D2 · AI Strategy and Business Value Creation

This is the heaviest domain on the exam — 28% of the blueprint, roughly 24 of 85 items on our mock. It is where business judgment meets arithmetic: which use cases to pursue, whether to build, buy or partner, how to prioritise a portfolio and when to scale, pause or terminate, how to design KPIs and a baseline, and how to calculate ROI and payback that a CFO will accept. Almost every item hinges on one discipline the exam rewards above all others: define and measure the value before you spend, and be willing to kill what does not pay. If you master the numbers on this page, they also underwrite the value arguments the governance and transformation domains assume.

What you need to know

Strategy starts from business outcomes, not from technology: you find high-impact use cases across customer operations, sales and marketing, R&D and software development, then map each AI capability to a measurable outcome. Build-buy-partner is a decision on budget, timeline, capability and regulatory constraint, with AWS Marketplace as the surface for evaluating buy and partner options. You prioritise on value, feasibility, sustainability and strategic alignment, and you decide to scale, pause or terminate on evidence. You establish a baseline before deployment, design KPIs that capture tangible and intangible benefit, and compute ROI and payback with real numbers. Leading indicators tell you early whether an initiative will work; lagging indicators confirm it later. Cost control means choosing the right pricing structure, using commitment discounts and levers such as batch and routing, forecasting with the AWS Pricing Calculator and tracking actuals with Cost Explorer.

Learning objectives

By the end of this page you should be able to:

Identify high-impact AI use cases across business functions and map capability to outcome.
Evaluate build-buy-partner decisions on budget, timeline, capability, vendor proposal and regulatory constraint.
Prioritise initiatives on value, feasibility, sustainability and strategic alignment, and decide to scale, pause or terminate.
Recognise when AI is not the appropriate solution.
Assess transition considerations: business continuity, cost, data readiness, performance impact and platform migration.
Design KPIs that capture tangible and intangible benefits.
Construct a baseline before deployment, and act sensibly when no baseline exists.
Calculate ROI, payback and a multi-year NPV-style case with real arithmetic.
Distinguish leading indicators that predict success from lagging indicators that confirm it.
Apply cost-factor controls: pricing structures, commitment discounts, batch and routing levers, Pricing Calculator and Cost Explorer.
Match investment level to industry maturity and competitive dynamics.

Task statements covered

Official task / skill	Where this page teaches it
2.1.1 High-impact use cases; capability-to-outcome mapping	2.1
2.1.2 Build-buy-partner decisions	2.2
2.1.3 Prioritise on value, feasibility, sustainability, alignment; scale/pause/terminate	2.3, Decision framework
2.1.4 When AI is not the appropriate solution	2.4
2.1.5 Transition considerations and platform migration	2.5
2.2.1 KPIs — tangible and intangible benefits	2.6
2.2.2 Baseline metrics before implementation	2.7
2.2.3 ROI frameworks	2.8
2.2.4 Leading indicators that predict success	2.9
2.2.5 Cost-factor controls; cost planning and optimisation	2.10
2.3.1–2.3.4 Competitive advantage, business-model transformation, investment level	2.11

2.1 Finding high-impact use cases and mapping capability to outcome

Strategy begins by asking where does AI create the most value here, not where can we bolt on AI. High-impact use cases cluster in four functions the exam names explicitly.

Function	Typical high-impact use case	AI capability	Business outcome
Customer operations	Automated triage, self-service resolution, agent assist	NLP, summarisation, retrieval	Lower handle time, higher deflection, CSAT
Sales and marketing	Lead scoring, personalised content, next-best-action	ML prediction, GenAI content	Higher conversion, revenue per rep
R&D	Literature synthesis, candidate generation, simulation	GenAI, ML	Faster cycles, more shots on goal
Software development	Code assist, test generation, documentation	GenAI (code)	Developer throughput, defect reduction

The discipline the exam rewards is the capability → outcome map: state the AI capability, the process it changes, the metric it moves and the size of the prize before choosing a tool.

A use case with no metric it moves is not a use case; it is a wish. When a stem describes buying a tool before naming the outcome, the correct answer reverses the order.

Exam signal

Stems that say “the CEO wants to use AI”, “a vendor demoed something impressive”, “we should adopt GenAI” with no stated outcome are testing whether you start from the business problem. The right answer names the outcome and metric first; distractors buy the technology first.

