You don't have AI. You have 30 uncoordinated ChatGPTs.

Before the map

I sit in a meeting with the CEO. It shows me the screen: ChatGPT open. Smile.

"Ricardo, we have finally entered the age of AI."

I take a deep breath. I choose my words carefully. This isn't the first time we've talked about AI — in previous diagnoses, the topic had already appeared. But saying it there, in a formal meeting, with the board present and the budget for the next cycle on the agenda, has another weight.

"Maybe not yet, the way you're thinking. You have 30 uncoordinated ChatGPTs — what the technology market today calls 30 personal agents without organizational coordination. It's not exactly the same thing as having AI in the company."

This CEO is not naive. On the contrary — he founded a medium-sized company that has been growing for fifteen years. What happened to him happens to most executives I know: the technology market has trained everyone to treat "buying AI" as a single decision, when in fact they are five distinct decisions, on different scales, with different consequences. And the menu that arrived on his table only has ChatGPT Enterprise.

So he bought it. Distributed access. Each manager started to use it in their own way. Each one discovered their own trick, built their library of prompts, adjusted their collection of instructions that took six months to get right. Each one became a small expert on their own tool — silent, isolated, individually valuable.

None of this goes back to the company. When the manager leaves, he takes everything. And then the company doesn't have AI: it pays for 30 personal tool licenses. The entire sector, from big tech companies to consultants, sold this as digital transformation — and almost no one stopped to say that something basic needed to be agreed upon first.

The numbers show that this scene is today the picture of most Brazilian companies.

CGI.br, in the ICT Companies 2024 survey, measured: only 13% of Brazilian companies have structured adoption of AI. The same percentage as in 2021 and 2023. Three years after ChatGPT, GPT-4, Claude, Gemini — the same number.

In the same period, Sebrae recorded that 44% of Brazilian entrepreneurs already use some form of AI. And Microsoft showed that 75% of micro, small and medium-sized companies in the country are "optimistic" about the impact of technology on their business.

Three numbers, three official sources. Seventy-five percent want it. Forty-four tried it. Thirteen organized. The vast majority of Brazilian companies today live exactly in the middle of these numbers — between the real will and the absent method. It's not an individual fault. It's a map error.

While most are in the middle of the road, a small group of companies that mapped out early are already paving the way.

Lu do Magalu, equipped with generative AI in partnership with Google, converts three times more than the company's own application and solves 70% of WhatsApp demands alone. Itaú has put 150 generative AI solutions into production by the beginning of 2026 — and launched Itaú Emps, an application that replaces human managers in serving micro and small companies. Brazilian industry, which had 16.9% of companies using AI in 2022, reached 41.9% in 2024 — jumping 25 points in two years, according to IBGE.

It's not magazine hype. It's IBGE, it's CGI.br, it's the financial statement of a listed company.

The question that CEO asked me — “do I need to buy AI?” — is the question that the market has trained every executive to ask. It's a binary question, yes or no. And that's precisely why it's interesting: because AI is not binary, and any answer that comes from this question unlocks at most one movement.

The question that unlocks is another: at what level of AI is each part of my company today, and what is the next move — by area?

Sales can be in one place. Tax in another. HR in yet another. Service in another. And the manager of each area — not the CEO alone — is the one who needs to understand the map to answer this question in his own chair.

This article is the map that CEO didn't have when he showed me ChatGPT open, smiling. It's the map that I didn't have when I started helping companies on the subject, years ago — I put it together while observing what worked and what didn't work in client after client.The next few acts describe five types of AI that coexist in any company that is taking the subject seriously. Each with a different function. Each with a different cost. Each one has a limit — and a clear signal of when your area needs to go up a notch. Along the way, I will share three things that I had to learn in practice and that few people talk about: that this ladder is not a mandatory path, that the biggest risk in implementation is not technical (it is human and silent), and that there is a path — the last one — that is not "the top of the ladder", but an architectural decision that can be made from day one.

Before paying for your next ChatGPT Enterprise license, read the map.

The five types of AI in your company

Before describing the five types, three important caveats.

First: I will use two names for each one. An editorial name, based on human hierarchy — Intern, Specialist, Trained Employee, Autonomous Executor, Second Brain. And a technical name, which is how the market in general, Anthropic, OpenAI, Salesforce and Brazilian suppliers call these categories — Conversational Agent, Specialized Agent, Customized Agent, Autonomous Agent, Organizational Memory Agent. They both talk about the same thing. Use whichever is most natural for you — and in internal conversations, both serve you.

Second: this list is a map, not a ladder. You don't have to climb all five. Your area could stop at one of them and be great at it. Another area of ​​the same company can jump straight to the fourth without going through the second one. I will return to this uncomfortable truth in the next act.

Third: none of the five replaces the others. In a company that is taking AI seriously, all five coexist. The question is not "which one to choose". It's "what's the next thing in my field."

With the caveats out of the way, let's look at the types.

