- Inteligência Artificial
- Governança de IA
- Liderança Técnica
- Requalificação Profissional
- Futuro do Trabalho
- Automação
AI Governance in Practice
Empower People, Don't Lay Them Off
Every week, another company announces it has cut part of its team "because of AI". Almost as often, another one freezes AI usage because the token bill scared someone in finance. These are opposite reactions, but they share the same flawed premise: that AI is, above all, a way to spend less on people. Since June 2026 I've led the AI Enablement area at InBot, and what I see day to day is the opposite. In this post I explain how AI governance works in practice here, why it exists to upskill the team rather than shrink it, and where you can start with your own team.
AI to Cut People or AI to Empower People? What I Heard at Lovable Day
On October 16 I attended Lovable Day São Paulo, and two talks summed up what I had already been advocating inside the company.
The first came from Carol Paiffer, investor and Shark on Shark Tank Brasil. She said that when she sees an entrepreneur wanting to lay off the team to replace it with AI, she sees a bad move.
The reasoning is simple: if you keep the team and upskill it, the company won't do the same with fewer people. It will do far more than it does today, with the same people. It's a gain that compounds, instead of a saving you collect only once.
The second came from Guilherme Junqueira, founder of Delta Academy, who also argues for training companies and their teams to use AI instead of replacing them. I'll come back to his talk in the next section, because it touches the other fear: cost.
Why laying people off to replace them with AI is a bad move
I agree with both of them, for two reasons that go hand in hand.
The first is that the team is who knows the business. A language model knows a lot in general and nothing about how your customer thinks, the exception support handles every week, or the shortcut sales uses to close a deal. At InBot, it's the teams themselves who build the skills and automations. AI only delivers what it delivers because someone with domain knowledge is telling it what to do.
The second is the multiplier effect. An NBER study of customer support agents found an average 14% productivity gain from generative AI, reaching 34% among the least experienced. In other words: AI doesn't replace people who are starting out, it shortens their learning curve. That's upskilling, not cutting.
And the data from those who went the other way isn't encouraging. An Orgvue survey found that, among leaders who laid people off after adopting AI, 55% admit they made the wrong call. Klarna, which became the poster child for "replacing human support with AI", started hiring people for support again after its own CEO acknowledged that the focus on cost led to lower quality. The company didn't abandon AI: it walked back the overreach. And Forrester predicts that half of the layoffs attributed to AI will be quietly reversed.
At InBot, nobody left because of the automations. The time they freed up became delivery. But for that to happen, we had to face the second fear.
Token Cost vs. the Cost of Inertia: Measure Before You Fear
Guilherme raised a point I hear often: the fear of token costs. The AI bill grows, someone sees the number, and the natural reaction is to cut usage. His counterpoint was the cost of inertia. When a company steps away from the tool because of cost, it saves now and pays later, because it stops learning while the rest of the market keeps learning.
I see token cost in two ways at the same time.
- As an investment, not an expense. When a 3-hour task drops to 13 minutes, the right comparison isn't "how much did it cost in tokens", it's "how much did it cost in tokens versus what those hours of a person's time cost". Three hours that become 13 minutes pay for a lot of tokens.
- As something to be measured, not feared. AI governance also means knowing what each thing costs. Fear without numbers leads to bad decisions in both directions: cutting what was paying off, or letting run what wasn't.
There's one more data point in favor of not freezing: AI prices drop fast. Stanford's AI Index 2025 shows that the inference cost for a model at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024. Epoch AI estimates that the price for the same performance falls between 9 and 900 times per year, depending on the task.
That doesn't mean the total bill will shrink, because we start using better models and doing more with them. It means that freezing usage today because of today's price is deciding based on a number that's already going stale. What doesn't go stale is how far behind a team falls once it stops learning.
Measuring, however, requires something organized to measure. That's where governance comes in.
AI Governance in Practice: Teams Build, the Area Makes It Last
"Governance" usually sounds like bureaucracy: committees, forms, approval queues. That's not what we do. AI enablement, the way I see it, means creating the conditions for the whole company to use AI well, not centralizing AI usage in a single team.