2.2 Build vs buy vs partner

Once a use case is chosen, decide how to source the capability. The four levers are budget, timeline, capability (do we have the skills and differentiation) and regulatory constraint.

Option	When it wins	Watch out for
Build	The capability is core, differentiating, and you have the skills and data	Slowest and most expensive; opportunity cost; ongoing maintenance
Buy	The need is common and non-differentiating; speed matters	Vendor lock-in, data terms, fit gaps; still needs integration
Partner	You need capability and shared risk on something strategic but not yet core	Dependency, IP and data-sharing terms, exit clauses

AWS Marketplace is the exam’s named surface for evaluating buy and partner options — a catalogue where you compare third-party solutions, terms and pricing against a build estimate. Evaluate a vendor proposal on more than price: fit to the use case, data-training and residency terms, security and certifications, total cost including integration, and an exit path.

Worked example. A common document-extraction need, non-differentiating, wanted in eight weeks.

The strategic rule: build only what differentiates you. Building a commodity capability is the classic value-destroying answer.

2.3 Prioritisation, and the scale / pause / terminate decision

You will always have more candidate initiatives than capacity. Prioritise on four axes — value, feasibility, sustainability, strategic alignment — and be willing to stop.

Value — size of the outcome, confidence in it.
Feasibility — data readiness, technical difficulty, organisational capacity.
Sustainability — can you run and fund it past the pilot; is the cost model viable at scale.
Strategic alignment — does it advance a stated business objective.

Once an initiative is running, the same evidence drives a periodic scale / pause / terminate call:

Decision	When	Signal
Scale	Meets or beats its success metric, sustainable at scale, aligned	Baseline beaten, unit economics hold
Pause	Promising but blocked on a fixable condition (data, capability, cost)	Fixable gap; revisit date set
Terminate	Fails value or feasibility, or no longer aligned; sunk cost is not a reason to continue	Metric missed, cost unsustainable, superseded

The exam’s hardest instinct here is terminating a sunk-cost initiative. Money already spent is gone whether you continue or not; the only question is whether future spend earns its return. An initiative that has missed its baseline and has no fixable path should be terminated, not nursed.

2.4 When AI is not the answer

A strategist earns credibility by saying no to AI when it is the wrong tool. AI is not the answer when:

The problem is deterministic and better served by a rule (see Domain 1).
The data does not exist or is too poor to support a model, and cannot be fixed in time.
The cost exceeds the value, or the value cannot be measured at all.
The risk is unacceptable for the tolerance for error (an irreversible, unregulatable use).
The real problem is a process or data problem that AI would only paper over.

“Fix the broken process first, then decide whether AI helps” is frequently the correct answer. Adding AI on top of a broken process usually scales the mess.

2.5 Transition considerations and platform migration

Moving a process to AI — or from one AI platform to another — is a change-management and continuity problem, not just a technical switch. Five considerations recur.

Consideration	The question	Risk if ignored
Business continuity	Can the process still run if the AI fails?	Outage with no fallback
Cost implications	What does the new model cost at real volume, including migration?	Pilot economics that break at scale
Data readiness	Is the data available, clean and accessible for the new process?	Garbage-in at scale
Performance impact	Does the change meet latency and quality needs?	Faster but worse, or slower than acceptable
Platform migration	What is the exit and switching cost of the current platform?	Lock-in; painful, expensive migration

The discriminator: a transition plan without a fallback and a phased rollout is the wrong answer. Cut-over-everything-at-once is a continuity risk.

2.6 KPI design — tangible and intangible benefits

A KPI is only useful if it is tied to the outcome and can be measured. The exam distinguishes two families.

Benefit type	Examples	How to make it measurable
Tangible	Cost reduction, revenue growth, time saved, throughput	Currency or count, against a baseline
Intangible	Customer satisfaction, employee productivity, brand, risk reduction	Proxy metrics (CSAT/NPS, engagement, retention)

Design a small KPI tree, not a dashboard of everything:

The guardrail is what stops you “winning” on cost while quietly wrecking the customer experience. Intangible benefits are real value; the exam wants you to proxy and track them, not to dismiss them because they are harder to quantify.

2.7 Baselines before deployment

This is the most examined discipline in the whole domain. You cannot prove value you did not measure before you started. A baseline is the current-state measurement of the KPIs you intend to move, captured before the AI goes live.

How to construct one: pick the KPIs tied to the outcome, measure them for a representative period (long enough to cover normal variation), record the method so the “after” is measured the same way, and note the conditions (volume, seasonality).

