Most enterprises I talk to have invested heavily in AI governance over the past two years. Model cards, risk assessments, pre-deployment reviews, compliance dashboards. The documentation is thorough. The approval workflows are real. Yet, no one can answer the question what their AI agents actually accessed in production last week, under what policy?
That gap — between governing the program and governing the execution — is the most consequential blind spot in enterprise AI right now.
Two different problems that often get conflated
Lifecycle governance is about the AI program: how models are selected, tested, approved, monitored, and retired. It answers questions like: Was this model risk-assessed before deployment? Did it go through the right review committees? Is it documented for regulators? Tools like model cards, governance registries, and compliance platforms live here. This work is necessary and most mature organizations have some version of it.
Runtime governance is different. It answers a different set of questions: What data is this agent accessing right now? On whose authority? Is that consistent with the policy it was approved under? Every time an AI agent retrieves data — from a database, a document store, an API — that is a governance moment. The agent is making a decision. Something should be verifying that decision is permitted before the data enters the context window.
Most enterprises have the former. Almost none have the latter.
Why the gap exists
In traditional software, governance at runtime meant access controls — role-based permissions, ACLs, audit logs from the database layer. That infrastructure exists and works reasonably well.
AI agents break those assumptions in three ways.
First, agents act autonomously. There is no human reviewing each retrieval before it happens. The agent decides what to access, accesses it, synthesizes it, and acts on it — often in a single chain of calls that happens in seconds.
Second, agents are polymorphic in their access. A single agent might touch a CRM record, a regulatory document, a real-time pricing feed, and an internal compliance database in the course of answering one question. Traditional access controls govern each of those systems separately. Nobody is governing the agent's behavior across all of them simultaneously.
Third, the LLM context window is opaque after the fact. If a customer complaint arises three weeks from now about a decision your AI made, can you reconstruct exactly what data the model saw at the time of that decision? In most architectures today, the answer is no.
What runtime enforcement actually looks like
The mental model I find useful is the difference between a compliance policy and a compliance control.
A policy says: "Agents handling customer PII may only access records relevant to the current transaction." That is lifecycle governance — it belongs in a governance registry, a model card, an approval workflow.
A control says: at the moment that agent makes a retrieval call, the system intercepts it, checks the policy, allows or denies the specific access, and writes a tamper-evident record of that decision with the exact policy version that governed it. That is runtime governance. The policy only matters if the control exists to enforce it.
The EU AI Act's high-risk enforcement provisions take effect August 2, 2026. "We have a governance framework" is not the same as "we can show you a cryptographic audit trail of every data access decision this agent made." Documented policies don't satisfy demonstrable runtime controls.
The practical implication
If you are leading AI programs in a regulated industry, I'd encourage you to ask two questions of your current governance stack:
One — where exactly does enforcement happen? Not where is the policy documented. Where does the system actually intercept a data access and make a decision about it?
Two — if a regulator asked you tomorrow to produce the complete audit trail of a specific agent's data access over the past 90 days, what would you hand them?
The answers tell you whether you have lifecycle governance, runtime governance, or both.
Both matter. Lifecycle governance without runtime enforcement is a policy that relies on good intentions. Runtime enforcement without lifecycle governance is a control system with no coherent policy behind it. The organizations that will navigate this period well are the ones building the connection between the two.
Anil Solleti is the founder of AutoPIL, a runtime governance platform for agentic AI in regulated industries, and a partner at VibrantCapital.ai.