Step 1 — The Intern (Conversational Agent)

It's the generic, one-to-one conversational AI tool without company context. ChatGPT, Claude Desktop, Gemini, personal Microsoft Copilot — they all come in here. Typical cost in Brazil: R$100 to R$150 per user, per month.

The function is individual productivity in administrative tasks — writing, summarizing, brainstorming, translating, quick analysis. The available literature (Microsoft Work Trend Index, McKinsey research) shows a gain of 20% to 40% in time freed up in these tasks, when the user knows how to ask.

Why Intern: It's like having a brilliant recent graduate in the room. He knows a lot in theory, but he doesn't know the company, he doesn't know the client, he forgets everything between one conversation and the next. You need to give context every time.

Structural limit: zero company memory, zero coordination between users, knowledge is lost when the manager leaves. It's exactly the CEO scene at the beginning of this article.

A sign that your area needs to improve: "everyone here uses ChatGPT, but everyone does it differently, and nothing we learn is recorded for the team".

When you don't need to go up: if your area uses the Intern only for individual productivity in tasks that don't require standardization — writing personal emails, individual creative brainstorming, exploratory research — Step 1 is enough. Don't force it up unnecessarily.

Illustrative case: widespread casual use is the case. Sebrae recorded in 2025 that 44% of Brazilian entrepreneurs already use some AI, and the overwhelming majority of these use exactly this. It's what makes Step 1 invisible — everyone uses it, no one notices.

Step 2 — The Specialist (Specialized Agent)

It's AI with deep expertise in a single domain — code, design, writing, research, sales, service, data. Examples: GitHub Copilot and Cursor (code), Midjourney (design), Jasper (marketing), Perplexity Pro (research), Microsoft 365 Copilot (writing). Typical cost: R$150 to R$1,400 per user, per month.

The function is categorically different productivity in a specific domain — it's not "20% faster", it's typically 5 to 10 times faster. Programmer with Cursor or Copilot builds in days what previously took weeks.

Why Specialist: it's like hiring a senior professional from a single area — he knows a lot about what he does, it's expensive, he doesn't know anything beyond that.

Structural limit: function silo. The Code Specialist knows nothing about marketing or the company. Each Expert in your organization is an independent silo.A sign that your area needs to move up: "I pay AI Specialist for ten professionals, I see a clear individual gain, but there is nothing organizational coming out of there — each one became a better silo than it was, without it becoming a standard for the house. And when one of them leaves, they take the gain with them".

When you don't need to go up: if your area is highly specialized (engineering, design, R&D) and technical gain is your biggest productivity factor, Step 2 may be where your area remains in the long term. Going up only makes sense when you need to mix the domain with the company context.

Illustrative case: RD Station launched Mentor AI within RD Station Conversas at the end of 2024 — an AI layer specialized in marketing and sales embedded in the product. It works because it serves marketers in a single domain, with its own vocabulary and its own metrics.

Step 3 — The Trained Employee (Custom Agent)

It is the agent configured with company-specific instructions, curated knowledge base (selected and organized), defined persona. Examples: Custom GPTs (OpenAI), Claude Projects (Anthropic), Microsoft Copilot Studio, agents configured by the team for internal tasks. Cost: setup time plus usage, usually included in the corporate plan. Realistic estimate: 20 to 40 hours to set up a Custom GPT that replaces 10 hours a week of repetitive team work.

The function is to standardize recurring company tasks. The Trained Employee has partial knowledge of the organization — you taught him what he knows.

Why Trained Employee: it's like hiring a person, teaching them how to do a specific task, and that person now does that well, always the same. But it only does that. And if the rule changes, you need to retrain.

Structural limit: knowledge frozen at the time of configuration. Don't learn alone. If company policy changes, someone has to update manually — and usually no one knows for sure what's out of date within the bot.

A sign that your area needs to improve: "I created a Custom GPT that answers HR questions, but every time the policy changed, someone had to update it manually, and no one can say for sure what is out of date in there today".

Illustrative case: Conta Azul launched the SME accountant's financial assistant in August 2025 — it spent six months testing with one hundred accounting companies before releasing it to the base of more than one hundred thousand customers. It's a well-done Trained Employee: specific role (assistance to the accountant), curated base (a knowledge base selected and organized by experts — not everything thrown in, but filtered, structured and validated for quality in the response), trained with real cases before going into production.

Step 4 — The Autonomous Executor (Autonomous Agent)

It is the agent that acts in the background, executes decisions, monitors processes, integrates with systems, and does things without you clicking. Examples: Devin (Cognition Labs), Claw, Hermes, Microsoft Copilot Studio Agents, Salesforce Agentforce, agents built on n8n with LLM. Cost: R$500 to R$10,000 per month, depending on scale and integration with internal systems.

The function is to execute decisions already made. When customer X does Y, they do Z — automatically. Monitors processes, acts within pre-defined rules, frees the human team from repetitive coordination work.