When no baseline exists: do not abandon measurement. Options in order of preference — (1) run a short pre-deployment measurement period; (2) use a control group (some units keep the old process) to compare concurrently; (3) reconstruct from historical records; (4) as a last resort, an A/B or holdout during rollout. What you must not do is deploy first and claim value later from memory. “We think it’s faster” is not value; a measured delta is.

Exam signal

Stems that say “leadership wants to know the ROI” after deployment, or “there is no way to measure the before state”, are baseline items. The correct answer establishes or reconstructs a baseline (control group, historical data, pre-measurement) rather than accepting an unmeasurable claim.

2.8 ROI frameworks with worked arithmetic

ROI turns a value story into a number a CFO will fund. The building blocks are time savings, cost reduction, revenue growth and productivity gains, netted against cost.

Basic ROI and payback.

Worked case — support automation.

Multi-year, NPV-style case. Money next year is worth less than money today; discount future benefits. Using a 10% discount rate on the same $110,000/yr net benefit, with the $60,000 spent up front:

A positive NPV of ~$213.5k over three years says the investment creates value even after discounting. The exam expects you to (a) net benefit against all cost including oversight, (b) not confuse gross saving with net, and (c) recognise that a payback under a year with a positive multi-year NPV is a strong case. The subtle trap is counting time saved as cash without asking whether the freed hours are redeployed — if the agents are not reassigned or reduced, the “saving” is capacity, not cash, and should be labelled as such.

2.9 Leading versus lagging indicators

You steer with leading indicators and you report with lagging ones. Confusing them means you find out too late.

	Leading indicators	Lagging indicators
Timing	Early, predictive	Later, confirmatory
Answer	“Is this going to work?”	“Did it work?”
Support example	Adoption rate, deflection rate, first-week usage	Annual cost saved, CSAT trend, revenue
Use	Adjust the initiative in flight	Prove value, decide scale

The exam wants you to pick leading indicators to manage the pilot (low adoption in week one predicts failure while you can still fix it) and lagging indicators to decide scale. An initiative reporting only lagging metrics discovers problems after the money is spent.

2.10 Cost-factor controls

Cost discipline is a strategy skill, not an afterthought. Start from the pricing structure, then apply levers.

Pricing structure	You pay for	Best when	Watch out for
Consumption-based	Usage (per token, request, page, image)	Variable or uncertain volume; scales to zero	Cost spikes with volume; hard to cap
Instance-based	Provisioned compute per hour	Steady, predictable, high utilisation	Idle capacity burns money
Seat-based	Per user per month	Predictable per-head budgeting; decouples from volume	Paying for low-usage seats

Levers, at a business level:

Commitment discounts — Savings Plans, Provisioned Throughput commitments and reserved tiers trade flexibility for a lower rate; only worth it at high, predictable utilisation.
Batch processing — non-urgent work run in batch can be substantially cheaper than on-demand.
Routing — sending simpler requests to smaller/cheaper models and only hard ones to premium models controls average cost.
Right-sizing context — RAG and summarisation cut tokens per request (see Domain 1).

Tools: the AWS Pricing Calculator produces a forward estimate before you commit; AWS Cost Explorer shows actuals, trends and anomalies after you are live. The pairing the exam rewards: forecast with the Calculator, verify and investigate with Cost Explorer.

Worked lever example.

2.11 Competitive positioning and investment level

The last task ties value to strategy. AI can create advantage three ways: operational improvement (do the same thing cheaper/faster/better), business-model transformation (do something the business could not do before — new products, new pricing, new channels) and defensive positioning (keep pace so rivals do not pull ahead).

Match investment level to industry maturity and competitive dynamics.

Situation	Appropriate posture
Early industry, few adopters, high uncertainty	Targeted experiments; option value, not big bets
AI becoming table-stakes in your sector	Invest to keep pace; falling behind is the real risk
A clear, defensible differentiation you can own	Invest to build; this is where “build” beats “buy”
Mature, commoditised capability everyone has	Buy the commodity; spend your budget on what differentiates

The trap is uniform investment regardless of context — either over-investing in a commodity or under-investing where AI is becoming table-stakes. Sustainable advantage comes from proprietary data, integration into the workflow and continuous improvement, not from having bought a model everyone can buy.

Decision framework

The value-feasibility-risk prioritisation grid

Score each candidate initiative 1–5 on three composite dimensions, weight them, and let the total plus the risk gate drive the scale/pause/terminate call.

