September 16, 2026 · 7:00 PM ET
Building an AI Team for International Marketing
Video Conference (Google Meet) · 3 hours
Three AI agents and an orchestrator, built live for a fictional Brazilian restaurant chain deciding whether to open in Miami, Lisbon or Buenos Aires - from raw spreadsheets to a board recommendation where every number traces back to its source.
Venue
Delivered live over Google Meet. Participants follow along on their own computers: the complete case package is downloadable from this page, and the tool used in the session, Claude desktop with Cowork, runs on Windows 10+ or macOS 11+ with a paid Claude plan (Pro, Max or Team). Ten minutes to install, four steps.
Audience
MBA students at the State University of Campinas (UNICAMP), in the state of São Paulo, Brazil. Established in 1962, UNICAMP was designed from scratch as an integrated research center, unlike other top Brazilian universities, which were usually created by consolidating pre-existing schools and institutes. unicamp.br
Guest session in the International Marketing course. The session is delivered in Portuguese.
Agenda
- 01
The Case: Cumari Cozinha Brasileira
A fictional São Paulo restaurant chain (38 own stores, R$ 152M in annual revenue, 22% store margin, a “contemporary Brazilian” brand) takes the board’s question: should we go international, and if so, where and how? Three candidate markets (Miami, Lisbon, Buenos Aires), a R$ 12M budget, a 36-month payback per unit, a 20% minimum IRR in BRL and a brand that cannot be licensed. No candidate wins on everything, which is what makes it a real decision.
- 02
Organizing AI as a Department
Why one agent per stage beats one giant prompt: a single owner per step, handoffs as files that appear in the project folder, an orchestrator that coordinates but never analyzes, and a result that reruns when the data changes. The anatomy of a reliable agent in five components: trigger, objective, process, context, rules. What makes an agent trustworthy is its structure, not the size of its prompt.
- 03
The Project: Folders, Context and Source Data
Building the case project step by step: five folders (company context, source data, agents, agent results, final deliverables) plus the master instruction file. Who authors each input and why it matters: the executive briefing, the expansion policy and the board decision framework come from the CEO, CFO/legal and governance, never from the department asking for the expansion; the three source spreadsheets (store performance, candidate market profiles, macro and cultural assumptions) come from operations, market intelligence, treasury and HR. Analysis without the company’s own rules is opinion; the documents turn analysis into verification.
- 04
Building the Three Agents and the Orchestrator
Live in Claude Cowork, one creation prompt per agent, each packaged as a reusable skill. Agent 1, the International Market Analyst: internal readiness, PEST, CAGE distances, Hofstede dimensions and a weighted attractiveness matrix - it compares and never recommends. Agent 2, the Market Strategy Manager: 60-month projections per market in local currency, entry modes, marketing mix (4Ps), pricing, payback, IRR, NPV and FX sensitivity, with a pass/fail check against the expansion policy - it verifies and never writes for the board. Agent 3, the International Marketing Director: challenges the previous agents’ assumptions, applies the five-dimension decision framework with green, amber and red signals, and delivers the verdict as a board deck. Then MASTER_INSTRUCTION.md, the orchestrator that discovers the files, invokes each agent in order and validates every handoff through gates: a missing file stops the flow, never “carry on anyway”.
- 05
Running the Flow and Auditing the Output
A new session, the folder connected and a single instruction: the whole team runs from raw data to Recomendacao_Diretoria.pptx in about 65 minutes, and you watch each deliverable appear in the folder in order. The Cumari verdict: proceed with conditions, Lisbon, two phased own stores - three greens, two ambers, zero reds, with four conditions that each have an owner, a deadline and a trigger. Then the audit script: placeholders zeroed, numbers cross-checked between files, boundaries respected, verdict driven by the framework’s rules rather than the agent’s taste. Closing with reuse: a new quarter, new markets, a new company or a new policy limit, same agents.
What participants take away
- Decompose a business decision into specialized AI agents, each owning exactly one stage, with file-based handoffs that make the work visible and auditable in a project folder.
- Write a creation prompt around the five components (trigger, objective, process, context, rules) and turn it into a reusable Claude skill, complete with references, scripts and a validator, so the agent decides and the script does the arithmetic.
- Set up a project structure and a master instruction file so that a brand-new session can run the entire team, in order, with one instruction and no manual coordination.
- Read AI output like a board member: check that no agent crossed its boundary, trace every headline number back to its source tab, and tell a verdict that follows explicit decision rules apart from one that merely sounds confident.
- Apply the international marketing toolkit - PEST, CAGE, Hofstede, attractiveness matrix, entry modes, 4Ps, payback, IRR, NPV and FX sensitivity - through agents rather than by hand, and rerun the same team when the data, the markets, the company or the policy change.
Lecture materials
Materials go live here after the session. Enter the password shared during the lecture to unlock downloads - bookmark this page.
Checking access…