MBA AI Strategy Concept Card Template

AI Strategy Concept Card: [Your course, case or company]

A concept card for the strategic side of artificial intelligence as an MBA core course and a management team need it: where AI creates value (lower cost, higher quality, new offerings), the "prediction machines" framing, scoring and prioritising a portfolio of use cases, building an AI business case with NPV, ROI and payback, and judging whether data and AI create a lasting advantage (moats and data network effects, and their limits). It is for MBA students, admitted students preparing for the core, and managers who sponsor AI projects. Worked example A is fully solved, Worked example B is half solved and Exercise C is left for you, with a separate answer key. Source basis: Agrawal, Gans and Goldfarb (2018), Hagiu and Wright (2020), Barney (1991), Porter (1985) and standard capital budgeting, written in original words with fictional examples (Corvane Robotics, Northfield Outdoor Co. and Tallow Creek Medical Center). Replace every [bracketed] value and every row marked Example. This is a learning aid, not investment advice.

Document control

Contents

• When to use it
• The idea in plain language
• Formulas and notation
• Assumptions and limits
• How managers use it
• Prioritising a use-case portfolio
• Worked example A (fully solved): Corvane Robotics exception handling
• Moats and data network effects
• Worked example B (half solved): Northfield personalised email offers
• Exercise C (unsolved): Tallow Creek no-show prediction
• Decision diagram
• Connections to other cards
• Common mistakes and misconceptions
• Self-test questions
• My notes and next steps
• Sources and further reading
• Answer key
• Revision history
• Standard and licence note

When to use it

One-sentence rule: fund an AI use case when it improves a decision or task that matters to the business by enough to repay its full cost, including data, integration, run cost and change, and be sceptical of any claim that the advantage will last unless you can show why rivals cannot copy it.

The idea in plain language

Where AI creates value. AI creates value in three ways. It lowers cost by automating or speeding up tasks (drafting replies, classifying documents, routing exceptions). It raises quality by making better or more consistent decisions (forecasting demand, detecting fraud, spotting defects). And it enables new offerings that were not economic before (personalised products at scale, conversational service, new data-based services sold to customers). Map each candidate onto one or more of these three and onto the firm's value chain (Porter 1985, see mba/mba-value-chain-card): which activity changes, and how does the change reach the customer or the cost line?

Prediction machines. Agrawal, Gans and Goldfarb (2018) offer the simplest economic lens. Predictive AI is a drop in the cost of prediction: filling in missing information from the information you have. When the price of something falls, we use more of it, and the value of its complements rises. The complements of prediction are data (to train and run the model), judgment (knowing the payoffs of each outcome and action, which a person or the firm's policy supplies) and action (the ability to act on the prediction). Break every decision into these parts:

A worked illustration. Corvane Robotics, a growth-stage warehouse-robotics company, services a fleet of 1,800 robots. On average 8% break down in a month. Without a prediction, Corvane compares visiting every robot (USD 400 each) with waiting (expected $0.08 \times 3{,}000 = 240$) and rationally waits: USD 240 per robot-month. A model sorts the fleet into a high-risk 10% with a 50% breakdown chance and a low-risk 90% with a 3.33% chance (the two average back to 8%). Now Corvane visits the high-risk robots (USD 400 is less than $0.5 \times 3{,}000 = 1{,}500$) and waits on the rest ($0.0333 \times 3{,}000 = 100$). Expected cost falls to $0.1 \times 400 + 0.9 \times 100 = 130$, so the model is worth USD 110 per robot-month, about USD 2.38m a year for the fleet. A perfect prediction would cost only $0.08 \times 400 = 32$, a value of USD 208. The gap between 110 and 208 is what better data and better models could still earn. Notice that the model has value only because Corvane can act (dispatch a technician) and knows the payoffs (judgment). Without either, the prediction is worth nothing.

Why this matters strategically. Because prediction gets cheaper for everyone, a firm's advantage shifts to the complements it controls: distinctive data, superior judgment embedded in processes, and the ability to act, such as an installed base, a distribution network or a service organisation. That is the link to the resource-based view (Barney 1991, see mba/mba-resource-based-view-card): a resource creates sustained advantage only if it is valuable, rare, hard to imitate and supported by the organisation.