Dimension	What it rolls up	1 (low)	5 (high)
Value	Outcome size × confidence × strategic alignment	Small, speculative, off-strategy	Large, proven, core to strategy
Feasibility	Data readiness × technical difficulty × capacity	Poor data, hard, no capacity	Ready data, straightforward, capacity exists
Risk (inverted)	Regulatory, reputational, reversibility, sustainability	Severe, irreversible, unsustainable	Low, reversible, sustainable

Scoring rule. Weighted score = Value × 0.5 + Feasibility × 0.3 + Risk × 0.2. Then apply gates: any dimension scoring 1 triggers pause (fix the gap) or terminate (if unfixable); a high weighted score with all gates clear supports scale.

Worked portfolio. Three live initiatives reviewed this quarter:

Initiative	Value	Feasibility	Risk	Weighted	Evidence	Call
A Support triage	5	4	4	4.5	Beat baseline (handle time −25%), unit cost holds at scale	Scale
B Sales content gen	4	2	3	3.2	Promising, but data access blocked pending a legal review	Pause
C Predictive maintenance	2	2	2	2.0	Missed baseline twice, sensor data too sparse, no fix in sight	Terminate

Reasoning trace. A clears every gate and has beaten its baseline with sustainable unit economics — scale it, and reuse its pattern. B scores well on value but a 2 on feasibility from a fixable blocker (legal sign-off on data access) — pause with a revisit date, do not kill a promising initiative for a temporary obstacle. C has failed its baseline twice on data that cannot be fixed in the relevant horizon; the money already spent is sunk and irrelevant — terminate, and redeploy its budget to scaling A. The framework’s discipline is refusing to let sunk cost keep C alive or let B’s fixable blocker be mistaken for failure.

Common mistakes

Mistake	Why it happens	What to do instead
Buying technology before defining the outcome	A demo is exciting; procurement is tangible	Name the outcome and the metric first; source the capability second
Deploying without a baseline	Eagerness to launch; baselines feel like delay	Measure the before-state, or use a control group / historical data
Counting gross savings as ROI	It flatters the case	Net against all cost including oversight and change
Treating time saved as cash automatically	Hours feel like money	Only cash if hours are redeployed or headcount changes; else it is capacity
Building a commodity capability	“Build” feels like control	Build only what differentiates; buy the commodity
Nursing a sunk-cost initiative	Loss aversion; “we’ve spent so much”	Judge future spend on future return; terminate what cannot pay
Scaling a pilot whose success was never measured	Enthusiasm and momentum	Scale only on a beaten baseline with sustainable unit economics
Reporting only lagging indicators	They look authoritative	Track leading indicators to steer while you still can
Choosing a pricing structure by habit	Familiarity	Match consumption/instance/seat to the volume and predictability
Cutting over to a new platform all at once	Speed and neatness	Phase the rollout with a fallback for business continuity
Uniform investment regardless of context	One rule is simpler	Match investment level to industry maturity and competitive dynamics
Dismissing intangible benefits	Hard to quantify	Proxy and track them; they are real value

Scenario walkthrough

Scenario. You advise a mid-sized insurer. The COO wants a decision on three AI initiatives and a defensible ROI story for the board. Initiative one: a claims-summarisation assistant that drafts a summary of each claim file for adjusters; a four-week pilot cut adjuster reading time from 20 minutes to 12 minutes per claim across 200,000 claims a year, and a vendor on AWS Marketplace offers it at $9,000/month consumption. Initiative two: a fraud-scoring model the team has been building for nine months; it has missed its accuracy target twice, the labelled-fraud data is sparse, and there is no path to more data this year — but $500,000 has been spent. Initiative three: an AI chatbot the CMO wants “because our competitor launched one”, with no stated metric. The board wants numbers, not adjectives, and the CFO is sceptical of AI ROI claims.

Expert reasoning trace.

