This page is password protected. Enter the password to read it.
Working notes for Session 3: draft assignment questions, a small-group exercise, questions for the Pam Habner visit, evidence caveats, and the build tracker. The real teaching window is 80 minutes, Sloan classes start five late and end five early.
Posted on Canvas with the case and the pre-read excerpt. Written company-agnostic until the case company is secured; the bracketed phrases get replaced once the case is drafted.
Use the primer's five-part breakdown: input, output, steps, context per step, data touched. Where is the one hard dependency you cannot skip, and where would you draw a privacy wall?
The primer argues the reflex is automating the existing process and the reframe is redesigning its shape. What would the reframe version have looked like, concretely? Which tasks collapse, which grow, and who decides what?
Use the four gauges (speed, cost, rigor, containment) and one organization-structure analogy from your own work experience: a firm you have worked in that already runs one of these shapes.
Based on the case, what evidence do you see of change anxiety or resistance among frontline domain experts, and what would you have done differently to enlist them before rolling anything out?
Deliberately not the billing-dispute example already worked in the primer, so groups must transfer the framework rather than recite it. Deliberately adjacent to Pam Habner's actual Citi Cards business, so it sets up sharper speaker questions. Groups of 4–5, ten minutes, sketched on a half-sheet or mini-whiteboard.
I added it deliberately. It is not in the primer. It forces groups to supply the human-enlistment dimension the reading omits, and it feeds assignment question 4 and the Habner question on rep anxiety.
Pick one group that chose a fast, lean shape and one that chose a heavier, review-laden shape, so the tension is visible on its own.
Then the probe, the same probe regardless of what the groups chose:
"Why not a pyramid here, given real money moves?"
The probe surfaces the underwriting-risk tension: a credit limit increase is a lending decision, and the write step carries real exposure. Strong groups will defend a lighter shape by pointing at where they placed the human commit point; weak groups will not have noticed the read/write line at all.
Citi Cards handles everything from balance checks to fraud and hardship calls. How do you decide, structurally, which calls get full agent autonomy versus a human in the loop, and has that line moved as your agents have matured?
The reading argues leaders should redesign the shape of the work rather than just make the old process faster. Can you point to a place at Citi where your team resisted the tempting "faster version of today's process" and instead restructured decision rights or the process itself?
Consumer credit is one of the most regulated environments for AI: fair lending, adverse action notices, data privacy. Where do you draw your internal "privacy wall," and has a regulatory requirement ever forced you into a less efficient topology than you would otherwise have chosen?
Separate from the technical design of the system, what did you do to address your frontline reps' anxiety or resistance when this rolled out, and what specifically worked or didn't?
Escalation and retention calls are exactly where emotion and stakes are both highest. Now that agents absorb the routine volume, is the reps' job becoming the harder, more concentrated job this reading predicts, and how are you preparing or incentivizing reps for that shift?
The participation rubric asks students to "apply course concepts to case evidence," so our own evidence has to survive the same scrutiny. Everything below is what I say when a sharp student checks the sources.
All three numbers are verbatim-accurate to Anthropic's engineering post, but they do not come from one test. The 90.2% is from Anthropic's internal research eval (Claude Opus 4 lead agent + Claude Sonnet 4 subagents versus a single-agent Opus 4, the same top-tier model as baseline, not a weaker one). The 80%-of-variance figure is from a separate benchmark, BrowseComp, which tests browsing agents on hard-to-find information. The ~15× is multi-agent versus a single chat (single agents run ~4×). If a student asks "variance in what test?", the honest answer is: a different one than the 90.2%. Say it before they do.
Harder tasks plausibly both need more tokens and score differently, so the 80%-of-variance figure is confounded, and Anthropic's own writeup treats it as suggestive rather than causal. Add the one-sentence caution whenever the stat is used live. The stronger framing dissolves the objection: topology (independent, parallelizable subtasks) is precisely what makes spending more tokens productive rather than wasteful. Shape and token spend are complementary explanations, not competing ones.
The same Anthropic post states plainly: "some domains that require all agents to share the same context or involve many dependencies between agents are not a good fit for multi-agent systems today" and singles out coding as the paradigm case. Cognition's "Don't Build Multi-Agents" names the failure modes a well-read student may quote verbatim: context fragmentation and dispersed decision-making. Note that the primer's own refund example is dependency-heavy (verify before pull, judge before size), which is why it decomposes into a pipeline, not a parallel fan-out. That is a feature of the example, not a bug: use it.
Google Research (2026), "Towards a science of scaling agent systems": on parallelizable financial-reasoning tasks, centralized multi-agent coordination gained 80.9% over a single agent; on sequential planning tasks (PlanCraft), every multi-agent variant degraded performance by 39–70%. Communication overhead fragments reasoning. Error amplification: 17.2× for independent agents without mutual validation versus 4.4× for centralized systems with orchestrator validation. Their 87%-accurate predictive model picks the architecture from measurable task properties, effectively an empirically grounded version of the primer's four design questions. This is the class's best defense against overclaiming and its best evidence that the four questions are the right questions.
The speed/cost/rigor/containment bars show relative intensity with no stated scale. The primer says "not a grade," and a quant-minded MBA will notice the bars look measured. If challenged, credit the objection: the gauges are a structured judgment aid, and "containment" in particular bundles distinct risks (data leakage, blast radius from write actions, output quality) that a rigorous design would split. Inviting the class to split it is a better teaching move than defending the bars.
Verified against primary sources: Galbraith (1974), "Organization Design: An Information Processing View," Interfaces 4(3), 28–36: structure exists to match information demands to processing capacity; the primer's compression is faithful. Mintzberg (1979), The Structuring of Organizations: the primer's five structures map 1:1 onto his five configurations. Joseph & Sengul (2025), Journal of Management 51(1), 249–308 is exactly cited per Crossref; if a student presses on "equifinality" specifically, the canonical source is Gresov & Drazin (1997, Academy of Management Review).
Fresh backup if wanted: OrgAgent (arXiv, April 2026), which itself cites Galbraith and Mintzberg, found company-style hierarchical multi-agent systems outperformed flat ones by 102.73% while using 74.52% fewer tokens on one benchmark: quantitative support for the pyramid ↔ layered-review-firm mapping.
The primer labels the closing forecast as personal prediction, not findings. Keep that framing in class. It is not assigned, not testable, and enters the session only as an optional one-liner in the debrief to seed reflection question 2. If it draws governance pushback, the serious anchor is Anthropic's 2026 "AI Organizations are More Effective but Less Aligned" result (diffusion of responsibility). Cite that as a caution alongside, not instead of, the performance evidence.
Claims about a JPMorgan "EVEE Intelligent Q&A" tool and "1,000+ AI use cases by 2026" appear only on low-authority aggregator sites and were not corroborated by JPMorgan's newsroom or top-tier press. Do not cite them in class or in any derived material. If a JPMorgan comparator is needed, Tearsheet's ~450 documented gen-AI use cases figure has better sourcing.
The primer is the source document; everything for the session and beyond gets built from it. This table is the running to-do. Update the chips as things land.