🛡️ Define AI Governance
Governance brings together agent identity, permissions, responsibilities, auditing, logs, cost controls, approved environments, accessible data, prohibited actions, human approval, security policy, information retention, and incident handling.
🆕 New here?
Governance it isn’t bureaucracy: it’s the set of decisions that lets you say "yes" safely. Without it, each team decides on its own what AI can access — and the organization discovers its rules after the first incident. Audit is being able to answer, after the fact, who did what, when, and with what authorization.
✓ Governance that works
- ✓Fits on one page and is read by the people who will use it
- ✓Says what's allowed, not just what's forbidden
- ✓It’s verifiable: the control exists in the system, not just in the text
- ✓Has an owner and a date of last review
✗ Governance in name only
- ✗A 40-page document no one has opened
- ✗Generic prohibitions that the team works around
- ✗A rule with no corresponding technical control
- ✗Written by someone who has never used an agent
Enables
Doesn’t just prohibit
Verifiable
Control in the system
Short
The one that’s read is the one that matters
With an owner
And a review date
🪪 Give the agent an identity and permissions
An agent running with someone's personal credentials makes auditing and accountability impossible—and grants access that no one reviewed. The agent is a system user, and it must have its own identity, minimal scope, and audit trail.
Action log (illustrative recreation, not an actual screenshot)
10:02 agente-ci-bot · read 14 files · repo: pagamentos
10:03 agente-ci-bot · ran “npm test” · output attached
10:05 agente-ci-bot · opened PR #4821 · 3 files
10:06 agente-ci-bot · DENIED: access to secrets/ (out of scope)
10:31 maria.souza · approved PR #4821 · cost: 0,42 USD
How to read: the yellow line is the most important — the denied attempt shows that the scope is real. And the last line closes the loop: an identified person approved it, with the cost recorded.
Identity
Never your own
Minimum scope
Default is deny
Visible denial
Prove there’s a limit
Traceable
Even the person who authorized it
💰 Manage cost, data, and environments
Three controls that are often remembered too late: how much you can spend, what data can enter the context, and which environments the agent can run in.
Capped, visible cost
Set limits per task and per team, with an alert before they’re exceeded. Track the cost per completed task — it’s the metric that supports or sinks the investment at the next budget meeting.
Data: what never goes into the context
Personal data, secrets, regulated information. This is a privacy and compliance issue, not a technical detail—and the response needs to be written down before anyone needs it in a hurry.
Authorized environments
A short, explicit list of where the agent can run. Production is excluded by default, and adding it to this list is a recorded decision, not a side effect of configuration.
Limit
By task and team
Cost/task
Management metric
Sensitive data
Outside the context
Production
Out by default
🪜 Learn the five levels of autonomy
This scale is the shared language that replaces vague discussions about “whether to use AI.” The right question is: for what kind of task, at what level?
How to read: each step brings the control that supports it. You don’t "choose" level 4— you reach it once you have gates, rules, a sandbox, and auditing in place. Level 5 appears in red because it’s a limit, not a goal.
By task
Not by company
Control first
Then the level
Level 5
Limit, not target
Plain language
Ends vague debates
📈 Level up with evidence
Autonomy granted in a burst of enthusiasm is taken back at the first scare — and then the team loses even what was already working. Promote a type of task based on the numbers.
📊 What to measure before promoting
- • Approval without correction: how many deliverables passed review without rework?
- • Attributable incidents: how many production issues came from this workflow?
- • Review time: did it go down or up since the agent joined?
- • Cost per task delivered: is it stable, going up, or going down?
- • Correct stop rate: when it didn't know, did it stop and ask?
💡 Practical tip
Promote in a limited pilot: one repository, one task type, for one month, with metrics agreed on before starting. And make it clear that moving down a level is a normal decision—not a failure of the team or the technology.
Measure before
To promote
Pilot
Limited in scope and short
Correct stop
Underrated metric
Downgrade
Valid decision
🧾 Build the governance checklist
One page that answers seven questions. It’s the artifact that lets you bring AI use to other teams without repeating the entire discussion every time.
🧪 Copyable exercise — your team’s checklist
Objective: move beyond theory with a document ready to share. Fill it out and take it to the person responsible for the system.
# Governança de agentes — <time / sistema> Dono: <nome> · Última revisão: <data> 1. QUEM é o agente? <identidade, onde as credenciais vivem> 2. O QUE pode acessar? <repos, ambientes, dados — e o que NUNCA entra no contexto> 3. O QUE faz sozinho? <lista curta: tarefas reversíveis e verificáveis> 4. O QUE exige aprovação?<deploy, migração, secrets, infra, dados de produção> 5. QUANTO pode gastar? <teto por tarefa e por mês; quem é avisado no estouro> 6. O QUE é registrado? <ações, comandos, custo, aprovações — e por quanto tempo> 7. QUEM responde? <pela mudança entregue e pelo incidente causado> Nível de autonomia atual, por tipo de tarefa: - correção de CI ......... nível <N> - feature pequena ........ nível <N> - migração de banco ...... nível <N>
How to verify: show the page to someone outside the team and ask them to answer: "can the agent modify the production database?" If they can’t find the answer in thirty seconds, the document isn’t ready yet.
Quick check: your team wants to raise automatic CI fixes from level 3 to level 4. What justifies it?
Seven questions
One page
Level by task
Explicit on the page
Discoverable
In 30 seconds
Owner
Without it, it doesn’t exist
📌 Module Summary
Next track:
Track 6 - Practical exercise: now you run the complete workflow in a real repository.