Claims summarisation — build the number properly. The pilot gives a measured baseline (20 min) and result (12 min), so the value is real, not asserted. Time saved: 8 min × 200,000 = 1,600,000 min = 26,667 hours/yr. At a loaded adjuster cost of, say, $40/hr, gross saving is $1,066,680/yr. Net against the AI cost ($9,000 × 12 = $108,000/yr) plus, say, $50,000 oversight and change: net ≈ $908,680/yr. But I flag the honesty point to the sceptical CFO: this is capacity unless adjusters are redeployed to clear backlog or headcount plans change — I present it as capacity that can be converted to cash or throughput, not as a cash cheque. Even discounted, the multi-year NPV is strongly positive and payback is a matter of weeks. Buy (Marketplace) beats build here because summarisation is non-differentiating. Scale, sourced by buy.
Fraud scoring — resist the sunk cost. The $500,000 is gone whether we continue or not; the only question is whether future spend earns a return. It has missed its target twice, the data is sparse, and there is no path to fix the data this year — feasibility scores a 1 on the grid, which gates to terminate when unfixable. I recommend terminate (or at most a hard-gated pause with a specific data-acquisition condition and a kill date), and I say to the board plainly that continuing because of the $500,000 already spent would be the sunk-cost fallacy.
Competitor chatbot — refuse the framing. “Because our competitor launched one” is not an outcome. Before any spend I require a defined use case, the metric it moves, and a baseline. If a real outcome exists (say deflecting 30% of a known 500,000 low-complexity contacts), it becomes a measurable case; if not, it is a wish and the answer is no. Pause pending a defined outcome and baseline — do not fund a metric-free initiative just to match a rival.
Give the CFO what earns trust. For the one initiative with a measured baseline I present net (not gross) benefit, an explicit capacity-versus-cash caveat, a payback figure and a discounted multi-year NPV, plus the leading indicators (adoption, minutes saved per claim in week one) we will watch to catch decay early. That is the story a sceptical CFO funds.

Exam-correct recommendation: scale claims summarisation via a Marketplace buy with a net, honestly-caveated ROI; terminate (or hard-gate) the sunk-cost fraud model; and refuse to fund the competitor chatbot until it has a defined outcome and baseline. Reject counting gross savings as cash, reject continuing fraud scoring because of the $500,000 spent, and reject launching the chatbot on competitive envy alone.

Exam traps in this domain

Trap	Why it is tempting	The discriminator
Buy the impressive tool now	Demos sell; action feels decisive	No outcome and metric first means no case; reverse the order
Report the ROI after launch from memory	Baselines feel like delay	Value needs a before-state; establish or reconstruct a baseline
Use the gross saving figure	It makes the case look bigger	Net against all cost; and check whether saved time is cash or capacity
Keep the initiative because “we’ve spent so much”	Loss aversion is powerful	Sunk cost is irrelevant; judge future spend on future return
Build the capability in-house	Building feels like control	Build only what differentiates; buy the commodity
Scale the pilot because everyone likes it	Momentum and enthusiasm	Scale only on a beaten baseline with sustainable unit economics
Match a competitor’s launch to keep up	Fear of falling behind	“Because they did” is not an outcome; require a metric and baseline
Pick the familiar pricing model	Habit and comfort	Match consumption/instance/seat to volume and predictability
Commit to a discount plan immediately	Discounts look like savings	Commitments only pay at high, predictable utilisation
Cut over all at once	It is neat and fast	Phase with a fallback; continuity beats speed

Practice questions

Each item states how many responses to select. Attempt before revealing.

Q1 · A CEO says 'I want us to use GenAI this year' but names no problem. What is the BEST first move for a strategist? (Select one)

A. Buy the leading GenAI platform to show momentum. B. Identify high-impact use cases, map each AI capability to a measurable business outcome, and prioritise from there. C. Fine-tune a model on company data. D. Announce an AI initiative to the market.

Answer: B. Strategy starts from outcomes, so you find use cases and map capability to a metric before choosing technology. Buying a platform (A) or fine-tuning (C) precedes the problem definition, and announcing to the market (D) commits before any value is established.

Q2 · A common, non-differentiating document-extraction need is wanted live in eight weeks. Build would take six months and three new hires. What is the BEST sourcing decision? (Select one)

A. Build, for full control. B. Buy a fit-for-purpose solution (e.g. via AWS Marketplace), because the capability is non-differentiating and timeline and budget favour buying. C. Delay until you can build. D. Fine-tune a custom model.

Answer: B. You build only what differentiates you; a commodity capability wanted fast is a buy. Building (A) or delaying to build (C) diverts scarce talent from differentiating work, and a custom model (D) is heavier than the need requires.

Q3 · An initiative has spent $500,000, missed its accuracy target twice, and has no path to the extra data it needs this year. What should a strategist recommend? (Select one)

A. Continue; too much has been invested to stop now. B. Terminate (or hard-gate with a kill date), because sunk cost is irrelevant and there is no fixable path to the required feasibility. C. Double the budget to push it over the line. D. Scale it to production to recoup the spend.