Generative AI and agents. The prediction lens still applies (a language model predicts the next piece of text), but managers usually meet generative AI as a cost of producing content and handling interactions. Its value comes from speeding up tasks and widening what each person can do; its risk comes from fluent errors. Agents add action directly, so the "action" element above is automated, which raises the value of good judgment rules and oversight (see mba/mba-ai-risk-governance-card).

Formulas and notation

Value of a prediction for one decision:

$V_{pred} = C_{no} - C_{with}, \qquad C_{no} = \min_a \mathbb{E}[\text{cost}(a)], \qquad C_{with} = \sum_s \pi_s \min_a \mathbb{E}[\text{cost}(a) \mid s]$

Weighted use-case score for use case $j$ across criteria $k$:

$S_j = \sum_{k=1}^{K} w_k \, x_{jk}, \qquad \sum_k w_k = 1$

Business case. Net cash flow in year $t$ with adoption ramp $a_t$:

$CF_t = a_t B - R, \quad t = 1, \dots, T, \qquad CF_0 = -I$

$NPV = \sum_{t=0}^{T} \frac{CF_t}{(1+r)^t}, \qquad ROI = \frac{\sum_t a_t B - (I + T R)}{I + T R}$

Payback is the time until cumulative cash flow turns positive: $PB = (n-1) + \dfrac{-\text{Cum}_{n-1}}{CF_n}$, where $n$ is the first year with positive cumulative cash flow. Discounted payback uses discounted flows.

Risk-adjusted (expected) NPV with probability of technical success $p$:

$\mathbb{E}[NPV] = p \cdot NPV_{success} + (1-p) \cdot NPV_{fail}$

Assumptions and limits

• The benefit is realised. Hours freed are not cash saved unless headcount, overtime or contractor spend falls, or growth is absorbed without hiring. Write the realisation path next to every labour benefit.
• The model performs in production as in the pilot. Data drift, new products and changed behaviour erode accuracy; the run cost must include monitoring and retraining.
• Adoption follows the ramp. People must trust and use the system; change costs are real (see mba/mba-change-management-card).
• Scores are judgments. A weighted score is only as good as its weights and the honesty of the scoring; test how the ranking changes when the weights change.
• ROI ignores timing and scale. It is a ratio without discounting; use it to communicate, NPV to decide.
• Vendor prices change. Usage-based pricing of generative AI can move in either direction; model the run cost with a range (see ai-data/inference-cost-model).
• The horizon is short. AI systems and their economics change fast; three to five years with no terminal value is a common and conservative choice.

How managers use it

In practice an AI strategy is a managed portfolio. A typical cycle: business units submit use cases on an intake form (ai-data/ai-use-case-intake); a cross-functional group scores them; the top candidates get a short discovery phase to test data readiness; those that pass get a pilot with pre-agreed evidence gates; and only pilots that meet the gates get a full business case and scaling budget. Committees look for: the decision or task that changes, the baseline, the realisation path for every benefit, the run cost, the risk classification and owner, and the kill criteria. Senior teams also ask a strategic question that no single business case answers: which capabilities (data platforms, skills, governance) should we build once and reuse across many use cases? That is where the "build, buy or partner" choice from mba/mba-ai-for-managers-guide comes in.

Prioritising a use-case portfolio

Score each use case from 1 (poor) to 5 (strong) on five criteria. For risk, 5 means low risk. Weights Example: value 30%, feasibility 20%, data readiness 20%, risk 15%, time to value 15%.

Excel, with weights in B2:F2 and scores in B4:F8: =SUMPRODUCT($B$2:$F$2, B4:F4); rank with =RANK.EQ(G4, $G$4:$G$8). UC1 and UC2 tie at 3.40 (RANK.EQ gives both rank 3); break ties on value, so UC1 takes third place and UC2 fourth.

Sensitivity. Move 10 points of weight from time to value to value (value 40%, time to value 5%). Scores become UC1 3.70, UC4 3.60, UC3 3.45, UC2 3.30, UC5 2.85: the demand forecast moves from third to first and the reply assistant stays fourth. The ranking depends on how much the firm values quick wins against large bets, which is a strategic choice, not a calculation. In the Northfield mini-case in mba/mba-ai-for-managers-guide the demand forecast also has by far the largest NPV, which is why a scoring grid should screen and a business case should decide.

