Two numbers, read together, describe most of the current state of AI in the enterprise. Gartner projects that 40% of enterprise applications will have AI agents embedded in them by the end of 2026. And separately, 92% of security leaders say they lack full visibility into what their own AI agents are doing — more than half of deployed agents run with no security oversight or logging at all, and the average enterprise now runs dozens of them. (Source: 6clicks, summarizing Gartner's GRC software research.)
Put plainly: AI adoption is running well ahead of AI oversight. Most organizations can tell you an AI agent imported some assets, scored a risk, or drafted a policy — they generally can't tell you exactly what it saw, what it changed, or who approved it, in a form that would survive an auditor asking "show me."
Automation was never the hard part
GRC platforms have been automating parts of compliance work for years — importing evidence, flagging control gaps, drafting first-pass risk scores. None of that required much governance of its own, because a human was still doing the actual work with the automation as a shortcut. AI agents change that relationship: they don't just assist a task, they propose an outcome — an asset entry, a risk score, a gap finding — that, left unreviewed, looks indistinguishable from something a person decided.
That's the actual audit and trust concern showing up in the numbers above. It's not "is the AI accurate enough" — it's "can you prove, after the fact, that a specific human looked at this specific AI output and approved it, with the reasoning intact." Most tooling wasn't built to answer that question, because it wasn't built with AI agents generating a meaningful share of the record in the first place.
What an actual answer looks like
This is the design question we started from when we built Govexa's AI agents — not "how do we automate more of a GRC program," but "how do we let AI do real work without any of it becoming a fact until a named person says so." Three agents, each scoped to one job, all sitting behind the same rule:
- Asset Agent— reads spreadsheets and existing exports, proposes structured, attribute-complete inventory entries, and flags anything it isn't confident about instead of guessing.
- Risk Agent— scores risks against the formula and thresholds your team configured, and proposes treatment options. It doesn't set the formula, and it doesn't close a risk on its own.
- Gap Analysis Agent— compares your current controls against a target framework and drafts a prioritized list of what's missing.
Every output from every agent is a proposal, not a record. It sits in a review queue with its reasoning attached until a named person on your team approves, edits, or rejects it. Nothing reaches the compliance or audit record unreviewed — and the approval itself, along with the AI's original input and reasoning, is logged in the same append-only audit trail as everything else. If an auditor (or a regulator, under something like the EU AI Act or ISO 42001) asks "who approved this and on what basis," the answer already exists; nobody has to reconstruct it from memory or Slack history.
What to ask any GRC platform selling you "AI features"
- When the AI proposes something, is there a real review step, or does it write directly to the record?
- Is the AI's reasoning — not just its output — logged and retrievable later?
- Can you tell, for any given record, whether it originated from an AI proposal, and if so, who reviewed it and when?
- Does the audit trail cover AI actions with the same rigor as human actions, or is it treated as a separate, lower-priority log?
The 92% figure isn't a talking point — it's a description of what happens by default when AI capability ships faster than the audit trail around it. Closing that gap isn't a matter of adding more AI. It's a matter of making sure every AI action has a name attached to the decision that followed it.