Answer: B. The $500,000 is gone regardless, so the decision rests on future return, which is negative without a data path. Continuing (A) and doubling down (C) are the sunk-cost fallacy, and scaling a failing model (D) multiplies the problem.

Q4 · Leadership wants the ROI of an assistant that was deployed three months ago, but nobody measured the before-state. What is the BEST response? (Select one)

A. Estimate the ROI from memory of how things felt. B. Reconstruct a baseline from historical records or set up a control group / holdout now to compare, then compute the delta. C. Report that ROI cannot be known and drop it. D. Assume 20% improvement as a standard figure.

Answer: B. A baseline can be reconstructed from historical data or approximated with a control group so a measured delta is possible. Guessing from memory (A) or a standard figure (D) is not evidence, and abandoning measurement (C) forfeits the value story.

Q5 · A pilot cut handle time from 12 to 9 minutes across 100,000 tickets/yr at $30/hr loaded cost; AI plus oversight costs $40,000/yr with a $60,000 one-off. What is the approximate Year-1 ROI? (Select one)

A. About 25%. B. About 83%. C. About 150%. D. Negative.

Answer: B. 3 min × 100,000 = 5,000 hrs × $30 = $150,000 gross; net $110,000 after the $40,000 run cost; Year-1 ROI = (110,000 − 60,000)/60,000 ≈ 83%. The other figures do not follow from netting the run cost and the one-off implementation correctly.

Q6 · Which are LEADING indicators you would watch to steer an AI support pilot early? (Select two)

A. Week-one adoption rate among agents. B. First-month deflection rate trend. C. Annual cost saved (reported at year end). D. Year-end CSAT. E. Total revenue for the fiscal year.

Answer: A and B. Early adoption and deflection predict whether the initiative will work while you can still adjust it. Annual cost saved (C), year-end CSAT (D) and fiscal revenue (E) are lagging measures that confirm outcomes only after the money is spent.

Q7 · A workflow has variable, hard-to-predict volume and may drop to near zero some months. Which pricing structure fits BEST? (Select one)

A. Instance-based provisioned compute running 24/7. B. Consumption-based pricing, because cost scales with usage and falls when volume falls. C. A large upfront commitment plan. D. Seat-based pricing for every employee.

Answer: B. Consumption pricing matches variable volume and scales toward zero when idle. Always-on instances (A) and a big commitment (C) waste money at low utilisation, and per-seat pricing (D) is unrelated to a volume-driven workflow.

Q8 · A CMO wants an AI chatbot 'because a competitor launched one', with no metric. What should the strategist require BEFORE funding it? (Select one)

A. Nothing; matching competitors is reason enough. B. A defined use case, the metric it will move, and a baseline to measure against. C. The largest available model. D. A press release first.

Answer: B. A defensible initiative needs an outcome, a metric and a baseline; competitive envy is not a metric. Matching competitors blindly (A), buying the biggest model (C) or announcing first (D) all commit spend before value is defined.

Q9 · An initiative is promising but blocked because legal has not yet cleared data access. Which portfolio decision is MOST appropriate? (Select one)

A. Terminate it. B. Pause it with a revisit date, because the blocker is fixable and the value is real. C. Scale it immediately. D. Ignore the legal review and proceed.

Answer: B. A fixable blocker on a valuable initiative calls for a pause with a revisit date, not termination. Terminating (A) discards real value over a temporary obstacle, scaling now (C) ignores the unresolved blocker, and bypassing legal (D) creates a compliance risk.

Q10 · 70% of 1,000,000 monthly requests are simple. On-demand is $0.010 each; a smaller model handles simple ones at $0.003. What lever reduces cost and by roughly how much? (Select one)

A. Commitment discount; about 10%. B. Model routing simple requests to the cheaper model; about half. C. Batch processing; no change. D. Larger context windows; about 5%.

Answer: B. Routing 700,000 simple requests to $0.003 ($2,100) and 300,000 to $0.010 ($3,000) totals $5,100 versus $10,000 — roughly 49% cheaper. A commitment (A) is unrelated to routing, batch (C) is a different lever, and bigger context windows (D) would raise, not cut, cost.