Why Autonomous Executor: is an employee who executes in the background without direct supervision, but only within the defined scope. Salesforce, when launching Agentforce in 2024, proposed a direct analogy with self-driving car levels — from 0 to 5 — to describe this type of agent.

Structural limit: executes very well what is ordered, but does not decide what should be done. And without an organizational memory layer (which is Step 5), you forget the history between sessions — each time you start from conceptual zero.

Sign that your area needs to grow: "I need something to happen every time client X does Y, without someone clicking — and the volume is big enough that this manual coordination is consuming good people in repetitive work".Illustrative case: Stone Payments was the first Brazilian acquirer to build its own anti-fraud system with AI. The system assigns a five-level risk score to each transaction — and autonomously decides to approve or block it. Critical decision, at scale (millions of transactions per day), but within clear and audited rules. It is Step 4 calibrated.

Critical counterexample: Klarna announced in February 2024 that it had replaced 700 human agents with AI. In May of the same year, he revised — he hired humans again for part of the volume. The problem was not the technology. It was skipping the previous internal sales stage and the organizational memory layer. I'll come back to this case in detail in Act 3, when we talk about the biggest silent risk of deploying AI.

Step 5 — The Second Brain (Organizational Memory Agent)

This is not "the most advanced tool". It is a distinct category. The Second Brain does not replace the four previous types: orchestra. It is the basis on which others begin to truly multiply.

Four components make up this architecture — this is OPEX's thesis, developed in the manifesto Memory before Intelligence (May 2026):

• Memory — unique, living, indexed truth of the company. Not in Drive, not in CRM, not in the manager's head.
• Policy — explicit, versioned rules, outside the leader’s head.
• Decision — immutable event with why, context, alternatives. Not revisitable due to guesswork.
• Cadence — auditable ritual of perceiving, planning, acting, reflecting.

There is no off-the-shelf product ready for Step 5. It's architecture — it combines tools (that change over time) with human curation and organizational discipline. OPEX OS is the name of our approach; There are variations in companies that have built their own path. Cost: initial investment in architecture plus ongoing curation. It's not a monthly license per user — it's a structural project. Typical range for a medium-sized company: R$50,000 to R$500,000 for setup, plus operating costs proportional to size.

The function is simple to describe and difficult to implement: the company remembers. Remember decisions with why. Enforces policy consistently. Records every relevant movement. Learn from your own history — and stop repeating the same mistakes from two years ago.

Why Second Brain: is organizational memory living outside people's heads. As the human brain remembers for the individual, the Second Brain remembers for the company.

Structural limit: requires discipline and time. It doesn't pay off on the first day — it starts paying off in 60 to 90 days and grows with the company. It is the only one of the five with a relevant implementation cost.

But it's the only one that scales without losing knowledge. And it is the only one that can be built from day one of the company's AI journey, instead of being the arrival point. I'll come back to this difference in depth in Act 5 — because it's where the blind spot in almost every discussion of enterprise AI lies.

Sign that your company needs: "we repeat mistakes from two years ago", "when someone critical leaves, they take away knowledge that I can't recover", "decisions become phantom rules that no one can explain", "I can't prove to a new partner why we do what we do the way we do".

Illustrative case: there are still few public cases in Brazil — the architecture is recent. OPEX OS, with a manifesto published in May 2026, is the reference I defend. I delve deeper into Act 5.

Five types. Five distinct functions. Five distinct costs. Neither replaces the other — they all coexist in any company that is taking AI seriously.

The question is not "which one to choose". It's "what's my next move in each area of ​​my company".

And to answer that, we need to look at a truth that few consultancies like to admit.

The trail is not linear

The popular expectation about AI in the company is the expectation of any technological wave: let's all move forward together, area by area, in a coordinated plan. That's not how it works. It never worked — not with ERP, not with CRM, not with BI, not with cloud migration. And AI will be no exception.

The practical truth is uncomfortable: in the same company, at the same time, different areas are at different levels. And that's normal — it's what to expect, not what to avoid.IBGE data shows this dispersion from above. Brazilian industry jumped from 16.9% of AI adoption in 2022 to 41.9% in 2024 — 25 points in two years. In the same period, the general average for Brazilian companies, measured by CGI.br, was 13%. In other words: the industry is three times ahead of the country's average. And within the industry itself, discrete manufacturing advances at a different pace than a continuous process, which advances differently than extractive manufacturing.

Each sector follows its own rhythm. Those who have greater pain, available data and the ability to apply it, advance faster. Those who don't have any of the three are stuck — and this explains why the national average remains at 13%, even with so many headlines about digital transformation.

The same dispersion appears when looking by size. According to CGI.br, in 2024, 38% of large Brazilian companies were already using structured AI, compared to 29% of medium-sized companies and only 10% of small ones. Almost four times more in large ones than in small ones.