How the area started
InBot's AI area wasn't born from a decree. It grew bottom-up. Before it existed, I convinced leadership to adopt Cursor, which at the time was a new tool used mostly outside Brazil. Until then, we did everything by hand. I brought the idea to the other devs, including a more experienced colleague who was wary of using AI at work.
Over time, the dev team itself started standardizing how we used skills so everyone was on the same page, because each person was doing it however they thought best. That's when the board decided to create the AI area: to take this initiative, which started inside the development team, to the rest of the company, with centralization and governance.
Notice that the whole story is about empowering the people who were wary, not replacing them. The area is a continuation of that.
The path to production
Today, on AI matters, the leads of Sales, Dev, Marketing, Planning, Projects, QA, UX and Support report to the area. The flow works like this:
- The team builds. The idea and the authorship belong to whoever has the problem and knows the domain.
- The area brings it up to company standards. We review, adjust and test with scenario suites. We've run suites of 800 automated tests and 250 CRM runs with 250 successes. If you've read my journey with TDD, you know I don't trust anything that hasn't been tested.
- The requesting team validates. Whoever asked for it confirms it solves the real need.
- The area approves the release. Only then does it go to production.
This design avoids two extremes. If everyone uses AI their own way, untested and with no owner, you get shadow AI: AI running in the dark, outside any control. The Microsoft and LinkedIn Work Trend Index showed that 78% of AI users bring their own tools to work, and only 39% had received any training from their company. If, at the other extreme, everything depends on a central team to build it, you create a bottleneck and take authorship away from the people who understand the problem.
The infrastructure that removes friction
Governance that depends on spreadsheets and goodwill doesn't last. That's why a large part of the area's work was building infrastructure. The main example is the Skills Hub: it distributes skills to Claude Code and Codex on Linux, macOS and Windows, with corporate SSO, group-based access, reviewed publishing and signed releases.
In practice, this turns governance into a product. Publishing a new skill isn't opening a ticket and waiting. It's following a flow that already has review, access control and versioning built in. Frameworks like the NIST AI Risk Management Framework and the ISO/IEC 42001 standard describe well what AI governance needs to cover. The challenge is doing it without friction for the people who are building.
But governance isn't only saying "yes, with standards". Sometimes it's saying "no".
Governance Also Means Saying "No": Shadow AI, Security and Process
Two cases illustrate this part of the job well, and neither of them ended with a skill in production.
When the risk is too high
Some teams wanted a skill that could interact with corporate email. The idea makes sense from a productivity standpoint: read, summarize, reply. After talking it through with the operations director, we decided it was too risky and turned the idea down.
The reason has a name. Simon Willison calls it the lethal trifecta: the combination of three things in a single AI agent, namely access to private data, exposure to untrusted content, and the ability to communicate externally. An inbox has all three. Or, in his words:
"An attacker can literally email your LLM and tell it what to do!"
Put simply: a single incoming email can become an instruction to your agent.
This isn't theoretical. EchoLeak (CVE-2025-32711) showed how data could leak from Microsoft 365 Copilot through a single email, with zero clicks from the victim. OWASP itself ranks prompt injection as the number one risk in LLM applications and recommends least privilege and human approval.
Turning down an idea like this isn't being against AI. It's what keeps a single skill from becoming an incident. IBM's Cost of a Data Breach 2025 report points out that high levels of shadow AI add an average of US$ 670,000 to the cost of a breach, and that 63% of breached companies had no AI governance policy. Gartner predicts that more than 40% of organizations will suffer an incident tied to unauthorized AI by 2030.
When the problem is the process, not the delay
The second case is subtler. Someone asked for a requirements-gathering skill to speed up the kickoff of client projects. When we looked closely, the problem wasn't the delay. It was the process itself: the way we collected information to start a project wasn't working well.
A skill on top of a bad process would only make the bad process run faster. So, together with the Planning and Projects teams, and led by the AI team, we built a requirements-gathering platform: InBoarding, which is currently in pilot.
This may be the most useful governance lesson I've learned so far: before automating, ask whether the problem is speed or design. AI accelerates what already exists. If what exists is wrong, it accelerates the mistake.
What Changed Without Anyone Leaving (and What the AI Maturity Assessment Says)
In the area's first four months, we shipped 77 initiatives and put about 50 skills and automations into production. Some concrete examples:
- The admin API operator turned a 3-hour task into 13 minutes.
