K/month)
- Steady cashable benefits €1.04M/year (€86.7K/month)
- Net monthly steady €55K/month
- Build €1.1M over 8 months avg €137.5K/month outflow before benefits
- Benefits start month 9 at 50% ramp → €43.3K/month month 9–11, full €86.7K month 12+
Cumulative payback: months 1–8 −€1.1M build; months 9–11 partial benefit +€130K; month 12+ need ~18 months from project start to recover remaining gap—payback ~20 months from kickoff. Finance may use discounted payback—state both if client asks.
Break-even STP: If each point of STP ≈ €24.7K/year benefit, break-even STP ≈ 380/1040 × 85% ≈ 31% STP to cover run alone—build recovery needs higher—useful sanity check when pilot STP is 62%.
External vendor comparison in business case
When build-vs-buy decision exists, model three options with identical benefit assumptions:
| Option | 3-yr TCO | NPV | Notes |
|---|---|---|---|
| Build custom | €2.4M | €1.8M | Highest control |
| Vendor SaaS | €2.9M | €1.2M | Faster time-to-value |
| Hybrid | €2.6M | €1.5M | SaaS + custom integration |
Prevents business case becoming build advocacy without finance-grade comparison—weighting includes exit cost, data residency, and eval control.
Tax, transfer pricing, and internal recharge (enterprise context)
Large enterprises often require internal recharge across BUs or countries. Document: cost center codes, transfer pricing policy for shared AI platform, VAT on managed services, and whether FinOps showback becomes mandatory recharge after pilot. AI Solution Engineers flag these early—misaligned recharge kills adoption when BUs see unpredictable bills. Work with client finance; do not invent tax treatment. Example: UK HQ recharges EU BU in EUR at monthly FinOps actuals + 5% platform admin fee per contract schedule—benefits case must use net benefit to BU after recharge for local steering truth.
Scenario summary table (quick reference for steering)
| Scenario | Sector | Key metric | Base NPV | Primary sensitivity |
|---|---|---|---|---|
| A | Retail banking | WAU 45%, review 25% | £4.2M | Adoption floor 36% |
| B | Manufacturing | Penalty attribution 40% | €0.9M+ | Attribution 30% |
| C | Health payer | Throughput not FTE cut | $3.1M benefit | Review 100% |
| D | Retail media | CTR +11% | £980K/yr margin | CTR 7% |
| E | Logistics | STP 85% | €1.8M (illustrative) | STP 55% fails |
Use this pattern in your own cases—executives remember one table more than thirty slides.
When presenting to CFO, lead with downside NPV and payback, not upside hero numbers—credibility on commercial topics is built from honest breakpoints, not peak-case optimism.
Document every assumption change in a model changelog—when steering asks "why did NPV move?", you answer with version diff (adoption 45%→41%, review 25%→30%), not memory.
Commercial checklist before steering (self-audit)
- Baselines sourced and dated—not estimates from sponsor memory
- TCO includes year-3 change line ≥15% of run for corpus/model drift
- At least one downside scenario presented with NPV
- Cashable lines have workforce or revenue mechanism named
- Pricing model matches FinOps maturity (no consumption without alerts)
- Benefits owners named with job titles, not "TBD"
- Sensitivity identifies non-API drivers first
- Contract assumptions mirrored in RAID—not only in legal doc
Integration with FinOps and product (topics 06, 22)
The commercial model breaks when disconnected from product and FinOps:
| Integration point | What to sync |
|---|---|
| PRD adoption targets | Benefits register adoption assumptions |
| Cost per successful task (22) | TCO run-rate line item |
| MVP scope (06) | Build cost and benefit start month |
| Review rate in product | Human review cost line |
| Chargeback rates (22) | Internal pricing recommendation |
| Eval gate failures | Benefit delay in downside scenario |
Example failure mode: Business case assumes £0.06/success; production £0.22/success after agent loops—steering must trigger case revision, not silent margin erosion.
Related playbook content
- Business case and prioritisation — approval narrative and portfolio scoring
- AI FinOps and Commercial Design — chargeback, showback, and commercial operating model
- Model FinOps roadmap — cost engineering depth path
- Performance Engineering and AI FinOps — unit economics inputs to TCO
- Product Management — adoption metrics drive benefits
- Change Management and Adoption — realisation depends on usage
- Presales and Solution Shaping — commercial shape in sales cycle
- RFP, Procurement and Contracting — contract assumptions
- Commercial value frameworks — benefits and value patterns
- 8D Framework — Fund/Commit and Operate gates
- How to use this Learning Map — study loop and artefact standards
Practice checklist
- I can explain steady-state TCO without counting API twice
- Human review and eval appear as explicit cost lines
- Cashable benefits have finance-agreed calculation and owner
- Sensitivity identifies adoption or review as dominant drivers
- Downside scenario is credible—not straw-man
- Pricing model matches uncertainty and FinOps maturity
- Benefits register includes baselines and start months
- I completed the primary exercise and filed models in my pattern library
Discussion
Comments
Share feedback or questions about this page. No account required.
Loading comments…