Portfolio view. Plot value against feasibility. Keep two or three quick wins (high feasibility, modest value) to build skills and trust, one or two larger bets (high value, lower feasibility) with staged funding, and drop the rest.

Worked example A (fully solved): Corvane Robotics exception handling

Situation (fictional). Corvane's warehouse robots stop when they cannot complete a pick (a crushed box, an item in the wrong bin). A remote operator then resolves the exception. Across 1,800 robots there are 6 exceptions per robot per day, each taking 1.5 operator-minutes at USD 45 an hour fully loaded. A vision model that resolves routine exceptions automatically is expected to cut operator interventions by 40% at full adoption and to reduce robot downtime, worth USD 0.60m a year in contribution from customer service-level credits avoided. The build costs USD 2.0m (labelling, model development, integration); compute, monitoring and retraining cost USD 0.45m a year. Adoption ramps 40%, 80% and then 100% from year 3. Corvane's hurdle rate for projects of this risk is 12%; horizon five years, no terminal value. Operators freed are redeployed to fleet growth, so Corvane hires fewer new operators as the fleet expands (the realisation path).

Step 1. Full-adoption benefit.

Step 2. Cash flows (USD m). Year 0 in B2, years 1 to 5 in C2:G2, rate 12% in B1.

Step 3. Measures.

Step 4. Risk adjustment. The team estimates a 70% chance that the model reaches the 40% cut. If it does not, Corvane learns this by the end of year 1, earns half of the year-1 benefit, pays the year-1 run cost and stops: cash flows $-2.0$ and $0.0248$, NPV USD -1.9779m. Expected NPV $= 0.7 \times 3.2850 + 0.3 \times (-1.9779) =$ USD 1.7061m. The project is still worth doing, but a staged design (a USD 0.5m proof of concept before the full build) would cut the loss on the failure path; see mba/mba-decision-trees-card and the real options logic in mba/mba-options-real-options-card.

Step 5. Sensitivity on the intervention cut.

The downtime benefit is scaled in proportion to the cut. The project breaks even just above a 20% cut, so the key question for the pilot is whether the model can resolve at least a quarter of exceptions safely.

Conclusion. At a 12% hurdle rate the exception-handling model has an NPV of USD 3.29m if it reaches a 40% cut and an expected NPV of USD 1.71m after allowing a 30% chance of failure. Recommend a staged build with a proof of concept whose gate is a safe automatic resolution rate of at least 25% of exceptions.

Moats and data network effects

A data network effect exists when more use of a product generates more data, the data improves the product, and the better product attracts more use. It is often claimed and less often real. Hagiu and Wright (2020) argue that data creates a lasting advantage only under specific conditions. Use these five tests:

Two more tests matter for generative AI. Foundation model commoditisation: if the capability comes from a model any rival can license, it is not a moat; the moat, if any, lies in proprietary data, workflow integration and distribution. Customer lock-in: whether switching costs come from the AI itself or from the surrounding system (integrations, trained users, stored history).

Example: Corvane's exception data comes from its own installed base of robots, is labelled as a by-product of operations, and keeps improving the model as new warehouse layouts appear. It passes the proprietary test and partly passes the diminishing-returns test. It fails the imitability test only if a rival with a larger fleet collects more of the same data. So the moat is real but bounded by fleet size, which links AI strategy to growth strategy (see mba/mba-growth-strategy-card). Northfield's product copy, by contrast, uses a licensed model and public product specifications: useful, but no moat at all.

Worked example B (half solved): Northfield personalised email offers

Situation (fictional). Northfield Outdoor Co. sends 60 million marketing emails a year with a 0.40% conversion rate. A recommendation model would personalise the products in each email. A randomised test suggests a 10% relative uplift in conversion. Average order value is USD 85 and contribution margin 42%. Build cost USD 0.9m, run cost USD 0.35m a year, adoption 50% in year 1 and 100% after. Hurdle rate 10%, horizon three years.

Questions:

1. Compute $B$ and the year 2 and 3 cash flows.
2. Compute NPV at 10% (=NPV(10%, C2:E2) + B2) and IRR.
3. Compute the three-year ROI and the simple payback.
4. ROI is positive. Is NPV? Explain the difference in one sentence.
5. What relative uplift is needed for NPV to be zero? (Use Goal Seek.)
6. What would you ask the marketing team before approving?

