Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Your AI agent has zero data retention and zero accountability

Five essential controls to implement

[Photo: Getty Images]

Profile photo of Sabina Anja

Chief Technologist

Zero data retention has become shorthand for responsible enterprise artificial intelligence: Sensitive prompts go in, answers come out, and the provider keeps nothing.

Companies rely on ZDR to address risks it was never designed to manage. As businesses shift from chatbots to agents that search systems, write records, run workflows, and use external tools, they rely on different privacy and cybersecurity protections in third-party contracts. ZDR governs whether a provider retains prompts and responses. It does not reveal whether an agent misbehaved, made improper queries, violated policies, or was compromised. Yet an enterprise must retain enough information to reconstruct an incident.

For translation, classification, extraction, and summarization, ZDR is a reasonable default. For an agent capable of consequential action, ZDR without a customer-owned audit trail can leave a company private by design and blind by accident.

TWO DIFFERENT PROBLEMS

ZDR addresses whether a provider keeps prompts and responses for monitoring, review, or operations. Under OpenAI’s ZDR, customer content is excluded from abuse-monitoring logs, and storage is turned off on endpoints. Eligibility varies, and some features follow different retention policies.

Agents introduce a second problem: reconstructability. An agentic workflow rarely consists of one prompt and one answer. It may involve an employee’s identity, retrieved documents, authorization policies, approval steps, retries, and an action. The agent might send a message, open a ticket, alter a customer record, change a cloud configuration, or initiate a payment.

When something goes wrong, the enterprise must distinguish among scenarios. Did a user make a request that the system executed badly? Did the model propose something that should have been blocked? Did a tool exceed its intended authority? Or did an attacker use a compromised key or service identity?

A billing record, token count, or “model invoked” event cannot provide the answer. Enabling ZDR does not solve both problems.

WHY AGENTS CHANGE THE SECURITY BOUNDARY

A chatbot can produce a wrong answer. An agent can produce a wrong answer and then act on it. The security boundary changes when a request can invoke a connector, query an internal database, call an application programming interface, run code, modify a record, or send a message outside the company.

The security-relevant event is no longer the text entering and leaving the model. It is the complete chain of decision and execution.

A system should record who initiated the request, the identity and permissions used, the model and policy versions involved, the data retrieved, the tools available, the authorization and approval decisions, the action proposed, the action executed, and whether it was interrupted, retried, reviewed, or rolled back.

That does not require storing every raw prompt and response. It requires telemetry: event IDs, timestamps, authenticated identities, application context, tool names, authorization decisions, approval records, and execution outcomes. These records can be separated from access, protected by an investigation process, and deleted according to a retention policy.

THE CREDENTIAL PROBLEM IS HERE

In 2024, Sysdig documented a campaign called LLMjacking, in which attackers used cloud credentials stolen through a Laravel vulnerability to access AI services. Their script tested the credentials against 10 providers and examined logging configurations to determine how much activity it would record. Sysdig estimated that model consumption could cost a victim more than $46,000 per day.

Cloud audit logs can confirm an Application Programming Interface (API) call occurred, identify the account and time, and record an IP address. But the prompt, output, and action context required separate logging that the victims had not enabled.

The gap becomes serious when a compromised identity belongs to an agent with tool access. A Chief Information Security Officer must be able to determine whether it read sensitive data, attempted a restricted action, sent instructions downstream, or bypassed an approval step.

Treating ZDR as a compliance checkbox makes such incidents harder to detect.

A BETTER DEFAULT

Enterprises face two distinct failure modes. The first is confidentiality: Sensitive content is retained or accessed beyond its purpose. Controls include data minimization, ZDR, encryption, access restrictions, and contractual limits.

The second is accountability: The enterprise cannot establish what an agent did, under whose authority, or why. Controls include customer-owned audit records, strong identity, policy enforcement, approval gates, and incident response.

The better principle is to retain less while keeping necessary evidence under the enterprise’s own governance. ZDR should be applied according to a workload’s autonomy and potential consequences. A translation service handling confidential documents is a strong candidate because the task is bounded and does not alter enterprise systems.

An agent that updates customer relationship management records, begins payment workflows, or changes cloud configurations requires a different posture. Provider-level ZDR can remain in place, but the enterprise should create a privacy-preserving, tamper-evident event record before granting the agent authority to act.

Five controls are essential:

1. Use an enterprise-controlled gateway that associates every request with an authenticated user, workload identity, tenant, policy, and risk classification.

2. Give tools narrowly scoped, short-lived credentials and explicit permissions. Avoid broad service accounts with access across business systems.

3. Record the model’s proposed action separately from what a tool actually executed.

4. Require human approval for irreversible or high-impact actions. Provide a way to stop the agent and record denials, overrides, retries, and rollbacks.

5. Preserve the minimum evidence necessary for an investigation: identity, authorization, policy version, tool invocation, result, and approval state. Redact sensitive fields, encrypt records, make them tamper-evident, and limit retention according to risk.

After an incident, can the enterprise reconstruct the chain of events without asking the model provider to recover data that was deliberately never retained? If not, the deployment may be privacy-conscious, but it is not operationally mature.

For agents, “we retained nothing” is not sufficient. The defensible standard is to minimize raw content, control access to the evidence that remains, and preserve enough customer-owned telemetry to establish who acted, with what authority, what the system attempted, and what happened.

Privacy without auditability solves only half the problem.

Sabina Anja is a chief technologist, executive director, and open-source entrepreneur.