But this is not the most revealing data. It's just that within the same company, the dispersion between areas is even greater than between sizes. In any reasonably sized organization I've seen in recent years, sales might be at Step 3, service at Step 4, tax still at Step 1, and HR navigating between Step 1 and 2 — all at the same time, under the same CEO. It's not a lack of coordination. It's the different nature of each role.

The visible case of this in Brazil today is Itaú. By the beginning of 2026, according to the specialized publication TI Inside, the bank had 150 generative AI solutions in production simultaneously — in different areas, at different maturities, with different governance. Generative investment AI, for example, was scaled in five waves over fifteen months: initial testing with 250 employees in November 2024, expansion to 2,500 employees in February 2025, release of the product to 10,000 customers in June 2025, and expansion to 100,000 customers in the second half of the year. Itaú Emps — the application launched in July 2025 that replaces human managers in serving micro and small companies — is an Autonomous Agent (Step 4) in full production. Bank risk areas have been using classic machine learning models since before the generative wave. Other areas are still experimenting.

There is no single "Itaú step". There are areas at Step 1, Step 2, Step 3, Step 4 — and continued investment in organizational memory architecture that orchestrates the whole. It's a real picture of a company that is taking AI seriously.

And the most important lesson for the reader of this article is the way Itaú advances: it didn't try to go all the way up together. Each area advanced in its own time, with its own governance, under its person in charge. What coordinates the whole is not a single schedule — it is a common architecture (memory, policy, decision, cadence) under which each area builds its own path.

The same dynamic happens on a smaller scale, and is even more evident. Conta Azul, a technology company for small and medium-sized companies with around 100 thousand customers, spent six months testing generative AI with one hundred accounting firms before releasing it to the market in August 2025. It started small, with tight governance, in a specific area — the accountant's financial assistant. Meanwhile, other areas of the company operated at other levels, with other dynamics. It is not the privilege of a large corporation; It's drawing. The company that recognizes that each area advances at its own pace organizes itself accordingly. The one that tries a single schedule crashes.

The practical consequence for the manager reading is direct. The question "where is my company in AI?" is the wrong question — it has no answer. The right question is: "What level is each area of my company at, and what is the next move for each one?"

Sales can be in one place. Tax in another. HR in yet another. And that is exactly what is expected.

Your company is not on an AI level. There are five — one per area. And that's not a problem. That's how it's done.

Before getting to the practical matrix that helps you make this diagnosis — which comes in Act 4 — I need to go through the most important truth in this article. The one that destroys the return of any implementation when ignored. The one that costs jobs when addressed poorly. And that no one has the courage to put on the Council's slide.

The Invisible Layer: The Silent Risk That Destroys ROIConversation with the operations director of a medium-sized industry. Ask for help — implemented generative AI eight months ago, zero ROI. The numbers are tough: productivity hasn't gone up, process time hasn't gone down, quality hasn't changed. It invested around R$180,000 in licenses, plus R$400,000 in consultancy. For him, the conclusion is obvious: technology doesn't work for his company.

I ask for permission to talk to the team directly, without him in the room. He authorizes.

Conversation one — senior manager, fifteen years with us. I ask how he uses the tool. Dry response: "I don't use it much. She doesn't understand our way of doing things." I push a little more. In a lower voice, he adds: "Look, if this thing really worked, in two years I wouldn't be here."

The tool works. What doesn't work is what is being fed into it.

The senior manager knows the client, knows the process, knows the way of the house. But when he sits down to configure the agent — to teach the AI ​​to do what it does — he gives a generic answer. It is not conscious bad faith. It's survival instinct. If you describe in detail how you qualify a lead, how you negotiate an exception, how you decide when to accept a late order, you are literally writing the document that can replace it. And when the AI ​​generates a bad answer, it doesn’t fix it — why would it? The more mistakes it makes, the more the company will conclude that "this doesn't work".

This has a name. I call it silent sabotage — not in the sense of bad faith, but of rational behavior by those who feel their jobs are threatened by a technology that the entire market is selling as a substitute. It's the biggest invisible risk of any AI implementation today, at any size company.

Silent sabotage occurs in four modes, all observable with some field effort.

Underreporting is the first way. The professional records less of what he does — just the result, without the path. Seller closes the deal and records "closed sale, R$X", without a history of overcoming objections, without the trigger that made it work, without the exception granted. Without the path, AI has nothing to learn — and critical knowledge continues to live only in the head of those who created it.

Evasive answer is the second. When asked to describe a process for setting up an agent, the practitioner gives superficial description. "How do you qualify a lead?" — "Oh, I'll see if it makes sense, then I'll call." Without the real detail, AI cannot replicate — and the general impression is that "our process is too complex for AI to understand". It is not. The description was missing.

Non-correction is the third, and perhaps the most corrosive. AI generates bad output. The professional knows it's wrong. And does not correct. Let the agent make mistakes day after day, week after week — because with each accumulated error, the company's likelihood of concluding that "this doesn't work" increases. It is rational behavior, and investment devastating.

