## 1. Executive Summary

Alan Nicolas is a generative-AI specialist who says he has built more than R$150 million in revenue across his companies and now runs Academia Lendaria, a school-cum-tooling company teaching businesses to survive and grow in the AI era. His operating position is unusual and specific: his companies have no human copywriter, 80-95% of his hiring processes are done by AI, and he replaced twelve launch support staff with a single AI trained on one day of WhatsApp question-and-answer transcripts in about two hours. He argues the AI market is still raw jungle, so an operator who studies the fundamentals a little more than the competition stands out the way copywriters did in 2017. He also predicts universal basic income is inevitable as 25-40% of current jobs become replaceable, and that within two to three years prompt engineering and junior programming will largely stop being needed. The R$150M figure and the job-displacement numbers are operator claims and projections, not independently verified facts.

## 2. Alpha / Sauce

- **Study ETL and RAG, not just prompts.** Extraction-Transform-Load into a vector database, plus Retrieval-Augmented Generation, is what reduces hallucination. Set up correctly, he says the AI becomes more accurate than humans in his tests. That knowledge is the difference between a toy chatbot and a real assistant.
- **Match the model to the job.** Claude (Anthropic) is best for copy; GPT-4 for flexibility and for making an AI swear/roleplay; Wolfram for math; Mixtral as the math-capable open LLM. Do not use one tool for everything.
- **Feed only scarce, private data.** The model already trained on the public internet, so loading the Bible or generic concepts is wasted effort. Load your company data, your product internals, or your personal definition of a concept like a Big Idea.
- **Chatbot versus agent.** A chatbot follows a fixed script; an agent adapts to the sales stage and can run techniques like SPIN selling. The distinction explains both the wins and the failures the Chevrolet chatbot that "sold" a car for one dollar was a scripted bot defeated by prompt injection.
- **Human in the loop is the operating model.** Do not fire the team; keep one or two people to review AI conversations and retrain the model with thumbs up/down and edited replies. That is how he cut twelve support seats to one AI without losing quality.
- **The lowest fruit is customer service.** Site chatbots, WhatsApp assistants, and SDR agents are the fastest AI wins little technical depth, high certainty of return.
- **Sell shovels.** He frames his business as selling the pickaxe, the jeans, the helmet, and the guide to the gold rush, plus taking a percentage of what students mine. Teaching businesses to make more money is easier to sell than teaching AI.
- **Copy the winners.** Asked where to start in copy, his answer: find who is selling the most, buy their VSL, run a spy tool, and duplicate. Study fundamentals and steal the proven structure.
- **YouTube beats TikTok and Instagram for lead quality.** The intent at the moment of use is different: people go to YouTube to learn or to listen to a two-hour podcast, which produces far better-qualified leads.

## 3. Methods & Setup - Step-by-Step

1. Pick the business function everyone shares sales and marketing first, then HR and target it rather than chasing the flashiest AI demo; setup: a single high-frequency workflow as the pilot.
2. Start with the cheapest fruit: a site chatbot or a WhatsApp attendant; setup: a simple plug-and-play stack such as Chatbase, Chatvolt, or an equivalent WhatsApp automation tool.
3. To beat the "it will hallucinate" problem, learn ETL and RAG before scaling; setup: extract your documents, transform and clean them, load them into a vector database, then retrieve them correctly at answer time.
4. Choose the model by task; use a cheaper model like Claude Haiku when you can, mid-tier Sonnet for copy, and reserve the expensive Opus for genuinely complex reasoning; setup: per-task model routing rather than one default.
5. Train a support AI on your own transcripts instead of guessing; setup: add the AI to the launch group, download the questions and answers from WhatsApp, paste the day's thread into Claude or Gemini, and extract a Q&A file.
6. Keep humans in the loop for quality; setup: a review dashboard where a thumbs-up re-trains the answer, a thumbs-down removes it, and an edit saves the corrected reply to the database.
7. Build assistants around real sales stages (an SDR agent) rather than a static script; setup: load a sales method such as SPIN selling and let the agent advance the conversation by stage.
8. Guard against prompt injection and over-promising; setup: hard constraints on what the agent may offer, and human approval for anything unusual.
9. When you must feed a model your data, import only what is scarce and private your company data, a niche author, or your internal definition of a Big Idea; setup: a curated, non-public document set.
10. Run creative variation as a production line: take a working ad and vary the format (podcast, question box, interview, hook), the edit, and the copy; setup: one proven creative as the seed and an AI to generate lateral variants.
11. Deploy AI social selling carefully: let an AI profile respond to followers, but keep your personal direct messages human when the relationship is sensitive; setup: scoped inbox access, not full WhatsApp access.
12. Do not hire junior programmers; instead staff senior/pleno engineers who can orchestrate several AIs and hold business context; setup: a hiring bar of pleno-and-up only.
13. Staff every business with three capabilities at least one automation person, one prompt engineer, and one data person (they can be one strong generalist); setup: a role map and an upskilling path.
14. Invest in durable human skills as insurance: emotional intelligence to handle constant change, clear thinking, reading, vocabulary, and a genuine hobby; setup: a personal learning routine independent of any single tool.

## 4. Tools & Stack

| Tool | Role |
| --- | --- |
| Claude (Anthropic) | Best-in-class copy model; Sonnet for output, Opus for complex reasoning |
| ChatGPT / GPT-4 | Flexible multimodal model; used for persona-swapping assistants |
| Gemini 1.5 | Long-context model he gives students free access to (large token window) |
| Wolfram / Mixtral | Math and scientific reasoning |
| Chatbase / Chatvolt | Low-code chatbot builders on site and WhatsApp |
| Flowise | Visual builder for ETL/RAG pipelines; its docs teach the fundamentals |
| Hugging Face | Nearly 800,000 open models; the library to browse before choosing |

## 5. Actionable Checklist

- [ ] Pick one shared business function and launch a single AI pilot this week.
- [ ] Stand up a site or WhatsApp chatbot as the first, cheapest win.
- [ ] Learn ETL and RAG so the assistant stops hallucinating.
- [ ] Route tasks to the right model instead of defaulting to one.
- [ ] Replace your next round of support hiring with an AI trained on existing Q&A transcripts.
- [ ] Keep a human reviewer in the loop with thumbs-up, thumbs-down, and edit feedback.
- [ ] Build the next assistant as a stage-aware SDR agent, not a fixed script.
- [ ] Add hard constraints so the agent cannot over-promise or be jailbroken.
- [ ] Feed the model only scarce, private data.
- [ ] Turn one winning creative into a lateralization line spanning format, edit, and copy.
- [ ] Hire pleno-and-up engineers, and add automation, prompt-engineering, and data capability to the team.
- [ ] Train durable human skills clear thinking, reading, emotional range as your hedge against displacement.