Q11 · Which tool pairing does the exam reward for cost planning versus cost tracking? (Select one)

A. Cost Explorer to forecast, Pricing Calculator to see actuals. B. AWS Pricing Calculator to forecast before committing, AWS Cost Explorer to track actuals, trends and anomalies afterwards. C. Marketplace for both. D. Neither is needed if you have a budget.

Answer: B. The Pricing Calculator produces forward estimates and Cost Explorer reports actuals and anomalies — the two are complementary. Option A reverses their roles, Marketplace (C) is for sourcing not cost tracking, and a static budget (D) does not forecast or detect anomalies.

Q12 · A team plans to switch a live process to a new AI platform by cutting over everything on one date. What is the MOST important improvement? (Select one)

A. Cut over faster to reduce disruption. B. Phase the rollout with a fallback path to protect business continuity. C. Skip data-readiness checks to save time. D. Remove the old process immediately.

Answer: B. Continuity requires a phased rollout and a fallback so a failure does not halt the process. Cutting over faster (A) increases risk, skipping data checks (C) invites garbage-in at scale, and removing the old process at once (D) destroys the fallback.

Q13 · Applying the value-feasibility-risk grid, initiative C scores Value 2, Feasibility 2 (sparse data, no fix), Risk 2, and has missed its baseline twice. What is the call? (Select one)

A. Scale. B. Terminate, because value and feasibility are low with no fixable path and it has failed its baseline. C. Pause indefinitely with no conditions. D. Increase investment.

Answer: B. Low value, unfixable feasibility and repeated baseline misses gate to termination. Scaling (A) or investing more (D) throws good money after bad, and an open-ended pause (C) is just avoidance without a decision.

Q14 · Which are TANGIBLE benefits suitable for a KPI, as opposed to intangible ones? (Select two)

A. Annual cost reduction in currency. B. Hours of staff time saved per year. C. Improved brand perception. D. Employee morale. E. Customer delight.

Answer: A and B. Cost reduction and time saved are directly measurable in currency or count. Brand perception (C), morale (D) and delight (E) are intangible benefits that must be tracked through proxy metrics.

Q15 · A CFO is sceptical that '5,000 hours saved' equals cash. What is the MOST honest way to present it? (Select one)

A. Report it as $150,000 cash saved outright. B. Present it as capacity that becomes cash or throughput only if staff are redeployed or headcount plans change, and label it accordingly. C. Inflate it to account for future growth. D. Drop the metric because it is not cash.

Answer: B. Time saved is capacity until it is converted through redeployment or headcount decisions, and saying so builds credibility. Booking it as cash (A) or inflating it (C) overstates value, and dropping a real capacity gain (D) understates it.

Q16 · An industry is rapidly making an AI capability table-stakes; your firm has not adopted it. What investment posture is MOST appropriate? (Select one)

A. Wait until the technology matures further. B. Invest to keep pace, because falling behind on a table-stakes capability is the real risk. C. Make one small experiment and stop. D. Build a fully custom version regardless of cost.

Answer: B. When a capability becomes table-stakes, under-investment is the danger and keeping pace is the appropriate posture. Waiting (A) or a single experiment (C) risks falling behind, and an unjustified custom build (D) over-invests in a commodity.

Q17 · Over three years a $60,000 upfront initiative returns $110,000/yr net; at a 10% discount rate, what is the approximate 3-year NPV? (Select one)

A. About $270,000 undiscounted only. B. About $214,000, because discounted PVs are ~$100k, ~$91k and ~$83k less the $60k upfront. C. Negative. D. Exactly $330,000.

Answer: B. Discounting $110,000 at 10% over years 1–3 gives ~$100,000 + ~$90,909 + ~$82,645 = ~$273,554, less the $60,000 upfront ≈ $213,554. The undiscounted view (A) ignores time value, the case is clearly positive not negative (C), and $330,000 (D) ignores both discounting and the upfront cost.

Q18 · When is AI NOT the appropriate solution? (Select two)

A. The decision is deterministic and must be exact and auditable. B. The data needed does not exist and cannot be obtained in time. C. The task is summarising unstructured customer calls. D. The task is extracting fields from varied documents. E. There is tolerance for occasional error and good data.

Answer: A and B. A deterministic must-be-exact decision belongs to a rule, and a use case with no obtainable data cannot support a model. Summarising calls (C), document extraction (D) and error-tolerant tasks with good data (E) are exactly where AI fits.