Exercise C (unsolved): Tallow Creek no-show prediction

Situation (fictional). Tallow Creek Medical Center, a regional hospital, has 120,000 outpatient appointment slots a year and a 12% no-show rate. A model predicting no-shows would target reminder calls and allow careful overbooking in high-risk slots. Operations expects to recover 25% of no-show slots at full adoption, each worth USD 140 in contribution. Build cost USD 0.6m, run cost USD 0.25m a year, adoption 50% in year 1 and 100% after. Hurdle rate 8%, horizon four years.

1. Compute the slots recovered and the yearly benefit at full adoption.
2. Lay out the cash flows and compute NPV, IRR, four-year ROI and simple payback, with Excel formulas.
3. Is the project worth doing on these numbers? How sensitive is the answer to the 25% recovery assumption?
4. Break the decision into prediction, judgment, action and outcome. Whose judgment sets the overbooking rule?
5. Name two non-financial risks (patient experience, fairness across patient groups) and how you would monitor them. Link your answer to mba/mba-ai-decision-making-card and mba/mba-ai-risk-governance-card.

Decision diagram

The diagram shows the order in which a portfolio review should ask its questions: framing and data first, scoring second, risk screen and pilot third, and the full business case only for pilots that meet their evidence gates.

Connections to other cards

Common mistakes and misconceptions

1. "AI strategy" as a list of tools. Strategy names the decisions and tasks to improve and the capabilities to build, not the vendors to buy.
2. Counting freed hours as savings. Without a realisation path the labour benefit is zero in cash terms.
3. Leaving out run costs. Compute, licences, monitoring, retraining and knowledge upkeep recur every year.
4. Ignoring the data cost. Cleaning, joining and labelling data are often the largest one-off cost.
5. Deciding on ROI. ROI ignores timing; a positive ROI can coexist with a negative NPV, as in Worked example B.
6. Choosing weights after scoring. It turns the grid into a justification for a choice already made.
7. Assuming a data moat. Most data advantages flatten, depreciate or can be bought; apply the five tests.
8. Confusing a licensed model with a capability. If every rival can license the same model, the advantage lies elsewhere or nowhere.
9. No kill criteria. Pilots without evidence gates rarely end, and they absorb scarce talent.
10. Forgetting judgment. A perfect prediction has no value if nobody has decided what to do with it.

Self-test questions

1. Name the three ways AI creates value and give one example of each.
2. In the prediction machines framing, what are the complements of prediction, and why does their value rise?
3. In the Corvane dispatch illustration, why is the model worth nothing if Corvane cannot dispatch technicians?
4. Two criteria are weighted 50% each. Use case X scores 4 and 2, Y scores 3 and 3. Which ranks first, and what does the tie-break depend on?
5. A project costs USD 2.0m in total and returns USD 3.0m in total benefits over three years. What is its ROI, and what does ROI not tell you?
6. Why does Worked example A use expected NPV, and how does a staged proof of concept change it?
7. State the five tests of a data moat.
8. Why is a licensed foundation model not a moat on its own?
9. A pilot frees 10 operator hours a day. Under what conditions is this a cash saving?
10. Give two reasons to keep quick wins in an AI portfolio even if their NPV is small.

My notes and next steps

Sources and further reading

• Agrawal, A., Gans, J., and Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.
• Hagiu, A., and Wright, J. (2020). When data creates competitive advantage. Harvard Business Review, 98(1).
• Barney, J. (1991). Firm resources and sustained competitive advantage. Journal of Management, 17(1), 99-120.
• Porter, M. E. (1985). Competitive Advantage: Creating and Sustaining Superior Performance. Free Press.
• Shapiro, C., and Varian, H. R. (1999). Information Rules: A Strategic Guide to the Network Economy. Harvard Business School Press.
• Brynjolfsson, E., Rock, D., and Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333-372.
• Textbooks for further reading (named only): Iansiti and Lakhani, Competing in the Age of AI; Agrawal, Gans and Goldfarb, Power and Prediction; Berk and DeMarzo, Corporate Finance (capital budgeting chapters).
• Open resource: OpenStax, Principles of Finance (CC BY-NC-SA 4.0), chapters on capital budgeting.

