MBA AI for Managers Course Guide Template

AI for Managers Course Guide: [Your programme, term or team]

A course guide for the "AI for managers" course that leading MBA programmes are adding to the core, and for working managers who have to decide where artificial intelligence fits in their business. It explains what a manager needs to know about machine learning, generative AI and AI agents at concept level, treats data as an asset, shows how to build a portfolio of use cases and choose between building, buying and partnering, summarises the best field evidence on productivity, and covers change, workforce, ethics and regulation. It gives a 12-session roadmap, a key concepts table that links the three AI concept cards, an original mini-case set at the fictional Northfield Outdoor Co. with a teaching note, a grading rubric, an exam plan and a self-test with an answer key. Source basis: Agrawal, Gans and Goldfarb (2018), Brynjolfsson, Li and Raymond (2023, 2025), Dell'Acqua et al. (2023, 2026), NIST AI RMF 1.0 (2023), ISO/IEC 42001:2023 and the EU AI Act, written up in original words. Replace every [bracketed] value and every row marked Example. This is a learning aid, not legal, investment or employment advice.

Document control

Contents

• What the course is for
• Prerequisites
• What managers need to know
• Session roadmap
• Key concepts table
• Mini-case: Northfield Outdoor Co. and the AI shortlist
• Typical assignments and how they are graded
• Exam preparation plan
• Common struggles
• Reading list
• Self-test
• My notes and next steps
• Sources and further reading
• Answer key
• Revision history
• Standard and licence note

What the course is for

AI has moved from a specialist topic to a general management question. Most decisions about AI in a firm are not about algorithms: they are about which problems are worth solving, which data the firm owns and can use lawfully, whether to build, buy or partner, how work and jobs change, how to measure whether the system helps, and who is accountable when it fails. Those are the questions this course trains you to answer. The aim is not to make you a data scientist. It is to make you the person in the room who can frame the business problem precisely, ask the technical team the right questions, read an evaluation honestly, write a business case that survives a finance review, and set up governance that a regulator, an auditor or a journalist would accept.

As of October 2026, several leading programmes have moved AI into the core rather than leaving it to electives. Yale SOM's own curriculum pages describe a required, hands-on AI course in the fall of the first year together with AI content in existing core courses; other schools integrate AI into existing courses or offer majors and electives. Formats change every year: check each school's current curriculum page, and see mba/mba-program-comparison for how to compare programmes. Nothing in this guide is any school's course material.

The course builds the competencies that business school accreditation frameworks and the docs for this library group as follows:

Prerequisites

You do not need to code. You do need four things, each covered by an existing template:

1. Probability and basic statistics: conditional probability, base rates, the normal distribution, a confidence interval and a hypothesis test for two proportions. Refresh with mba/pre-mba-quant-refresher, academic/stats-probability-basics and academic/stats-proportion-tests.
2. Time value of money and NPV: an AI use case is an investment. Refresh with mba/pre-mba-finance-primer and mba/mba-npv-irr-card.
3. Spreadsheet fluency: NPV, IRR, SUMPRODUCT, NORM.S.DIST, data tables. Drill with mba/pre-mba-excel-modeling-drills.
4. Strategy vocabulary: value creation and capture, competitive advantage, resources and capabilities. Read mba/mba-strategy-guide and mba/mba-resource-based-view-card in parallel with sessions 3 and 4.

What managers need to know

1. Machine learning, generative AI and agents at concept level

Machine learning (ML) is software that learns a mapping from examples instead of following rules written by a programmer. Most business value from ML so far comes from supervised learning: the system sees many historical cases with a known outcome (a transaction that turned out to be fraud, a customer who churned, the demand that materialised) and learns to predict that outcome for new cases. The output is a prediction, usually a score or a probability, and its quality can be measured against outcomes that arrive later. Unsupervised methods find structure (segments, anomalies) without labelled outcomes, and reinforcement learning learns actions from rewards.

Generative AI uses very large models trained on broad data (foundation models, including large language models) to produce new content: text, code, images, audio, video. A language model predicts the next piece of text given what came before; scaled up and tuned, that simple objective yields systems that draft, summarise, translate, classify and answer questions. Two properties matter to managers. First, the outputs are fluent but not guaranteed to be true: the model can produce confident, wrong statements (NIST calls this confabulation). Second, the same general model can be pointed at many tasks with instructions and examples, so the cost of trying a new use case is low while the cost of making it reliable can be high.