Shadow AI is the fourth. The professional uses personal ChatGPT on his cell phone, outside the corporate system, for tasks that should go through the company's agent. The productivity gain remains with the person (who will use this as an individual differential). Sensitive company data is leaked to a system that the company does not control. And corporate AI seems "underused" — which reinforces the argument that it doesn't work.

Each of these four modes is understandable. These are rational reactions from those who feel their own jobs are threatened. The solution is not to demand more professionalism from employees — it is to address the perception that drives behavior.

And this perception is not addressed by the tool. Not even for the big tech that sold the license. Not even by the consultant who implemented it. It is addressed by the direct manager. Internal sales is a parallel track to technical implementation, independent, and non-delegable. The professional will only truly collaborate when the direct manager has answered — reliably, repeatedly, demonstrated — three questions that he has in his head but rarely verbalizes: "will this replace me?", "what do I gain from this?", "how will they evaluate me from now on?".

Three practical moves answer these questions and disable the four sabotage modes.First: position AI as an amplifier, not a substitute. The phrase is simple: "let's free you from repetitive work so you can do what only humans can do — relationships, judgment, creativity". But there is one non-negotiable condition: the sentence must be true. If the real intention is to reduce staff, the internal sale will fail — and it deserves to fail. Deliberate lying at this point destroys the team's trust in leadership permanently.

Second: involve the team in configuring the agent. Those who teach feel ownership. Whoever feels dominion does not fear. It's an old principle of change management — John Kotter, in the 1995 classic published in Harvard Business Review, "Leading Change: Why Transformation Efforts Fail", already called it "empower others to act". Applied here, it becomes something concrete: the senior manager who silently sabotaged the agent above becomes the curator of the team's own agent. The relationship with technology changes from day to night.

Third: show individual gains before organizational ones. The team only buys change when it itself wins. "The company will gain 20% efficiency" doesn't move anyone. "You will earn 5 hours a week to do the work that matters" move. Quantify the individual gain first, in each person's mind. Organizational gain comes as a consequence — not as a sales argument.

The most publicly known case of those who ignored this entire layer is Klarna. In February 2024, the Swedish fintech announced that it had replaced 700 human agents with AI. In May of the same year — three months later — he announced a review and returned to hiring humans for part of the volume. The CEO publicly admitted that the aggressive replacement had degraded service quality and the customer experience. It was not the technology that failed. Klarna's AI still works today. What failed was skipping the inside sales layer and skipping the organizational memory layer (which I'll cover in Act 5). Leadership decision to transform AI into a replacement, without mediating the transition.

The biggest risk of AI in your company is not technical. It's human. And it's not overt resistance — it's silent resistance, which destroys ROI without you ever knowing why.

If you understand why this matters, you are ready for the practical map that comes now — which helps the manager to diagnose, area by area, what step they are on today and what the next move is for each one.

The role of the manager: the matrix that guides decisions by area

The area manager does not need to be an AI expert. You need to know, at each step, four practical things:

• What changes in my area when the team uses this level.
• What decision can I delegate to this level.
• What's the risk if I jump this step?
• What a clear sign that my area has risen — and is ready for the next one.

This is what makes the manager lead adoption from his own chair, without becoming an AI engineer. The matrix below is the visual version of this vocabulary — each column is a common area of ​​the company, each line is a step, each cell translates what that combination means in practice.

How to read the matrix

Reading is simple: take your column, find the line that describes where your area is today, and look at the line immediately above it — that's your next move.

Example: you are a sales director. Your column is first. If your salespeople today only use ChatGPT to draft proposals (Step 1, first cell in the column), the next natural move is Step 2 — adopting AI built into CRM. It's not Step 5. It's not Step 4. It's the next one. If you already have CRM with AI (Step 2), the next one is Step 3 — Custom GPT pricing. And so on.

And the rule from Act 2 continues to apply: each area can be in a different place. Sales can be at Step 3 while tax is at Step 1 and service is at Step 4. This is not a problem — it is expected.

Important: The examples listed in each cell are representative of the type, not exhaustive. Each step has dozens of variations by sector, size and maturity. At Step 2, for example, tools for technical sales are different from those for consultative sales; what works for one industry may not make sense in retail; M&A lawyers use different platforms than tax litigation lawyers. The rule for the manager is to understand the type of movement that each step represents, not to memorize the brand of the tool — because the brand changes every semester.### What decision can I delegate to AI?

The question that comes next is practical: what kind of decisions can I delegate to AI in this area of ​​mine? The answer goes through four criteria that must be answered before any delegation.

Codifiability. Can the decision rule be written clearly, with objective criteria? The more codeable it is, the more the AI ​​can perform. The more nuance and unique context it requires, the more the decision remains human — with AI preparing the material.

Volume. How many times does this decision happen per day, per week, per month? High volume justifies investment in automation. Not low volume. Single strategic decision — merger, partnership, critical layoff — is never AI.

