🍞 Brief: Padaria Pão&Cia
Neighborhood bakery, 12 employees, 1 owner (Marta) who handles admin. Sells at the counter and via WhatsApp. Receives ~40 orders/day via WhatsApp, issues invoices manually, tracks inventory in a spreadsheet, and answers the same customer questions all day.
Use this context for all six exercises. Wherever there’s a number, make up a plausible one and stay consistent.
🎯 Find 3 tasks that take 10 hours
Your mission: find, at Padaria Pão&Cia, three repetitive tasks that add up to about 10 hours a week and are strong candidates for automation.
🎯 Objective
Leave with a list of 3 tasks, each with its frequency, duration, and pain level—ready to become a project.
Step by step
List everything Marta does on a typical day (write down at least 10 activities).
For each one, estimate frequency (x/week) and duration (min/use).
Multiply to find hours per week and rate the “pain” (1–5). Choose the 3 with the highest hours × pain.
✅ "Done" criterion
You have 3 tasks, each with calculated hours per week and a pain score from 1 to 5, and the total for all three is between 8 and 12 hours.
Sample answer key
Tarefa | freq/sem | min/vez | h/sem | dor --------------------------------|----------|---------|-------|---- Responder dúvidas no WhatsApp | 200 | 1.5 | 5.0 | 4 Lançar pedidos na planilha | 200 | 1.0 | 3.3 | 3 Emitir nota fiscal manual | 60 | 2.0 | 2.0 | 5 --------------------------------|----------|---------|-------|---- TOTAL 10.3 h/sem Top 3 (horas × dor): WhatsApp (20) · Nota (10) · Planilha (10)
How many times per week
Minutes per run
Frustration rating, 1–5
hours × pain ranks
📊 Audit and calculate ROI
Take the top task from exercise 1 (responding to WhatsApp) and calculate the return on automating it.
🎯 Objective
Arrive at an ROI and payback figure that justifies the project—or doesn’t.
Formulas you’ll use
economia_mensal = horas_mes_economizadas × custo_hora roi_% = (economia_anual - custo_projeto) / custo_projeto × 100 payback_meses = custo_projeto / economia_mensal
✓ Do it this way
- ✓Use the fully loaded hourly cost (salary + payroll taxes and benefits), not the raw salary
- ✓Count only the hours that the automation actually saves
- ✓Include API/tool costs in the project cost
✗ Avoid
- ✗Overstate the savings by assuming 100% automation
- ✗Forget the monthly maintenance cost
- ✗Ignore the remaining time spent on human review
✅ "Done" criterion
You have monthly savings, annual ROI in %, and payback in months, with all assumptions documented.
Sample answer key
Premissas horas_mes (WhatsApp) = 20 h % automatizável = 70% -> economiza 14 h/mês custo_hora (Marta) = R$ 45 custo_projeto = R$ 2.500 (build) + R$ 120/mês (API+ferramenta) Cálculo economia_mensal = 14 × 45 = R$ 630/mês (bruto) economia_líquida = 630 - 120 = R$ 510/mês economia_anual = 510 × 12 = R$ 6.120 roi_% = (6120 - 2500) / 2500 × 100 = 144,8% payback = 2500 / 510 = 4,9 meses
Loaded, not raw
Be conservative
% of the investment
How many months to pay for itself
🗺️ Diagram the as-is → to-be
Map the process customer support via WhatsApp as it is today (as-is) and how it will be automated (to-be).
🎯 Objective
Two flows side by side, making clear which steps AI takes over.
Step by step
- Write the current workflow as a sequence of steps (each step: actor + action).
- Mark each step an AI could take over with 🤖.
- Rewrite the to-be workflow by removing or condensing the automated steps.
- Indicate the decision points (where a “yes/no” is needed).
Sample answer key
AS-IS (hoje)
cliente → manda msg no WhatsApp
Marta → lê a msg
Marta → identifica o tipo (pedido? dúvida? reclamação?)
Marta → responde manualmente
Marta → se for pedido, copia pra planilha 🤖
Marta → confirma valor e prazo 🤖
TO-BE (automatizado)
cliente → manda msg no WhatsApp [trigger]
IA → classifica (pedido/dúvida/outro) 🤖
decisão → é dúvida comum? --sim--> IA responde da base 🤖
--não--> encaminha p/ Marta
IA → se pedido, extrai itens e grava na planilha 🤖
IA → devolve confirmação ao cliente 🤖
✅ "Done" criterion
Two complete flows, automated steps marked with 🤖, and at least one explicit decision point.