Answer key

Worked example B. (1) $B = 24{,}000 \times 85 \times 0.42 =$ USD 0.8568m. Year 1: $0.5 \times 0.8568 - 0.35 = 0.0784$; years 2 and 3: $0.8568 - 0.35 = 0.5068$. Cumulative: -0.3148 after year 2, 0.1920 after year 3. (2) NPV at 10% = USD -0.0291m; IRR = 8.48%. (3) Benefits $0.4284 + 0.8568 + 0.8568 = 2.1420$; costs $0.9 + 3 \times 0.35 = 1.95$; ROI = 9.8%. Payback = $2 + 0.3148/0.5068 = 2.62$ years. (4) NPV is slightly negative: ROI ignores the time value of money, and the return of 8.48% is below the 10% hurdle. (5) A relative uplift of 10.17% gives NPV of zero, so the case sits on a knife edge. (6) Ask how the test was randomised and how long it ran (novelty), whether the uplift holds outside the test segment, whether it cannibalises orders that would have happened anyway, the unsubscribe and complaint rates, and whether a longer horizon or lower run cost is realistic.

Exercise C. (1) Slots recovered $120{,}000 \times 12\% \times 25\% = 3{,}600$; benefit $3{,}600 \times 140 =$ USD 0.5040m. (2) Cash flows -0.6000, 0.0020, 0.2540, 0.2540, 0.2540; NPV at 8% = USD 0.0079m (=NPV(8%, C2:F2) + B2); IRR = 8.48%; ROI $= (1.764 - 1.6)/1.6 = 10.2\%$ on benefits $0.252 + 3 \times 0.504 = 1.764$ and costs $0.6 + 4 \times 0.25 = 1.6$; payback $3 + 0.090/0.254 = 3.35$ years. (3) Barely: NPV is close to zero and falls below it if recovery is a little under 25%, so the decision turns on evidence from a pilot and on non-financial value (access, waiting lists). (4) Prediction: will this patient miss this slot? Judgment: the cost of an empty slot against the cost of an overbooked clinic, and the fairness rules; set by clinical and operations leadership, not by the model team. Action: reminder call, overbooking. Outcome: attendance records. (5) Overbooking could lengthen waits for patients who attend; the model could flag some patient groups more often (for example by neighbourhood or transport access), shifting the burden of overbooking onto them. Monitor waiting times and flag rates by group, and offer help (transport, reminders) rather than only overbooking.

Self-test.

1. Lower cost (automated reply drafting), higher quality (better demand forecasts), new offerings (personalised services at scale).
2. Data, judgment and action. When prediction becomes cheaper, firms use more of it, so the inputs that make predictions useful become scarcer relative to demand and more valuable.
3. The value of a prediction comes from changing an action. If no action can change, the best action is the same with or without the prediction, so $C_{with} = C_{no}$.
4. Both score 3.0. Break ties on the criterion that matters most strategically (often value) or on evidence quality; the tie shows the grid alone cannot decide.
5. ROI = $(3.0 - 2.0)/2.0 = 50\%$. It ignores timing, scale and risk.
6. Because the model may not reach its target, and failure has a cost. A proof of concept spends less before the uncertainty is resolved, so the failure-path loss shrinks and expected NPV rises.
7. Value added by data, diminishing returns, depreciation, proprietary access, imitability of improvements.
8. Every rival can license the same model, so it creates no rare or hard-to-imitate resource; advantage must come from proprietary data, integration, distribution or judgment.
9. When the firm reduces paid hours (fewer hires, less overtime or contractor spend) or absorbs growth without hiring.
10. They build skills, data infrastructure and trust cheaply, and they produce evidence about adoption that improves the estimates for larger bets.

Revision history

Standard and licence note

Based on the economic and strategy works listed above (Agrawal, Gans and Goldfarb 2018; Hagiu and Wright 2020; Barney 1991; Porter 1985) and on standard capital budgeting, checked in October 2026. The text, examples, numbers and questions are original; no textbook or case text is copied or paraphrased, and Corvane Robotics, Northfield Outdoor Co. and Tallow Creek Medical Center are fictional. The template is independent and not affiliated with or endorsed by any business school, publisher or software vendor. Template licence: MIT.