AI agents are systems in which a model plans and takes actions through tools (search, databases, email, payments, code execution) over several steps towards a goal, with varying degrees of human approval. Agents move AI from producing advice to taking actions, which raises the stakes for permissions, monitoring and accountability. A manager should ask of any agent: what can it do without a human, what can it never do, how is each action logged, and how is it stopped.

2. Data as an asset

Data becomes an asset when it is relevant to a decision, accurate, available when the decision is made, lawfully usable for that purpose, and hard for rivals to reproduce. Most firms overrate the first and underrate the last three. Ask of every data source: who owns it, what consent or contract governs its use, how often it is refreshed, what its error rate is, and whether a competitor could buy or collect something equivalent. The cost of cleaning, labelling and integrating data is usually the largest one-off cost in a predictive project, and the Northfield case below makes this visible. Whether data creates a lasting advantage is covered in mba/mba-ai-strategy-card.

3. A portfolio of use cases

Treat AI as a portfolio, not a single bet. Map candidate use cases on value and feasibility, keep a balance of quick wins that build capability and a few larger bets that could change the business, and kill projects that miss agreed evidence gates. The scoring method, with weights and a worked example, is in mba/mba-ai-strategy-card; the intake form for each use case is ai-data/ai-use-case-intake, and ai-data/ai-opportunity-canvas helps frame one opportunity.

4. Build, buy or partner

Most firms end with a mix: buy for generic productivity, partner or build where data and process are distinctive. Check vendor risk with compliance/ai-vendor-risk-questionnaire.

5. What the productivity evidence says

The strongest evidence comes from field experiments and staggered roll-outs in real firms. Brynjolfsson, Li and Raymond studied 5,179 customer support agents at one firm who gained access to a generative AI assistant. Issues resolved per hour rose by 14% on average, by 34% for novice and low-skilled agents, with little effect for the most experienced (NBER working paper 2023; published in the Quarterly Journal of Economics in 2025). Dell'Acqua and co-authors ran a pre-registered experiment with 758 consultants at Boston Consulting Group (working paper 2023; published in Organization Science in 2026). On tasks inside what they called the "jagged technological frontier", consultants with AI completed more tasks, faster and with markedly higher quality; on a task chosen to lie outside it, consultants with AI were less likely to reach the correct answer than those without. Noy and Zhang (2023) found faster and better professional writing with a chatbot in an online experiment.

Three lessons for managers follow. Gains are real but uneven across tasks and people, so measure your own. The less experienced often gain most, which changes training, hiring and career ladders. And performance falls off a cliff at the edge of the model's competence, where people trust it most easily; that is why human oversight and evaluation (mba/mba-ai-decision-making-card) matter. Firm-level gains also take time: Brynjolfsson, Rock and Syverson (2021) describe a productivity J-curve in which intangible investment in process and skills comes before measured output.

6. Change and the workforce

AI changes tasks before it changes jobs. Break each role into tasks, mark which are automated, augmented or unchanged, and redesign the role around what remains. Decide early how freed capacity will be used (growth, quality, lower hiring, redeployment) and say so to employees; silence breeds fear and quiet resistance. Build AI literacy for everyone who uses or oversees AI (see ai-data/ai-literacy-plan). Kotter's (1995) eight steps and the material in mba/mba-change-management-card apply directly.

7. Ethics

Ask four questions of every use case: is it fair to the people it affects, can its decisions be explained to them, is someone clearly accountable, and was the data obtained and used with the consent or legal basis it needs? Add a fifth for generative AI: could it deceive people about what is machine-made? Ethics also includes the decision not to use AI where the stakes, the lack of good data or the need for human judgment make it inappropriate. See mba/mba-ethics-guide.

8. Regulation and standards

As of October 2026 the main reference points are the EU AI Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744, the "Digital Omnibus on AI"), which sets risk-based obligations with staged application dates; the voluntary NIST AI Risk Management Framework 1.0 (January 2023) and its Generative AI Profile, NIST AI 600-1 (July 2024); and the certifiable management system standard ISO/IEC 42001:2023. Other jurisdictions have their own rules. The details, with a risk register example, are in mba/mba-ai-risk-governance-card. Check the consolidated texts and the official pages before you rely on any date.

Session roadmap

Key concepts table

Mini-case: Northfield Outdoor Co. and the AI shortlist

Situation

This case is fictional and was written for this template. Any resemblance to a real company is accidental.