Reversibility. If the AI ​​makes a mistake, is the decision repairable? Wrong transaction approval is reversed in minutes. No decision to abandon a strategic client. The more reversible it is, the more autonomy you can delegate. The less reversible, the more humane oversight.

Auditability. Can you later inspect what the AI ​​decided and why? Without an auditable trail, no delegation makes sense — because you'll never know if it's calibrated or not.

Stone, mentioned in Act 1, is an example of Step 4 calibrated in these four criteria: the rule is codifiable (risk score in five levels), the volume justifies it (millions of transactions per day), each individual decision is reversible (wrong blocked transaction is undone in minutes), and the trail of each decision is auditable on the dashboard. Critical decision in aggregate, within the range where AI performs better than humans.

Act 3's Klarna failed at least two of the four criteria: the rule for "generic human service" is not easily codifiable (each case has unique emotional nuance), and the error has limited reversibility (customer lost to poor service returns with friction). It wasn't the technology that failed — it was the caliber that was wrong.

The rule of thumb for the manager: always start with decisions that score well on the four criteria. Move up to bigger decisions only when you have an auditable trail established. And never delegate to AI what your area makes as a unique strategic decision.

Why the manager is the key

Anyone who understands the line of their own area does not need to understand the others in depth. But you need to understand the entire matrix to know where your area connects with others — and where the company, as a whole, is ready for the next coordinated move.

And here we come to the most important part of this article. The rung that appears as the last row of the matrix, the Second Brain, is not simply "the top" — it is a different category. And it is the piece that makes all the previous ones really multiply the value they promise.

Your own path: organizational memory

Different in nature, not in degree

Look at the matrix we just described. Steps 1, 2, 3 and 4 have one thing in common: they all accelerate processes that already existed in the company before AI arrived. ChatGPT speeds up writing that humans were already doing. CRM with AI accelerates sales that humans were already making. Custom GPT accelerates responses that humans already gave. The autonomous agent executes processes that humans already perform.

Step 5 is different. It doesn't accelerate anything that existed — builds something that didn't exist. The company's living organizational memory. Decisions recorded with why. Versioned policy with trace. The auditable cadence of learning from what you do.

This is not "more advanced AI". It's another category of thing. It's organizational architecture — supported by AI, but which would exist as a management concept even if AI weren't good enough to implement it. Today is good enough. That's why we were able to talk about it seriously.

Both valid paths

And here's the part that changes everything in the way you think about this step: it can be constructed as an arrival point or as a starting point.

Path A — arrival point: the company climbs the ladder step by step. At some point — usually between the third and fourth year of using AI — someone realizes that the memory layer is missing. That each previous step is powerful, but does not accumulate. That knowledge keeps slipping out the front door when critical talent leaves. Then it builds Step 5 to coordinate what is already running.Path B — starting point: the company understands the destination from the beginning. Generally because the founder has already experienced, elsewhere, the pain of losing critical knowledge. Decide to build Step 5 on day one, and use it as a spine that receives Steps 1, 2, 3 and 4 as each area enters.

Both paths are legitimate. The second saves two to three years of expensive learning — but requires a combination of technological maturity (clearly understanding what AI can and cannot do well) and prior organizational implementation experience (having already led similar structural transformation before, in any technological wave) that not every company has at the start. Knowing this choice is the first step to making it consciously.

What delivers, in three horizons

Unlike the other steps — which deliver immediate and relatively clear gains — the Second Brain delivers in three distinct horizons. Knowing which gain to expect over which horizon is what avoids the frustration that kills implementations.

Short term (3 to 6 months) — immediate operational gain. Onboarding a new employee shortens from months to weeks because the context is available, not in the head of the person who needs to be in a meeting. Meetings become shorter — no one needs to "remember" why an old decision was made, the system shows. The decision remains consistent because the rule is written and applied equally by everyone. The leader unlocks because he stops answering the same question fifty times.

Maria, hired as a sales manager, joins on a Monday. Instead of six meetings to understand the history of the house's main client, consult the system. In half an hour you have the complete context: why this client has been with the company for seven years, what pricing rule was agreed, what exceptions have already been granted and why. When you sit down with him in the first meeting, he speaks like someone who knows the relationship — because he does, even without having experienced it.

Medium term (6 to 24 months) — structural gain. Citable culture under pressure: in internal conflict, someone opens the original decision and dissolves the debate. Brutal reduction in rework — the company stops repeating mistakes from two years ago. Real standardization: rule applied consistently, not according to the manager's mood. Succession with reduced risk: departure of critical talent does not empty the company. And Steps 1 to 4 gain context to work with — they stop being generic tools and become extensions of the house.

Beginning of 2027. New team wants to give 20% discount to win large customers. System shows the decision recorded a year before: "discounts above 15% require validation from the CFO. Reason: three deals in 2025 with aggressive discount lost margin for eighteen months." A conflict that would last an entire meeting is resolved in a few minutes — based on a recorded decision, not on the opinion of the loudest speaker.