Who does what
What AI assumes
Yes/no points
What triggers the workflow
📐 Design the blueprint
Turn the to-be into a one-page blueprint: trigger, tools, model, logic, and actions.
🎯 Objective
Technical specifications detailed enough for someone to start building without asking you anything else.
Blueprint template (fill in)
BLUEPRINT — Atendimento WhatsApp Pão&Cia
Trigger : nova mensagem no WhatsApp Business (webhook)
Ferramentas : n8n · WhatsApp API · Google Sheets · modelo de IA
Modelo : LLM rápido/barato (classificação) + 1 chamada p/ resposta
Lógica :
1. classificar intenção -> {pedido | duvida | outro}
2. se duvida -> RAG na base de FAQ -> resposta
3. se pedido -> extrair itens (JSON) -> append na planilha
4. se confiança < 0.6 -> encaminhar p/ humano
Ações : responder cliente · gravar planilha · notificar Marta
Pontos de falha: msg ambígua · item fora do cardápio · API fora do ar
📋 Blueprint checklist
- • Is exactly 1 trigger defined?
- • Does every decision have an "otherwise" path?
- • Is there a human fallback for low confidence?
- • Are the failure points listed?
✅ "Done" criterion
The blueprint fits on one page and includes a trigger, numbered logic, actions, and at least 3 failure points.
The single trigger
Tools and model
Numbered steps
Where things can break
🔧 Build a simple flow
Build a minimum 3-node workflow: trigger → AI → action. Here you can build it for real in n8n/Make or simulate it with pseudocode.
🎯 Objective
A flow that receives a message, calls AI, and performs an action — end to end.
Pseudo-flow (copy and adapt)
// NÓ 1 — TRIGGER (webhook WhatsApp)
input = { from: "+55...", text: "quero 2 pães de queijo e 1 café" }
// NÓ 2 — IA (extração estruturada)
prompt = `Extraia o pedido como JSON do schema:
{ "itens":[{"produto":string,"qtd":number}], "obs":string|null }
Responda SÓ o JSON. Texto: ${input.text}`
ia_out = chamarLLM(prompt) // -> { "itens":[...], "obs":null }
pedido = JSON.parse(ia_out)
// NÓ 3 — AÇÃO (gravar + responder)
appendPlanilha("Pedidos", pedido.itens)
responderWhatsApp(input.from,
`Anotado! ${pedido.itens.length} itens. Confirma? (sim/não)`)
💡 Practical tip
Test the AI node in isolation first, with 3 different real messages, before connecting the action. 80% of bugs are in JSON parsing.
✅ "Done" criterion
The workflow runs with 3 different test messages and records/responds correctly to at least 2.
Incoming webhook
Extract JSON from text
Validates the output
Records and responds
💼 Put together a business proposal
Package everything in a 1-page proposal for Marta: problem, solution, ROI, price, and next steps.
🎯 Objective
A proposal that anchors the price to the ROI from exercise 2 and ends with a clear CTA.
Proposal template (fill in)
PROPOSTA — Atendimento automático no WhatsApp Para: Marta · Padaria Pão&Cia 1. O problema ~20 h/mês respondendo as mesmas dúvidas e lançando pedidos à mão. 2. A solução Assistente no WhatsApp que classifica, responde dúvidas da base e registra pedidos na planilha automaticamente. Você só revê exceções. 3. Resultado esperado ~14 h/mês liberadas · menos erros de lançamento · resposta em segundos. Economia estimada: R$ 510/mês líquidos · paga em ~5 meses. 4. Investimento Setup: R$ 2.500 (única vez) · Manutenção: R$ 120/mês. 5. Próximo passo Aprovação até [data] -> entrego rodando em 2 semanas. Responda "topo" para começarmos.
⚠️ Attention
Never price by the cost of your time. Price at a fraction of the value the client gains. Here, the price (R$ 2.500) is less than half the annual savings—so it’s easy to say yes.
✅ "Done" criterion
One-page proposal with the 5 sections, price anchored to ROI, and a CTA with a date and clear action.
In the client’s pain point
Anchors the price
Share of the value
Clear date + action
📌 Module Summary
Next Module:
6.2 — Ready-to-use prompt library (speed up each exercise with copy-and-paste prompts)