Q19 · A leader wants to commit to a discounted throughput plan for a workload whose volume is unpredictable and often low. What is the risk? (Select one)

A. None; discounts always save money. B. A commitment only pays off at high, predictable utilisation; low or variable use leaves you paying for capacity you do not use. C. The model will be slower. D. The data will drift faster.

Answer: B. Commitment discounts trade flexibility for a lower rate and only win when utilisation is high and predictable. Discounts do not always save (A), commitments do not change speed (C), and they are unrelated to data drift (D).

Q20 · Which KPI design BEST protects the customer experience while cutting cost per ticket? (Select one)

A. Track only cost per ticket. B. Set cost per ticket as the north-star with deflection as a leading indicator and CSAT as a guardrail that must not drop. C. Track deflection alone. D. Track number of AI requests made.

Answer: B. A guardrail metric such as CSAT stops you ‘winning’ on cost while degrading service. Cost alone (A) or deflection alone (C) can hide a service collapse, and counting AI requests (D) measures activity, not value.

Q21 · How should AI create SUSTAINABLE competitive advantage rather than a temporary one? (Select one)

A. By buying the same model competitors can buy. B. Through proprietary data, deep workflow integration and continuous improvement that rivals cannot easily copy. C. By announcing AI adoption loudly. D. By using the largest available model.

Answer: B. Advantage that lasts comes from assets and integration rivals cannot replicate, not from a purchasable model. A commodity model (A), marketing (C) or model size (D) are all easily matched by competitors.

Q22 · A vendor proposal is the cheapest option but trains on your data and offers no exit clause. How should a strategist weigh it? (Select two)

A. Weigh data-training and residency terms and lock-in, not just headline price. B. Require an exit path and clarity on IP and data use before selecting it. C. Choose it purely because it is cheapest. D. Ignore regulatory constraints since it is a vendor’s responsibility. E. Assume all vendors have identical terms.

Answer: A and B. A build-buy-partner evaluation weighs terms, lock-in and an exit path alongside price. Choosing on price alone (C), ignoring regulatory constraints (D) and assuming identical terms (E) all miss the material risks in the proposal.

Q23 · A team reports only year-end lagging metrics for a pilot and is surprised it 'failed'. What was the MISTAKE? (Select one)

A. They should have used a bigger model. B. They tracked no leading indicators, so they could not detect and fix low adoption while there was still time. C. They measured too many things. D. They should not have set a baseline.

Answer: B. Without leading indicators there is no early signal to steer the pilot, so problems only surface after the spend. Model size (A) is unrelated, measuring more (C) is not the fault, and dropping the baseline (D) would make evaluation worse.

Q24 · A board wants a defensible ROI story for a claims-summarisation pilot that cut reading time from 20 to 12 minutes over 200,000 claims at $40/hr, costing $108,000/yr plus $50,000 change. Which presentation is MOST defensible? (Select two)

A. Net the ~$1.07m gross saving against the ~$158k annual cost and show payback and a discounted multi-year NPV. B. Flag that the saving is capacity unless adjusters are redeployed or headcount changes, and label it as such. C. Report the gross $1.07m as immediate cash profit. D. Present adjectives about efficiency instead of numbers. E. Omit the AI running cost to make the case stronger.

Answer: A and B. A credible case nets against all cost, shows payback and NPV, and states honestly whether the saving is cash or capacity. Reporting gross as cash (C), using adjectives (D) or hiding the running cost (E) destroys credibility with a sceptical board.

Key takeaways

Start from the business outcome and a measurable metric; source the capability (build/buy/partner) only after the outcome is defined, and build only what differentiates you.
Establish a baseline before deployment — or reconstruct one with historical data or a control group; value you did not measure you cannot defend.
Compute ROI on net benefit against all cost including oversight, and be honest about whether time saved is cash or capacity; show payback and a discounted multi-year NPV.
Prioritise on value, feasibility, sustainability and alignment, and use the grid to decide scale, pause or terminate — sunk cost never justifies continuing.
Steer with leading indicators, decide scale with lagging ones; an initiative with only lagging metrics finds out too late.
Match the pricing structure (consumption, instance, seat) to volume and predictability, and use commitment, batch and routing levers only where they genuinely pay; forecast with the Pricing Calculator, verify with Cost Explorer.
AI is not always the answer — a deterministic problem, absent data, or a broken process are reasons to say no.
Match investment level to industry maturity, and build sustainable advantage from proprietary data, workflow integration and continuous improvement, not from a purchasable model.