Priya Raman had been chief operating officer of Northfield Outdoor Co. for eight months when the chief executive, Tomasz Kowal, asked her to bring "an AI plan" to the October board meeting. Northfield designs and sells outdoor gear (tents, packs, footwear and technical clothing) through 85 stores and its own website, which now generates 38% of revenue. The company is profitable but growth has slowed, a competitor has launched an AI shopping assistant, and two board members have asked why Northfield seems to be "behind".

Priya's team had collected more than forty ideas from across the business. After a first screen, three remained. The first was a machine learning demand forecast at store and SKU level, replacing the spreadsheet forecasts that the planning team built each season. Northfield's planners were good, but seasonal products were often overstocked in some stores and sold out in others, and end-of-season markdowns had been rising. The data science lead, Daniel Osei, warned that three years of point-of-sale data sat in two systems with different product codes, and that cleaning and joining them would take most of the first six months.

The second was a generative AI assistant for the contact centre. Northfield employed 220 agents who answered questions on orders, returns, sizing and product care by phone, chat and email. A vendor offered an assistant that suggested replies and pulled the relevant policy and product information into the agent's screen. Northfield had run an eight-week pilot with 60 agents using the assistant and 60 comparable agents without it, assigned by team. The contact-centre manager, Ana Lucía Ferreira, was enthusiastic: "My new starters are performing like six-month agents." Her most experienced agents were less impressed, and two of them had found the assistant suggesting a returns window that had changed in the spring.

The third was generative product copy: drafting product descriptions, care instructions and translations for the website from technical specifications, with a copywriter editing each draft. It was cheap and the e-commerce team wanted it immediately.

The chief financial officer, Hamid Sultani, set two rules for the board paper. Every use case had to show a four-year NPV at Northfield's 10% hurdle rate, and any claimed labour saving had to say how the saving would actually be realised. "Hours freed are not money saved," he said, "unless we hire fewer people, grow without hiring, or do something more valuable with the time."

The general counsel, Mei Lin Zhou, added a third rule: each use case had to be screened for legal and reputational risk before the board saw it, including the EU AI Act for Northfield's stores and customers in the EU.

Priya had two weeks. She knew the board would ask three questions: which project first, what could go wrong, and how the company would know whether it was working.

Exhibits

Exhibit 1. Northfield Outdoor Co. key figures (fictional, last financial year)

Exhibit 2. Contact-centre pilot, eight weeks, resolutions per agent hour (fictional)

Pilot notes: 3.1% of assisted contacts were flagged in quality review for a policy error in the suggested reply, mostly outdated returns rules; the error rate in the comparison group was 1.4%.

Exhibit 3. Cost and benefit estimates by use case (fictional, USD m)

All yearly benefits reach 50% of the full level in year 1 and 100% from year 2.

Exhibit 4. First legal and risk screen (fictional, general counsel's notes)

Discussion questions

1. Compute the four-year NPV of each use case at 10%. Which ranks first, and is NPV alone the right basis for the ranking?
2. The pilot showed an 11.5% average gain. What gain should Northfield use for its own workforce, and why does the tenure mix matter?
3. What productivity gain would U2 need to break even? What other benefits or costs are missing from Exhibit 3?
4. What does the 3.1% policy-error rate tell you about the assistant, and what would you change before scaling it?
5. Was the pilot design good enough to support a board decision? How would you improve it?
6. Draft Priya's one-sentence recommendation and the three metrics the board should track.

Teaching note

Read this only after you have answered the questions.

Question 1. Cash flows in USD m, years 0 to 4, with the Excel layout year 0 in B2 and years 1 to 4 in C2:F2:

Excel: =NPV(10%, C2:F2) + B2 and =IRR(B2:F2). U1 year 1 is $0.5 \times (1.8 + 1.2) - 0.8 + 4.5 = 5.2$, where 1.8 is 20% of the USD 9.0m inventory reduction. U1 dominates, but most of its value is the one-off working-capital release of USD 9.0m, which depends on planners trusting the forecast enough to cut stock. Its IRR is high because the release arrives early; quote NPV, not IRR, to the board. NPV is the right first lens but not the only one: U3 is small but fast and builds generative AI capability with low risk, and U2's main effects (quality, customer experience, agent onboarding) are not in the cash flows.

