Track · AI in Product4 min read

Gartner: 70% of Firms Running Agentic AI Face Material Incidents by 2029

Gartner forecasts material incidents for 70% of firms running agentic AI by 2029. Identity tools cover permission; none cover authority, the gap where CIO, CISO and CRO collide.

AI is blurring the lines between CIO, CISO and CRO
AI is blurring the lines between CIO, CISO and CROStartupStockPhotos / pixabay

Context

  1. Gartner predicts at least 70% of organizations running agentic AI in production will suffer a material service, security or cost incident by 2029

  2. Gartner VP Analyst George Spafford: written corporate policies cannot physically stop an agent from making a destructive error

  3. A 40-minute approval workflow fails in a five-minute decision window

  4. Regulators from the UK SMCR to DORA, NIS2 and the EU AI Act require a named accountable person

  5. The fix names owners for five decisions: protective action, business consequence, residual risk, restart, and approval validity

Gartner predicts that by 2029, at least 70% of organizations running agentic AI in production within infrastructure and operations will suffer a material service, security or cost incident, tied partly to weak runtime controls. The forecast, published in September, frames a governance gap that existing identity and access tools cannot close: they verify which agent acted and what it was permitted to do, but not whether it was entitled to act with these consequences, here and now.

The scenario Gartner's analysis rests on is concrete. At 09:24 on a Tuesday, an AI agent spots suspicious privileged activity, decides containment counts as approval, and shuts off the identity service. By 09:30 the call centre and the payments team cannot log in. Three executives arrive with three instincts: the CIO wants the service restored, the CISO wants it kept down until the threat is proven gone, and the CRO calculates how long the firm can operate before breaching a board-approved risk tolerance.

Vice President Analyst George Spafford states the core problem plainly: "Written corporate policies cannot physically stop an agent from making a destructive error."

What does Gartner's prescription actually solve?

Gartner's recommended controls are conventional and sensible:

  • Give every agent its own identity and an accountable owner
  • Grant access for the task, then revoke it
  • Let the agent propose while a separate control layer decides execution
  • Add circuit breakers and budgets
  • Keep humans in the loop for high-impact or irreversible actions

Run the Tuesday incident through them, though, and the gap reappears. The control layer pauses the agent and asks a human to approve. Which human? A CISO can approve containment, and that approval settles nothing about stopped payments, missed customer deadlines or the regulator's questions. Identity, permission and authority are three different things, and only the first two have tooling.

Where does ownership break?

The seams fall into three moments. Before the agent acts, conflicting authority dominates: a group CISO may authorize containment while a subsidiary CEO objects because it halts a critical service, and if nobody settled precedence in advance, the loudest voice or the clock wins.

During the incident, three failures compound. The person who can order protective action is rarely the one who should accept the cost. Latency destroys process: an approval workflow that takes forty minutes is useless in a five-minute window, and a slow decision in that window is a governance failure dressed up as diligence. And the governance machinery itself often fails with the incident — if the identity provider, approval workflow and chat tool all sit on the platform that is down, the authority chain is out of reach.

Afterward, recovery splits three ways: the CIO declares the platform available, the CISO says it is not trusted, the CRO says residual exposure remains too high. Then drift sets in as models, suppliers and threats change. Using James Carse's distinction between finite and infinite games, AI approvals are the infinite kind: kept alive deliberately, or quietly going stale.

Why do the standard fixes fail?

Each popular fix answers "Who is in charge of AI?" while leaving "Who owns this decision, right now?" untouched:

  • A Chief AI Officer holds the decision without the three lenses
  • An AI committee spreads accountability until nobody can find it
  • Making the CRO the owner, per the IIA's Three Lines Model, removes the independent challenger
  • Folding the CISO under the CIO means the board hears only agreement
  • An agent security platform solves control while leaving authority open

What works instead?

The workable approach splits the incident into decisions, each with a named owner assigned before Tuesday arrives: one person authorizes protective action, one accepts the business consequence, one accepts residual risk, one authorizes restart, one decides whether the original approval stands. Agents get executable mandates — purpose, permitted decisions, cost and consequence limits, escalation triggers, revocation conditions — checked by the control layer for authority, not just access.

The audit trail also needs upgrading. A decision trail records what evidence existed at 09:24, which authority applied, what consequences were expected, who accepted the remaining risk and what would force a rethink. Six months later, the organization should not depend on interviewing whoever was in the room.

One executive must answer for how the decisions fit together, a requirement regulators already encode from the UK's Senior Managers and Certification Regime to DORA, NIS2 and the EU AI Act's human oversight duties. Boards, under the UK Corporate Governance Code 2024, need five answers on one page: what changed, what sits outside appetite, which consequential decisions matter and who owns them, what evidence supports management's position, and what should trigger renewed challenge.

Agentic AI leaves human accountability in place. It removes the time organizations once had to exercise it — and the firms that pre-assign decision ownership, with tested tie-breaks and decision windows, are the ones that will keep millisecond agents governable.

via cio.com (Source)

More from Priya Raman

Priya Raman

Show full bio

News editor covering money, funding and market shifts around product management at Roadmap File.

78 articles