The Governance Confidence Gap

AI agents are the shiny thing right now. The hype says the enterprise has them under control — deployed, governed, audited, safe.

The data says otherwise.

A recent OutSystems report (via CIO Dive), surveying 858 IT decision-makers, put numbers on the gap:

• 83% of firms still govern AI agents through their existing IT, security, or compliance processes — the same processes built for software that sits still and waits for instructions.

• Only 5% have automated controls, audit trails, and formal human review points in place. All three. Five percent.

• Only about a quarter can handle conflicts between AI agents — what happens when two autonomous systems disagree about what to do with your data.

Read that again: the overwhelming majority of companies have handed autonomous systems the ability to act — read inboxes, move information, trigger workflows — while governing them like they're still just software licenses. An agent that can message your clients on your behalf is being overseen by the same change-control process that approves a SaaS renewal.

Where I stand

My rule with anything new: learn it, watch how others adopt it, examine whether it fits or just adds noise, then adopt. That cycle takes hours or months. We don't just add things.

With agents, I've run the cycle. The verdict is adopting — but only behind approval gates. And here's the reframe: the gate isn't overhead. In intelligence work, no asset acts without tasking authority and a human in the loop. The approval gate is the control surface. It's the product.

What good looks like

The 5% have three things, and they're the same three things I build for clients:

1. Automated controls — the agent can't exceed its mandate. Permissions are enforced by the system, not by a policy document nobody reads.

2. Audit trails — every action logged, attributable, reviewable. If you can't reconstruct what the agent did, you don't govern it.

3. Formal human review points — a person signs off before anything irreversible happens. Not after. Before.

For the families and executives I work with, the exposure is concrete: data pasted into consumer AI tools is now a flagged leak vector. The fix isn't to ban the tools — it's private infrastructure, compartmentalized AI, and approved tools with human review before anything sensitive moves.

The 83% aren't wrong to move fast. They're wrong to confuse motion with control.