Data Governance Before Deploying AI Agents: What to Classify First

Data Governance Before Deploying AI Agents: What to Classify First

Before giving an AI agent read access, classify data by sensitivity and name an owner for each set. Here are the four levels and the practical order.

•

Want to apply this in your company?

Book a call with Juan →

Data governance for AI starts with a direct question: which company information can an agent read, who owns it, and what happens if it is exposed? Before you connect any agent to your CRM, spreadsheets, email or WhatsApp history, classify each data set by sensitivity and record a responsible owner for it. That classification decides what the agent can see, what requires human approval, and what stays out of reach.

Key takeaways

  • Classify data by the impact of exposure, not by the system where it is stored.
  • Every data set needs a named business owner with a role and the authority to decide who gets access.
  • Personal and sensitive data follow their own rules, aligned with LGPD, before they enter any automated flow.
  • Reading is a permission granted per flow: start with the minimum needed for one specific task.
  • Without access logs and periodic review, governance becomes a document nobody consults.

Why classify before you give the agent access

Many companies start with the agent. They pick a tool, connect the customer database, and only then ask what it should be allowed to do with that data. That order inverts the risk. A support agent that reads the full conversation history can answer a question about delivery times well, but it can also repeat a health detail the customer mentioned in passing, or reveal a discount negotiated with another client. The problem is not the technology on its own. It is the absence of a rule about what may be read.

We treat classification as part of mapping the process, before any pilot. It does not require a multi-year project. It requires knowing, for each flow, which sources are necessary, who answers for them, and what the consequences are of a reading or answering error. This is the same reasoning that supports a well-organized base, as we describe in Data Ready for Superintelligence: What to Standardize First.

The four levels that fit most companies

A simple model is usually enough to begin. It does not need to be perfect, but every area has to apply it the same way.

  • Public: information already published, such as the catalog, opening hours, the return policy posted on the website and official FAQs. The agent can use it freely.
  • Internal: procedures, price tables meant for the team, service scripts and manuals. The agent may read them, but should not reproduce them word for word to a customer.
  • Confidential: proposals under negotiation, margins, customer-specific commercial terms, contracts and financial data. Access is restricted by role, and as a rule the agent receives only the field needed for one specific answer.
  • Restricted: sensitive personal data as defined by LGPD, information protected by professional confidentiality, credentials, and data about children and adolescents. These do not enter automated flows without a specific design, legal review and, in many cases, a decision not to automate. For decisions that affect the data subject, it helps to know what the law provides in LGPD automated decision-making and the right to review.

Credentials are not business data. Passwords, API keys and tokens should never sit in a knowledge base, a prompt or a shared spreadsheet. They belong in a secrets vault and should reach the agent only through a tool with a narrowly defined permission.

Who owns the data and why it matters

Data without an owner becomes a problem at the first question. When the agent gives a wrong answer about a warranty, someone has to say whether the rule was wrong, outdated or misread. That person is the data owner: the person in the business area who is accountable for the information, decides who may access it, and approves changes.

Usually the owner sits in the area that generates the data: sales for the customer portfolio, finance for collections, operations for inventory, legal for contracts. IT is typically the technical custodian, not the owner of the content. Keeping these roles separate prevents the decision to expose a data set from falling to someone who does not know what it means.

Record the owner in a simple table with the name of the data set, the system where it lives, the responsible person, the sensitivity level and the date of the last review. This table is the first governance artifact. It can live in a spreadsheet, as long as it has a responsible person and a routine for updates.

Step by step: from inventory to first access

This is the sequence we usually follow before a pilot. It works for companies of different sizes, though the detail changes with the number of systems involved.

  • List the processes the agent must handle, naming the expected result of each one, such as confirming a delivery date or booking a visit.
  • For each process, list the data sources needed to answer well, and mark the ones that would be unnecessary.
  • Assign an owner to each source, with name and role, and ask them to confirm its sensitivity level.
  • Classify each source at the agreed levels, and archive or remove whatever has no use in the chosen process.
  • For each source, define the minimum fields the agent may read. If the answer depends only on the delivery date, the agent does not need the full address.
  • Separate what is read-only from what requires human approval, such as changing an order, granting a discount or promising a deadline.
  • Record who requested access, when, for which flow, and with whose approval.
  • Test with fictitious data or in a controlled environment. Only then connect a real source.

The principle of least privilege and deny by default, described in OWASP's authorization guidance, supports this list: every permission needs a reason and an owner.

Illustrative scenario: a service company with scattered data

Illustrative scenario: a small service company wants an agent on WhatsApp to report the status of work orders. The data sits in a scheduling system, in a quotes spreadsheet kept by finance, in folders of inspection photos, and in the sales team's conversation history.

When classifying, the team notices that the status of an order is internal, but the final status can be shown to the customer. The quote is confidential, because it carries margins and a negotiated discount. The photos may include people and addresses, which calls for separate handling. The sales history mixes personal remarks with operational information and needs filtering before any use.

The practical decision is this: the agent reads only the order status and the expected date, with no access to the quotes spreadsheet. The photos stay out of the initial scope. When a customer asks for prices, the agent passes the conversation to the responsible person with a summary, without repeating internal numbers. None of these cuts depends on whether the tool is good or bad. They are business decisions made before any code is written.

How our method applies

Our method has four stages: map the process and its baseline, limit data and actions, pilot a real flow with records and an agreed success metric, and then transfer the operation and decide whether the company expands, adjusts or stops. Data classification belongs to the second stage, but it depends on the mapping, which shows which sources actually exist. When Be Human runs the operation, access records and owner reviews continue as routine work, not as a document filed after delivery.

Which references to use as a base

General frameworks help you name risks without inventing rules. The NIST AI Risk Management Framework organizes AI risk management into governance, mapping, measurement and management functions, which gives structure to the owner and level table. For access control, OWASP's authorization guidance reinforces principles such as denying by default and granting the least privilege necessary.

In Brazil, ANPD is the reference body for guidance on personal data protection. It is worth following its publications before deciding what counts as sensitive personal data in each case. When data is sent to providers outside the country, the analysis should also consider the requirements described in International data transfer under LGPD when using AI providers.

Frequently asked questions

Do I need to classify all company data before starting? No. Classify the data for the processes the agent will touch in the first flow. The rest comes in as new flows are chosen, which avoids a classification project that never ends.

Who should approve the classification? The owner of each data set approves the level and the fields released, with support from IT and, when sensitive personal data is involved, from legal or privacy staff. When approval rests only with IT, the business area loses control over what its own data means.

Is read-only access enough? Not necessarily. Even reading a confidential data set can leak into a customer answer. For that reason, read permission should come with rules about what may appear in a response and with a record of what was consulted.

Can I use a spreadsheet for the data inventory? Yes, as long as it has an owner, a review date and a clear classification criterion. The format matters less than the routine. Over time the inventory may move into a dedicated tool, but it should not wait for one to get started.

Sources

AI governanceData governanceLGPDAI agentsData classification