What the Board Will Ask About Your AI Agents
C-level corner · Executive brief
Sooner or later the board, a regulator or an auditor will ask three questions: who is this agent acting for, what was it allowed to do, and can you prove what it did? If the answers rely on the agent's instructions or on a log nobody can join together, they won't hold. Treat agents like any other workload, with controls enforced outside the agent and one trail of evidence.
The business problem#
AI agents decide on their own which tools to use and which data to read. That makes them useful, and makes “who did this?” harder to answer. In Europe it is no longer only a security team question. The technical guide sets out three overlapping regimes:
- The EU AI Act. By the guide’s timeline, prohibited practices have applied since February 2025, obligations for general-purpose AI models since August 2025, and the full regime for high-risk AI systems from August 2026. The EU has since proposed moving some high-risk deadlines later, so check the current dates. The Act includes risk management, automatic logging and human oversight. The guide puts penalties at up to EUR 35 million or 7% of global annual turnover.
- DORA. For financial institutions, AI supporting decisions, risk assessment or customer interaction falls under operational resilience rules, including incident reporting and oversight of AI providers.
- NIS2. For essential and important entities, significant incidents must be reported within 24 hours (early warning) and 72 hours (full notification), and senior management must approve and oversee cybersecurity risk measures, including those governing AI.
All three ask for the same things in different words: risk management, logging and audit, incident reporting, supply chain oversight and access control.
What changes#
The central principle: you cannot manage an agent through its instructions. An agent will follow whatever it is told, including instructions hidden in a malicious document. Data it must not reach has to be blocked by access controls. The policy lives in the platform, not in the prompt.
In plain terms, the guide describes five layers:
- Network. The agent can only reach approved destinations, and only approved callers can reach it.
- Identity. Every action traces back to a verified identity: the agent’s own, or the person it acts for.
- Permissions. Enforced by the data platform: if a manipulated agent asks for restricted data, the platform refuses. The OWASP Top 10 for AI explains how agents get manipulated.
- Data protection. Encryption, limits on what can leave, and human approval before data is exported.
- Audit. Every login, query and tool call is logged in a trail that cannot be quietly altered.
No single layer is enough. Together they make misuse hard and contain the damage.
One blind spot is worth naming. Security monitoring may collect the database queries an agent runs, but it sees a query from a user: not that an agent planned it, what the person asked, which tool it chose, or what it read next. For an investigation, that context is the whole point. Agents don’t need a new security stack. They need to show up in the one you already run.
What it’s worth#
- Lower regulatory risk. One set of controls maps to the logging, oversight and access control duties of all three regimes, instead of a separate answer for each.
- Faster incident response. When an investigator can follow a conversation from the question asked to the data touched, an agent stops being an invisible user.
- Faster approvals. Agents that inherit these controls pass review sooner. The case is made in Keep Shipping.
Decisions for leadership#
- Make “on whose behalf does this agent act, and how does it log in?” a required answer before any agent goes live.
- Rule that restrictions on agents are enforced by access controls, never only by instructions to the agent.
- Require agent activity, not just database queries, to reach the security monitoring you already run.
- Name an owner who can show the board, on request, what any agent did and for whom.
Questions to ask your team#
- For each agent, can we name the person or service it acts for?
- If an agent were tricked into asking for restricted data, what would stop it?
- Can our security team see what an agent was asked and which tool it chose, or only the queries it ran?
- Are AI defaults on our platforms open or closed, and where does AI processing physically happen?
Technical background, for your architects
- Governing AI Inference in the Data Cloud: A Security Architecture Guide for EMEA Enterprises: the regulatory mapping and the five layers in depth. Platform-agnostic principles with Snowflake examples.
- Cortex Agents in Microsoft Sentinel: Closing the AI Blind Spot in Your SIEM: how to get agent activity into the security monitoring you already run. Snowflake- and Microsoft Sentinel-specific.
- Snowflake Cortex AI -- Governance, Audit & Inference Region Guide: defaults, audit coverage and where AI processing runs. Snowflake-specific.
Figures and dates come from the linked articles and their sources, as of when they were written; check them, and the current status of each regulation, against your own situation. Snowflake-related articles explore the art of the possible and are not official Snowflake solutions or endorsed by Snowflake.