We use cookies to optimize your experience, analyze traffic, and personalize ads. Please choose whether you accept our use of non-essential cookies. Read More
In the Panorama Vibe Coding Brasilsurvey, conducted by Skyone with the news, governance appears as the first of three gaps that prevent the safe scaling of AI in companies (the other two being technical knowledge and IT infrastructure). The numbers illustrate the scale: among the 1,982 respondents in the total database, only 26.4% confirm having clear and formalized rules on how to input data into AI tools. Among those who create solutions in the dominant profile, 60.9% say they have no clear guidelines on data, while 52.1% develop automations and 49.5% build internal integrations.
The challenge for those leading in technology, security, and finance is twofold. If the company restricts usage, it loses productivity and pushes teams towards shortcuts, as discussed in the previous article in this series. If it allows unrestricted access, it exposes corporate information, assumes regulatory risks, and loses control over spending. AI governance exists to resolve this tension without sacrificing either side.
This text details what constitutes applicable governance, what technical controls support it, how to divide responsibilities, and how to structure it to enable its use rather than create barriers. The data comes from a survey with our own sample, which reflects the group consulted and not the entire set of Brazilian companies.
AI governance is the set of policies, roles, controls, and processes by which an organization defines how artificial intelligence can be used, by whom, with what data, and under what supervision, as well as how this use is monitored and corrected.
The concept is often confused with two neighboring aspects. Information security protects assets and data against unauthorized access. Compliance verifies adherence to laws and regulations. Governance is the layer that decides who is responsible, why, and how the other two apply to the specific case of AI. This distinction is the focus of this article.
Speed matters because adoption doesn't wait. According to the overview, 22% of Brazilian companies surveyed already use AI or LLMs in their development cycles, and more than 80% intend to increase their AI and automation budget in the coming year. When usage spreads faster than the rules, each new use case emerges without criteria and needs to be corrected later, which costs more.
The central risk is losing control over where the information goes, who can access it, and how it will be handled. A contract with sensitive clauses, a compensation spreadsheet, the source code of a product, or a customer database can be sent to an external service in seconds by someone who had no intention of causing harm.
The figure of 26.4% with formalized rules shows that, in most of the sample, this decision still depends on individual judgment. By difference, 73.6% do not confirm having clear rules. And the risk is not limited to chat: when half of those who create it develop automations (52.1%) and internal integrations (49.5%), the information also starts circulating between systems, without each flow having been evaluated.
In the case of personal data, the LGPD (Brazilian General Data Protection Law) applies, and the company is responsible for its processing, including when done by employees using non-approved tools. Sending personal data to an AI tool must have a defined purpose, basis, and protection, and cannot depend on the goodwill of each user. Specific cases should be evaluated with the legal department and the data protection officer.
There is also a sign of awareness: among those who create more than one type of solution, 27.3% spontaneously express concern about leaks and security. This is an indication that the issue gains importance as maturity increases.
An effective AI usage policy answers practical questions in language the user understands. It typically includes:
The most common mistake is writing the policy as a legal text. A short document with concrete examples (for example, "it can summarize minutes of an internal meeting; it cannot be sent to the customer database") is more likely to be read and applied.
Politics without control depends on individual discipline. Four technical fronts support what was decided on paper.
Who can use which models and assistants, and with what permissions, should be managed as in any enterprise system. The point of attention is the AI's connection to email, document repositories, CRM, and other tools: each integration expands what the assistant can read or do. The governance question is who authorizes the connection, with what scope, and how it is revoked. Connections with individual authentication tend to keep the AI's access within what each person can already see, but this needs to be verified in each implementation.
In Skyone Astro, connections are made through connectors, individual OAuth, and dedicated MCP servers, with administration controlling permissions and connections. In practice, the IT team decides which tools can be integrated with AI and monitors what is connected, instead of leaving each user to configure it on their own. For those building integrations and automations, the landscape also presents Skyone Studio, which centralizes connections with iPaaS and Lakehouse and provides governance from the source.
Without records, there's no way to investigate an incident or understand how AI is being used. Auditing allows you to consult prompts and responses, filter by user, period, or topic, and, when necessary, determine what happened. There's a precaution to take: this monitoring deals with employee data and must have a clear purpose, restricted access to records, and transparent communication with users, in alignment with the legal department and the data protection officer.
This is the feature that Skyone Astro offers in the audit layer: command logging and querying of prompts and responses with filters and CSV export, in addition to supporting investigations when it is necessary to ascertain a fact. The export is useful for security, legal, and compliance areas that work with their own analysis tools.
Automatic rules (guardrails) check prompts and sent files and can block sending or simply issue an alert, depending on the severity. A practical approach is to start with alerts, observe false positives, and only then tighten restrictions to block the highest-risk categories. Blocking everything from day one tends to generate frustration and workarounds.
In Skyone Astro, guardrails act on prompts and uploads and can operate with blocks and alerts, allowing for the progression described above. The overview adds that the platform anonymizes sensitive data (DLP) before requests reach the models.