- A bot content package dropped from about 35 to 13 to 16 minutes.
- 1,300 of a client's employees were imported by a skill.
- InSuporte generates AI insights, approved by humans, on more than 1,800 support tickets, and replaced 2 dashboards and 4 automations.
In none of these cases was the result "now we need fewer people". It was "now this person does in 13 minutes what used to take 3 hours, and spends the rest of the time on something else". That's exactly the compounding gain Carol talked about. The World Economic Forum's Future of Jobs Report 2025 points the same way: 85% of employers plan to prioritize upskilling their workforce, and 59 out of every 100 workers will need training by 2030.
A validation, not a finish line
At Lovable Day, Delta Academy offered participating companies a free AI maturity assessment. I filled it out on behalf of InBot and we got a perfect score, along with a comment that stuck with me: InBot already does, at a corporate level, what they teach their mentees.
I share this carefully, because this result isn't mine. It belongs to the people in every area who stopped what they were doing to learn, build and test new things. The AI area organized the path, but the teams are the ones who walked it.
And it's a validation of the path, not of the finish line. There's still a lot in pilot, a lot to measure better, and a lot that will change with the next models. The score says the direction is right. It doesn't say the work is done.
Where to Start Tomorrow: 3 Steps for Your Team
You don't need an AI Enablement title to get started. The story of InBot's AI area began with one person insisting that the team try a new tool. If I had to sum up what worked in three steps, they would be:
- Turn your personal usage into something for the team. That prompt you repeat every week can become a versioned skill the whole team uses. Many of InBot's internal systems were born exactly this way. If you want to see how I work with AI day to day, I wrote about it in my AI development workflow.
- Map what already exists and give it an owner. List the AI automations your team already uses, including the ones nobody has mentioned out loud. For each one, define who owns it, how it's tested and who approves changes. That's the minimum viable governance.
- Before automating, ask whether the problem is the process. If a task is slow because it's poorly designed, automating it just makes the problem faster. Sometimes the right answer isn't a skill, it's redesigning the process.
Conclusion
The debate about AI in companies tends to get stuck between two fears: losing your job and losing money on tokens. Both fade when the question changes from "how much can we save with AI?" to "how much more can our team do with AI?".
AI governance in practice is what makes that second question answerable. It gives authorship to the people who know the problem, ensures that what goes to production has been tested, knows how to say "no" when the risk is high, and recognizes when the problem isn't about AI at all. None of this requires cutting people. It requires empowering them.
If you lead a team, or you're just the person who uses AI the most on it, start with the mapping: what does your team already do with AI today, and who owns each thing? The answer will probably surprise you. I've written before about the impact of AI on professions; this post is the practical side of that conversation, told by someone applying it. In upcoming posts, I want to show in more detail how some of these skills became internal systems.
References
- VibeDays. "Lovable Day São Paulo", 10/16/2026.
- Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey. "Generative AI at Work". NBER Working Paper 31161.
- Orgvue (via HRreview). "AI regret: over half of firms say their redundancies were 'a mistake'", 2025.
- Entrepreneur. "Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO", 2025.
- Maginative. "Klarna Dials Back its AI Customer Service Strategy, Now It's Hiring Humans Again", 2025.
- Forrester. "Predictions 2026: The Workforce Muddles Through Ambient Disruption", 2025.
- World Economic Forum. "The Future of Jobs Report 2025".
- Stanford HAI. "AI Index Report 2025".
- Epoch AI. "LLM inference prices have fallen rapidly but unequally across tasks", 2025.
- Microsoft and LinkedIn. "2024 Work Trend Index".
- IBM. "Cost of a Data Breach Report 2025".
- Infosecurity Magazine. "Gartner: 40% of Firms to Be Hit By Shadow AI Security Incidents", 2025.
- NIST. "AI Risk Management Framework".
- Willison, Simon. "The lethal trifecta for AI agents: private data, untrusted content, and external communication", 2025.
- Reddy, Pavan; Gujral, Aditya. "EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System". arXiv, 2025.
- OWASP. "LLM01:2025 Prompt Injection".
- Mierzwa, Matheus. "Minha Jornada com TDD".
- Mierzwa, Matheus. "Meu Fluxo de Desenvolvimento com IA".