Skip to the stories
Microsoft 365 & SharePoint

AI Governance Needs Enforcement, Not Just Another C-Suite Title

  • Pam Baker
  • 925 views

Authority over AI is consolidating at the top while day-to-day governance spreads across teams. Enforcement at execution time is where the gap closes.

AI Governance Needs Enforcement, Not Just Another C-Suite Title

Authority over AI decisions is moving up the org chart, while the work of governing AI is moving out into the business. That split is the real problem. Executives hold the sign-off but often don't control the systems that act on it, and the people who run those systems rarely have the authority to stop them. Adding another title doesn't fix that.

Titles are multiplying faster than accountability

IBM's 2026 CEO study found that 76% of surveyed organizations now have a chief AI officer (CAIO), up from 26% in 2025. Proposed titles such as chief AI agent officer and chief regulatory officer are also appearing, which points to more C-suite expansion. Here's the thing, though: the same study found that 79% of the executives surveyed are decentralizing decision-making and distributing accountability as AI takes on a bigger role across the enterprise. A title on a org chart doesn't mean the holder owns the outcome.

The chief governance officer title is even less settled. In corporate governance circles, the CGO traditionally handles board oversight, and the Society for Corporate Governance's 2025 survey found the function sits inside the legal department or general counsel's office at 84% of companies it surveyed. South Korea's SK Telecom uses the same title for a role inside its AI governance structure. Same label, different job. If you're trying to map responsibilities across vendors, peers and board packs, that mismatch causes real confusion.

For CIOs, the useful question isn't who gets the next title. It's who can make increasingly distributed controls work as a single system.

Policy documents don't stop agents

Agentic AI breaks the old model. A written AI policy sets expectations, but it can't independently stop an agent from reading restricted data or executing a transaction nobody approved. Interviews with CIOs, CTOs and governance specialists point to a shift: AI governance is moving out of policy documents and committees and into the operational systems that enforce decisions.

Bapuji Koraganti, a cybersecurity engineer at a large American mass media, telecommunications, entertainment and tech company, describes the disconnect bluntly. "In most enterprises, AI governance lives in two places that rarely talk to each other: a policy PDF owned by a committee, and a set of OAuth scopes, service accounts, and API keys configured by whoever shipped the integration." In his view, the consequential approval often happens when someone grants credentials, not when a governance committee signs off on deployment.

That matters more with agents than with conventional software. Agents can work across applications, reach sensitive information and start transactions without a person in the loop for each step. The credential grant effectively becomes the authorization, whether or not anyone intended it that way.

Authorize the transaction, not just the agent

Patrick Parker, chief innovation officer at identity governance provider Omada, argues for a sharper distinction. Authorizing an agent to use a financial system is not the same as authorizing a specific transaction. An agent with access to finance tooling shouldn't automatically be able to change supplier banking details or release payments.

Parker's point is that these limits have to be enforced when the action happens, not just described in policy. "An enterprise can have a complete library of AI policies and still lack the wiring between governance decisions and enforcement," he said.

In practice, that means moving away from blanket approvals and attestations toward enforceable permissions, transaction limits, monitoring and the ability to revoke an agent's authority quickly. Parker put it this way: "Sensitive actions need authorization enforced outside the model, at execution time. Where approval is required, it must bind to the actual beneficiary, amount, destination or data scope. Approval to pay one supplier cannot become permission to pay another simply because the same agent and tool are involved."

That's the detail most governance frameworks skip. A approval scoped too loosely becomes a standing permission, and standing permissions are exactly what agents exploit by accident.

Insurers and contracts are starting to ask questions

Governance is also becoming a financial and contractual issue. Insurance brokerage Gallagher reports that some insurers are using tier-based pricing and adjusting coverage limits based on applicants' answers to AI governance questions. Its spring 2026 commercial insurance report says deficiencies in AI risk management readiness are beginning to affect pricing.

Don't read that as a universal surcharge. Insurers' approaches vary, and general cyber pricing trends don't always match how individual AI exposures get treated. Still, the incentive is clear: be able to demonstrate that governance controls exist and actually run. CIOs also shouldn't assume stronger governance automatically means lower premiums. Insurers price specific exposures, policy terms and coverage limits, and governance is one input among several.