The use of AI models is usually charged based on processing volume, making costs variable and difficult to predict without monitoring. With over 80% of surveyed companies planning to expand their AI and automation budget, visibility into where money is being spent is no longer a detail. User limits, alerts when approaching a ceiling, and consumption reports allow finance to see the investment and leadership to compare costs and results by team. This is the aspect that usually brings the CFO closer to the governance conversation.
Skyone Astro displays usage in Brazilian Reais (BRL), with user limits and alerts, making it easier for those responsible for the budget to track their investment. Skyone describes two subscription models: a standard model for small and medium-sized businesses and an enterprise model for large companies, with detailed conditions outlined in a commercial proposal.
Responsibility is shared, and the lack of a clear owner for each part is one of the most common causes of governance that exists only on paper. One possible division:
| Paper | Primary responsibility |
|---|---|
| IT and technology | Platforms, integrations, access management and connections |
| Information security | Risk analysis, data protection rules, incident response |
| Legal and data protection officer | LGPD, contracts with suppliers, monitoring limits |
| Finance and controlling | Budget, limits and monitoring of consumption |
| Area leaders | Use cases, quality criteria, and human review on your team |
| Users | Follow the guidelines, report questions and incidents |
In practice, it's worthwhile to appoint a person or committee to coordinate the other areas and make decisions in case of conflict between speed and control. The current situation shows that overloaded IT is one of the factors pushing users towards shortcuts, which reinforces the need to distribute the workload and not concentrate it in a single area.
The principle is proportionality. Low-risk cases, such as rewriting a public text or summarizing non-sensitive internal material, deserve a fast track. Higher-risk cases, such as analyses involving customer data or decisions affecting people, deserve review and additional controls.
Three practices help maintain balance. First, offer an approved and accessible alternative, because rules without an option for use push people back to personal tools. Second, track simple indicators (number of exceptions requested, alerts generated, consumption per team, incidents) to adjust rules based on data. Third, combine rules with training: 52.7% of the total user base points to a lack of technical knowledge as the biggest obstacle, indicating that rules only work if people know how to apply them. Skyone associates this approach with Builders by Skyone, an executive education and practice program.
A corporate AI platform can bring together, in a single environment, the production resources of teams and those of administration. Skyone Astro is an example of this approach. The overview describes it as a corporate AI Gateway: a single portal where the company controls employee access to multiple LLMs, with compliance taking precedence over free experimentation. According to Skyone's reference material, for administrators it offers prompt and response queries with filters and CSV export, guardrails for prompts and uploads with blocks and alerts, consumption control with values in reais (Brazilian currency), limits per user and alerts, as well as connections via connectors, individual OAuth and its own MCP servers, with permission control. For teams, the environment provides access to multiple models and assistants.
It is prudent to treat this as a starting point for evaluation, not as a guarantee. No platform eliminates the risk of data exposure on its own, nor does it produce automatic compliance. Infrastructure, data retention, and operational requirements need to be verified on a case-by-case basis, and the evaluation should use synthetic or authorized data, with the controls that the company committee deems indispensable. Human review of critical decisions also remains the organization's responsibility.
The overview from Skyone and the news helps to define the problem: the use of AI and automation is growing, and investment is keeping pace, but only 26.4% of the sample confirms formalized rules regarding data. Effective governance combines concise policy, technical controls (access, auditing, data protection, and consumption), and defined responsibilities in IT, security, legal, finance, and leadership. When well-calibrated, it allows the company to say "yes" to more use cases with known criteria, instead of relying on generic prohibitions.
Want to understand how to structure AI governance in your company? Discover Skyone Astro and Skyone's content on corporate AI, and evaluate the approach with a real-world example from your operation.
It is the set of policies, roles, controls, and processes that defines how artificial intelligence can be used in an organization, by whom, with what data, and under what supervision. It includes rules for use, access control, auditing, data protection, cost tracking, and assignment of responsibilities.
Information security protects data and systems against unauthorized access and use. AI governance is broader: it defines which uses are permitted, who is responsible for them, and how security, privacy, quality, and cost risks will be addressed. Security is one component of governance.
Authorized tools, permitted and prohibited data categories, permitted and prohibited uses, requirement for human review in critical decisions, those responsible for questions and incidents, and review frequency. Concrete examples by area increase buy-in.
It depends on the purpose, transparency, and proportionality of the monitoring, and should be evaluated with the legal department and the data protection officer. Good practices include clearly communicating with users, restricting access to records, and using data only for the stated purposes, such as security and incident investigation.
By setting limits per user or team, configuring alerts when the limit is approaching, and monitoring consumption reports by period, regular monitoring allows you to correlate spending and results before expanding access.
Transform Your Business with Skyone. Request a demo or schedule a call with our experts to discover how Skyone can accelerate your digital strategy.
Have a question? Talk to a specialist and get all your questions about the platform answered.