Question 2. Weight the tenure bands by Northfield's workforce mix: without the assistant $0.3 \times 1.90 + 0.4 \times 2.40 + 0.3 \times 2.85 = 2.385$, with it $2.661$, a gain of 11.6% (=SUMPRODUCT(mix, with)/SUMPRODUCT(mix, without) - 1). It is close to the simple average here only because the mix is nearly balanced; a contact centre staffed mostly by experienced agents would see a gain nearer 2.5%. At 11.6% the same contact volume needs $220 \times (1 - 1/1.116) = 22.82$ fewer agent equivalents, worth USD 1.187m a year at full adoption.

Question 3. U2 breaks even at a gain of 14.6% (found with Goal Seek on the NPV cell). At 5% its NPV is USD -2.474m; at 20% it is USD 1.224m. Missing items include refunds and complaints from wrong answers, quality-review effort, faster onboarding (lower training cost), effect on agent attrition, customer satisfaction and the vendor's pricing changes. The CFO's rule matters: the saving is real only if Northfield hires fewer agents as people leave, which needs a hiring plan, not just a pilot.

Question 4. The assistant more than doubled the policy-error rate (3.1% against 1.4%). The cause is stale retrieval content, not the model: the assistant is only as current as the policy library it draws on. Before scaling, assign an owner for the knowledge base, add a review date to every policy article, show the source next to each suggestion, and train agents to check high-stakes answers (returns, safety, refunds). This is the jagged frontier at small scale: the assistant is strong on routine questions and weak on recently changed rules, where agents trust it most.

Question 5. Assigning by team risks confounding (a stronger supervisor or a different queue). Eight weeks may capture novelty effects. Improve it with randomisation at agent level within teams, a longer run, pre-registered metrics (resolutions per hour, quality errors, customer satisfaction, repeat contacts) and a power calculation (see mba/mba-ai-decision-making-card and academic/stats-power-sample-size).

Question 6. Example: "Start the demand forecast now, with data clean-up as phase one; launch product copy with human editing as a low-risk capability builder; fix the knowledge base and re-run a randomised assistant pilot before any scale decision." Track forecast accuracy against planner forecasts by category, inventory and markdowns against plan, and the assistant's error rate and resolutions per hour in the new pilot.

Typical assignments and how they are graded

Common assignments are an individual AI use-case memo, a team business case with a pitch, a critique of a vendor or academic evaluation, a governance or policy brief, and a final exam with a short case.

Exam preparation plan

See mba/mba-exam-prep for the general method.

Common struggles

• Starting from the tool. "We should use a chatbot" is not a problem statement. Start from a decision or a task and its cost.
• Counting hours as cash. Freed capacity has value only when it is realised (fewer hires, growth without hiring, redeployment to higher-value work).
• Accuracy as the only metric. With rare events, a model that always predicts "no" can be 97% accurate and useless. Use precision, recall and the cost of each error.
• Generalising one study. A 14% average gain in one contact centre does not transfer to a different task or workforce; the effect was 34% for novices and close to zero for the most experienced.
• Ignoring the run cost. Licences, usage charges, monitoring, evaluation and knowledge-base upkeep recur every year.
• Treating regulation as a checklist at the end. Classification under the EU AI Act changes the design (human oversight, logging, documentation), so do it at intake.
• Assuming the data moat. Many data advantages are small or short-lived. Test them.
• Over-trusting fluent output. Generative AI sounds right when it is wrong; design checks where errors are costly.

Reading list

• Agrawal, A., Gans, J., and Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press (updated edition 2022). Starred.
• Agrawal, A., Gans, J., and Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press.
• Iansiti, M., and Lakhani, K. R. (2020). Competing in the Age of AI. Harvard Business Review Press. Starred.
• Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942. Starred.
• Dell'Acqua, F., McFowland III, E., Mollick, E. R., et al. (2026). Navigating the jagged technological frontier. Organization Science, 37(2), 403-423. Starred.
• Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Portfolio.
• Russell, S., and Norvig, P. (2020). Artificial Intelligence: A Modern Approach, 4th edition. Pearson (reference for concepts).
• Kohavi, R., Tang, D., and Xu, Y. (2020). Trustworthy Online Controlled Experiments. Cambridge University Press.
• NIST AI 100-1 (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), and NIST AI 600-1 (2024), Generative Artificial Intelligence Profile. Free from nist.gov.