Long term (2 to 5 years or more) — gain in nature. The company learns from its own history, and stops repeating the mistake of three years ago. Speed ​​of organizational change: changing a policy propagates in days, not months, with a trace of who did it, when, why. Immunity to critical turnover: founder can leave without disrupting the business. Culture as defensible architecture, not as folklore that only survives as long as someone still remembers. Intangible asset auditable — relevant in M&A, in funding, in regulatory governance. Strategic decision time drops by 50% to 70% because the context is readily available.

2031. Founder leaves the company after twenty years. Successor takes over with access to every recorded strategic decision — why each important client is with us, what is the reason for each exception, which old rule is still valid, which has become a legacy to review. The company continues without missing a beat. What mattered was never in the founder's head — it was in the architecture of the house.

The concrete difference — company WITH vs WITHOUT organizational memory| Dimension | No organizational memory | With organizational memory |

|----------|---------------------------|-----------------------------| | Onboarding again C-level | 3 to 6 months | 2 to 4 weeks | | Repeated decision | Frequent — company rediscovers what it already knew | Detected before repeating the error | | Conflict over politics | Solved by hierarchy (who shouts the most) | Resolved by consultation (original decision) | | Departure of senior talent | Empties critical knowledge | Knowledge remains in architecture | | Rule change | Spreads in months (and badly) | Propagates in days (with trail) | | Internal audit | Expensive manual reconstruction | Direct consultation | | Due diligence (M&A) | Fragile — buyer discounts value for risk | Demonstrable asset — supports the multiple | | AI applied (Steps 1 to 4) | Generic, disappoints | Contextual, multiplies |

Each row in this table is a concrete operational measure. A company on the right side is cheaper to operate, more resilient to exits, more valuable in transactions, quicker to make decisions. Company on the left side pays repeatedly — in hours of meetings, in redone decisions, in valuation lost in due diligence.

The margin that leaves your company for big tech

One last point, perhaps the most important for those who think about AI from the financial risk and governance side: companies with Step 5 implemented transform the relationship with any AI supplier.

Today, those who depend on a single supplier face the same pattern that we saw with ERP in the 2000s and with cloud platforms in the 2010s: annual price increase without proportional value adjustment, increasing exit costs, progressive margin capture by the platform. It's an old, well-known model — and one that many companies have already experienced in previous technological waves. The difference, in AI, is the speed with which lock-in forms and the impact it has on critical business decisions.

Company with Step 5 breaks this pattern at the origin. The knowledge remains in the company, in an auditable and exportable format. Switching the underlying model — from Claude to GPT-5, from OpenAI to Anthropic, from vendor X to vendor Y — doesn't destroy anything that's been built. The Second Brain is the house's proprietary architecture, powered by tools that come and go over the years.

The practical result inverts the power relationship: the supplier starts to compete for your business, you do not compete to change suppliers. And the margin that today leaves your company, month after month, for big tech — starts to stay at home, where it should always be.

Knowing the destination, however, is not the same as knowing how to get there. The next act — the last practical one before closing — is the implementation plan. Three concrete ways to start Monday, in any of the real Brazilian company scenarios.

Implementation plan: three paths to start Monday

Knowing the map is one thing. Implementing it within the company is another. This penultimate act is the operational plan: three concrete paths, each suited to a type of company, with a real beginning and a defined turning point.

Path A — Structured (top-down, with Board approval)

Path A is for a company with formal governance — family company with an active board, listed S/A, group with private equity, multinational. The starting point is a Council or Board decision, and the typical sequence is:

1. Management, HR and technology area conduct mapping of areas and maturity (using the matrix from Act 4).
2. Structured plan is designed with a horizon of 12 to 24 months.
3. Plan is presented to the Board with budget, KPIs and aligned cultural change plan.
4. Formal executive approval.
5. Phased implementation by area, with person responsible and calendar.
6. Monthly monitoring, quarterly adjustment, scaling according to results.

Time until first measurable result: 6 to 9 months. Cost: structural investment that varies greatly by size — typically from R$100,000 to R$1 million in initial implementation, plus transformation consultancy.

Where an external consultancy helps: structure the mapping, write the executive plan for the Board, model KPIs comparable to industry benchmarks, sustain the pace of implementation over the 12-24 months.

Path B — Educational (bottom-up via leaders)Path B is for companies that want to start before deciding everything at once — medium-sized companies, startups with 30 to 500 employees, innovation areas within large companies. The typical sequence is:

1. Management training on the map (this article is a reasonable start; it can be complemented by a structured workshop).
2. Each manager diagnoses his own area using the step × area matrix.
3. Pilots by area — fast, cheap, measurable in 30 to 90 days.
4. Cross-learning between areas (internal case presented to other areas).
5. Eventually consolidates into organizational architecture (and migrates to Path A).