David Gildea, VP of product for generative AI at cyber resilience provider Druva, compares this to earlier shifts in cyber insurance. "Much like requiring MFA, insurance companies are starting to look at evidence of protections for AI in place and not just attestation. That includes proof that an organization can recover its data and systems if an AI agent or AI-driven attack causes damage, just as insurers began asking for proof of tested backups during the rise of ransomware." If that pattern holds, the questions will move from "do you have a policy" to "show us the recovery test."

Procurement and legal teams feel the same pressure through vendor contracts. Allison Spagnolo, senior managing director, deputy general counsel and chief privacy officer at security services firm Guidepost Solutions, points to data-use provisions, audit rights, liability allocation and contractual protections around AI capabilities that change over time. Those terms are where an enterprise either takes on undue risk or pushes it back to the vendor.

Savio Lobo, CIO at managed services provider Ensono, goes a step further. He argues AI vendor agreements belong inside enterprise risk management, not in a separate bucket of isolated technology purchases. Procurement runs commercial negotiations, legal sets contractual terms, security and privacy define risk requirements, and business or technology owners decide whether the leftover risk is acceptable. Contract language alone still won't stop an AI system from taking an unauthorized action, though. Those obligations need technical controls and operational oversight behind them.

Lobo's summary of the vendor side is worth keeping: "The important thing is that AI vendor review cannot become a separate universe from third-party risk management. It needs additional AI-specific questions, but it should plug into the enterprise controls."

Who actually holds the authority?

Ensono offers one working model. Lobo works alongside a CTO and a CAIO. The CAIO, supported by the CIO and CTO, sets strategy, standards and adoption priorities. Privileged access, contractual liability and business process automation stay with the executives who own those areas. Lobo's view is that "creating a new C-level title for every emerging issue is unlikely to solve AI governance," and that new titles won't erase existing accountabilities either.

The harder case is a company without that structure. As AI decisions move into individual business units, responsibility gets split in ways that leave gaps. A business leader approves an automated workflow. Security controls its permissions. Technology runs the environment. Legal sets the obligations. Nobody necessarily outranks the others, and when a problem surfaces, the lack of a single accountable authority can become a serious issue on its own.

Anthony Habayeb, CEO and co-founder of AI governance platform Monitaur, places ultimate accountability with the CEO and board, since they set the organization's risk appetite. His answer to who owns AI accountability is direct: "Can one executive own AI accountability for the whole company? No. Can a CAIO orchestrate it, with risk and audit verifying it? Yes."

He adds that the right structure ends with an accountable business executive saying yes or no, informed by experts across technology, risk, security and other disciplines who each own specific governance controls. That's a useful test for any org chart. If you can't name the person who says no, and the people whose work informs that call, the structure isn't finished.

A common operating model, not another executive

Tom Jenkins, vice president of strategic partnerships at decision intelligence platform RelationalAI, says there's no single operating model for AI governance. His company is seeing more federated approaches, with standards set centrally while business units keep responsibility for their results. That fits the organizational reality, but it creates a technical problem. How do you ensure the same policies, business rules and constraints apply across different applications and agents?

Jenkins sees the CIO, or a chief data and AI officer, increasingly responsible for the common architecture that makes distributed governance workable. Lobo makes the same point from the operations side. He recommends a unified control plane that brings together AI inventories, risk classifications, named owners, identity standards, monitoring and operational evidence. "Then the individual functions can innovate inside those boundaries," he said.

Measure whether controls work

Counting approved policies or governance meetings tells you very little. A better question for CIOs is how many AI systems have accountable owners, meet their access requirements, and can show their controls are working. Parker suggests a metric worth tracking: governance latency, the time between an authorization or revocation decision and its effective enforcement across production systems. A short latency means governance actually controls AI behavior. A long one means governance exists mainly on paper.

The next phase will likely produce more executive titles, and the reporting lines will vary. Some enterprises will widen the CAIO's remit. Others will put coordination under the CIO or spread authority across established technology and risk leadership. Either way, many organizations will be tempted to solve the problem by redrawing the hierarchy.

That usually backfires. A new role without budget, decision rights and operational authority to settle conflicts just adds another participant to a fragmented approval process. The more consequential shift is already underway: governance is becoming an operating discipline. The open question for enterprises is who has the authority to keep all those moving parts working together, and whether that authority reaches the systems where AI actually acts.

Source: Pam Baker · www.informationweek.com

Neighbouring cells