Self-test

1. In one sentence each, distinguish predictive ML, generative AI and an AI agent, and name the main failure to watch for each.
2. Brynjolfsson, Li and Raymond report an average gain and a gain for novices. State both and explain what the difference implies for a firm's training budget.
3. What is the "jagged technological frontier", and what management practice does it argue for?
4. A team of 40 agents gains 14% in resolutions per hour at constant contact volume. How many agent equivalents does that free, and under what condition is it a cash saving?
5. Name five tests that data must pass to be a strategic asset.
6. When would you buy rather than build an AI capability? Give two conditions.
7. Why should the board see NPV rather than IRR for Northfield's demand forecast?
8. Name the four functions of the NIST AI RMF and the standard against which an AI management system can be certified.
9. Give two reasons why a team-level pilot may overstate or understate the effect of an AI assistant.
10. What is the productivity J-curve, and why does it matter when a board asks for results after one quarter?

My notes and next steps

Next steps: [1. ] [2. ] [3. ]

Sources and further reading

• Agrawal, A., Gans, J., and Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.
• Brynjolfsson, E., Li, D., and Raymond, L. (2023). Generative AI at work. NBER Working Paper 31161. Published as Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889-942.
• Dell'Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403-423. Working paper version 2023, SSRN 4573321.
• Noy, S., and Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187-192.
• 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.
• Kotter, J. P. (1995). Leading change: Why transformation efforts fail. Harvard Business Review, 73(2), 59-67.
• National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1; (2024) Generative Artificial Intelligence Profile, NIST AI 600-1.
• Regulation (EU) 2024/1689 (Artificial Intelligence Act), as amended by Regulation (EU) 2026/1744; consolidated text on EUR-Lex.
• ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system.
• Textbooks for further reading (named only): Iansiti and Lakhani, Competing in the Age of AI; Russell and Norvig, Artificial Intelligence: A Modern Approach; Provost and Fawcett, Data Science for Business.
• Open resource: the NIST AI Resource Center (airc.nist.gov) and the NIST AI RMF Playbook, free online.

Answer key

1. Predictive ML learns a score from labelled history (watch drift and biased labels); generative AI produces content from instructions with a general model (watch confabulation and data leakage); an agent plans and acts through tools over several steps (watch wrong or unauthorised actions at scale).
2. 14% on average and 34% for novice and low-skilled agents. The tool partly transfers the know-how of the best agents to new ones, so the firm can shorten onboarding and may shift training spend towards judgment and escalation skills rather than routine knowledge.
3. The uneven boundary of tasks AI does well: inside it AI raises speed and quality, just outside it AI can lower the chance of a correct answer, and the boundary is hard to see. It argues for testing AI task by task, keeping human checks where errors are costly, and training people to recognise where the frontier lies.
4. $40 \times (1 - 1/1.14) = 4.91$ agent equivalents. It is a cash saving only if the firm hires fewer people (for example through attrition) or handles growth without hiring; otherwise it is capacity that must be put to other valuable use.
5. Relevant to a decision, accurate, timely (available when the decision is made), lawfully usable for that purpose, and hard for rivals to reproduce. Ownership and refresh cost are good additions.
6. Buy when the capability is common across firms and not a source of advantage, and when a vendor's product meets the need with acceptable data terms; also when you lack the talent and the time to build.
7. Most of its value is an early one-off working-capital release, which makes IRR very large (158%) and hard to compare; NPV states the value created in money and adds across projects.
8. Govern, Map, Measure, Manage; ISO/IEC 42001:2023.
9. Overstate: stronger teams or supervisors in the treated group, novelty effects, easier queues. Understate: learning curve not finished in eight weeks, spill-over of tips from treated to untreated agents in shared teams.
10. Firms must invest in process redesign, data and skills before output rises, so measured productivity can dip first and rise later. A one-quarter review may wrongly kill a project that is on track; agree leading indicators and evidence gates in advance.

Revision history

Standard and licence note

Based on the published research and public frameworks listed in "Sources and further reading" (Agrawal, Gans and Goldfarb 2018; Brynjolfsson, Li and Raymond 2023 and 2025; Dell'Acqua et al. 2023 and 2026; NIST AI RMF 1.0 and NIST AI 600-1; ISO/IEC 42001:2023 named by number only; the EU AI Act as amended by Regulation (EU) 2026/1744), checked in October 2026. The text, the Northfield case, its exhibits and numbers, and all questions are original; no published case or textbook text is copied or paraphrased, and Northfield Outdoor Co. and every person in it are fictional. Curriculum facts about named schools describe public pages only. The template is independent and not affiliated with or endorsed by any business school, standards body, regulator or publisher. Check the current official texts before you rely on any date or requirement. Template licence: MIT.