Time until first result: 30 to 90 days per area. Cost per initial pilot: typically between R$20,000 and R$100,000, plus investment in leadership training.

Where external consultancy helps: train management to understand the map in depth, design low-risk, high-signal pilots, mentor managers in driving adoption within each area.

Path C — Hybrid (recommended for most)

Path C combines the previous two and is the one that makes the most sense for most Brazilian companies with between 50 and 5,000 employees. It starts with Path B (management training and rapid pilots by area) and, when a critical mass of areas shows measurable results, migrates to Path A (formalization of the plan and presentation to the Council).

The decisive advantage is that the plan presented to the Council is not theoretical: it comes with real pilots, field data, documented lessons learned. Executive approval has low political risk, because there is already an internal precedent. It's the most robust path in practice — it reduces approval risk, reduces implementation risk, and maintains momentum from the first month.

In any of the three paths: three non-negotiable guarantees

The three paths are distinct in how, but identical in what needs to stand on any of them. Three non-negotiable guarantees. Failing either one defeats the plan, either way.

Guarantee 1 — Inside Selling as a Parallel Track. Either way, the Act 3 inside selling layer runs in parallel to the technical implementation — never in series afterward. The three movements described there apply continuously: position AI as an amplifier, involve the team in the configuration, show individual gain before organizational gain. Skipping this layer is the fastest way for silent sabotage to destroy ROI six months after the approved investment.

Guarantee 2 — Clear Ownership: three defined roles. Without clarity on who owns what, the project dies in the first territorial dispute. Three roles must be defined from day one:

• Executive sponsor — C-level that defends the budget, prioritizes, responds for the aggregate impact.
• Strategic operator — manager or director who drives adoption in his own area, measures and adjusts it on a daily basis.
• Curator of memory — new role, specific to Step 5. Records decisions, keeps policies updated, keeps the organizational memory alive over time.

In a small company, these roles can be fulfilled by one or two people. In a larger company, there are three different chairs — and the confusion between them is the main reason for institutional paralysis in AI projects.

Guarantee 3 — Security and LGPD addressed: three questions. For any Brazilian company, three questions must be answered before adopting any of the five steps:

1. Where does the data end up? Provider, geographic region, encryption in transit and at rest.
2. Who has access? Auditable logs, segregation by function, output and copy control.
3. How ​​do we respond to LGPD requests? Right of revocation, portability, forgetfulness.

The ANPD published specific guidelines on the use of AI with personal data — it’s worth following. For Step 5, this point is especially critical: organizational memory can include sensitive data and the company is responsible for controlling what is inside it. Compliance is not a technical detail — it is part of the architectural design from the first line.

Where OPEX comes in

If your company wants practical help to lead this journey in any of the three paths, OPEX Consultoria operates on three fronts aligned with this article: strategic advice, organizational memory architecture, and management training. Closing detail.

The map and the manifestoThis article is the map — it describes how it works, where each area of ​​your company is today, what changes when you move up one, how to start Monday on any of the three paths.

The question he doesn't answer is the most important: why does this architecture matter in the long term, beyond the immediate operational gain. Why organizational memory is not an "advanced tool" — it is the only way for a company to learn from its own history in an era where everything is accelerating. Because without it, AI disappoints. Because with it, culture stops being folklore and becomes defensible architecture, quotable under pressure, transmissible between generations of leadership.

This question is the basis of everything described here. And it has an answer in a sister piece that I wrote a few months before this article: the manifesto Memory before Intelligence. Seven acts, seven thousand words, a thesis — memory before intelligence, that’s the order. The manifesto develops in depth why this order is the only one that works in any company, in any sector, of any size.

If you want:

• Understand the practical map and use it with your team — this article is the piece. It can be a starting point for internal discussion with the board or for a workshop with area managers.
• Understand the philosophical reason for all this — the manifesto Memory before Intelligence develops the basic thesis, in seven acts with cases and field tests.
• Practical support for your company at any part of the journey — OPEX Consultoria operates on three fronts: strategic advice on adoption (planning, area mapping, plan for the Board), organizational memory architecture (implementation of Step 5 customized for your sector and size), and management training (enabling leaders to drive adoption in their areas, without needing to become AI engineers).

The CEO's opening question — "have we finally entered the age of AI?" — had a longer answer than he wanted to hear in that meeting. I hope you now have the vocabulary, the map and the roadmap to answer this same question in your company, in your area, in your next move.

Memory before intelligence. That's the order — and this is the piece that shows the map of how to get there, area by area.

Ricardo Segura is founder of OPEX Consultoria. Manifesto "Memory before Intelligence" v1 published in May 2026. Professional contact via LinkedIn.

✅ Full article — Acts 0 to 7 delivered. Suggested next steps: full final review, conversion to LinkedIn-friendly HTML, rendering of the matrix as a PNG image, and planning of derivatives (IG carousel, threads, workshop deck).