
AI Agent Logs as Evidence in Customer Disputes: What to Keep
Which AI agent conversation and decision logs are worth keeping in a customer dispute, how to protect privacy under LGPD, and when to ask a lawyer first.
Want to apply this in your company?
Book a call with Juan →AI agent logs can serve as evidence in a customer dispute when they show, with a reliable timestamp, integrity and context, what was said, which rule was applied, which data was consulted and who approved each decision. Not every record is worth keeping, and keeping everything creates privacy risk. The practical recommendation is to decide in advance which records you keep, for how long and with what access, and to consult a lawyer before using any excerpt in a negotiation or a legal proceeding.
Key takeaways
- A useful log records a decision, not just text: input, output, the rule applied, the data consulted and human approval when there is one.
- Keep what creates a commitment to the customer, such as a delivery date, a quoted price, a warranty condition and consent, and discard what has no clear purpose.
- Editable screenshots carry little weight as evidence. System-generated records with a timestamp and the version of the configuration are more consistent.
- Personal data inside logs is still personal data, and Brazil's LGPD (general data protection law) applies to it.
- Before citing a log for or against a customer, take the question to a lawyer.
What a record needs to contain to work as evidence
An AI agent record works as a set of events. Observability practitioners describe this kind of data in three complementary forms: logs, which are timestamped events, aggregated metrics, and traces, which link a sequence of steps to the same operation. For a dispute, what matters is usually an event with enough context to reconstruct what happened.
In practice, a useful record of a conversation that involved a decision should contain:
- Date and time recorded by the system, with the time zone, not typed in afterward.
- An identifier for the conversation and the customer, preferably in pseudonymized form.
- The version of the configuration in force: instructions, business rules, available tools and the model used.
- Actions taken, such as a query to an order system or a message sent, with the result returned.
- Decision points: what the agent resolved on its own, what was escalated to a person and who approved it.
Without the configuration version, you cannot tell whether the agent was answering under today's rules or those from several weeks earlier. That detail often decides the argument.
Which records deserve to be kept and which do not
A good rule is to keep what proves a commitment or a decision, not the entire raw content of every conversation. Commitments made by the agent, such as a stated deadline, a quoted amount, a warranty condition or authorization to use data, should have their own durable record. Automated decisions that affect the customer, such as refusing an order or applying a billing rule, also deserve retention, together with the justification and the possibility of human review. The same applies to escalations and changes to customer records made by the agent.
Conversations with no consequences, raw audio with no purpose, test attempts and fields with sensitive data that the process does not need should have a short retention period or not be stored at all. Keeping everything may seem prudent, but it increases the volume of personal data, the surface exposed to a leak and the effort of finding a record when one is actually needed.
Privacy: keeping proof without turning the customer into a file
Privacy care starts with the question of purpose. Before recording a field, you need to know what it is for and who will need it. The LGPD is built on principles such as purpose, necessity and security, and Brazil's National Data Protection Authority publishes guidance worth reviewing. Each company should turn these principles into written operational rules.
In practice, this means masking fields such as full tax ID numbers and phone numbers when the log does not need them, restricting access to the people who actually investigate incidents, recording who viewed or exported the records and setting a retention period with a deletion procedure. It is also important to separate the technical log, used for operations, from material that might be attached to a dispute, because they often do not have the same retention period or the same custodians.
If the conversation involves recording, transcription or third-party data, the care required increases. This topic is covered in more detail in Recording Meetings With AI: Consent and Privacy Law (LGPD). When a customer asks for a review of an automated decision, the subject connects to what LGPD Automated Decision-Making: The Right to Review (Art. 20) covers.
Integrity: how to show that a log was not altered
A log that anyone with server access can edit quietly loses its weight. Several measures help: store records in a medium that does not allow silent changes, compute a cryptographic hash of exported files, keep the exported copy with its export date and maintain a history of who accessed what. The NIST AI Risk Management Framework treats governance and traceability of AI systems as part of risk management, which helps justify these choices internally.
None of these measures alone guarantees acceptance in court or in a negotiation. Because the way to demonstrate integrity depends on the case, the type of system and the dispute, this point should be assessed with a lawyer before any use.
Illustrative scenario: a dispute over a deadline
Illustrative scenario: a services company uses an agent on WhatsApp to inform delivery deadlines. A customer says they were given a different deadline and demands compensation. What helps the company is a set of records: the exact message sent, the version of the instructions in force on that date, the query made to the order system with its time and the response received, and the record showing that the conversation was not escalated. What weakens the company's position is an edited screenshot taken on a phone, a log with no configuration version or an export made weeks later with no record of who produced it. Whether to use these records and how to present them is a decision for the company's lawyer.
When to consult a lawyer before using a log
The role of whoever operates the agent is to organize the evidence and the operation. Legal analysis is a different kind of work. Consult a lawyer before using any record in the following situations:
- Before citing a log in a formal notice, a written negotiation or a response to a formal complaint.
- When the record contains personal data of third parties or sensitive data.
- When the customer contract does not define who keeps logs, for how long and under what conditions.
- When the customer asks for access, correction or review of an automated decision.
- When your sector has its own rules on document retention or communication with consumers.
This text is operational guidance and does not replace a legal opinion. On questions like these, a lawyer also helps decide what should not be disclosed or sent to the customer.
How this fits into our working method
In our method, deciding how to keep logs comes at the start, not after an incident. First we map the flow and what is already recorded today, because many platforms already store part of the data. Then we limit data and actions so the agent only accesses what it needs. Next we pilot a real flow with defined records and an agreed metric, and we review with the client which records were actually useful. When the operation is handed over, the retention policy, access rules and export procedure are documented, and the decision to expand, adjust or stop the use rests on that data. For teams defining rules for using AI tools, a good starting point is Company Policy for Employee ChatGPT Use: What to Allow and Ban.
Frequently asked questions
Are AI logs accepted as evidence? It depends on how they were generated, kept and presented. Automatic records with timestamps, access control and configuration context tend to carry more weight than editable screenshots. The assessment belongs to a lawyer, case by case.
Do I need to keep every WhatsApp conversation? No. Keeping everything increases privacy risk and makes it harder to find what matters. The safest approach is to keep the records that prove commitments and decisions, with a defined retention period and restricted access.
Can I use a log without telling the customer? That depends on the contract, the legal basis that justifies processing the data and the LGPD requirements for your specific case. It is not advisable to decide this without legal guidance.
How long should I keep logs? There is no single period for every case. Define retention by purpose, by any legal obligations in your sector and by the period in which a dispute could arise. Record that decision in a written policy and review it with your lawyer.