AI Governance for Commercial Teams: The Operating System, Not the Policy PDF

An approved AI tool can still return data its user should never see. Access can fail where one system hands work to another. IBM's 2026 breach research makes that a concrete governance problem: enforce the controls across the workflow, then test that they hold.

This opens a six-part series for commercial teams running AI in production. The published articles cover access, review, logging, audit scope, and vendor data handling. A written policy helps define the rules; the workflow must enforce them.

What the breach sample tells us

IBM's 2026 Cost of a Data Breach Report studied breaches at 602 organizations globally between March 2025 and February 2026. IBM's accompanying analysis supplies the AI-incident comparisons below. These are findings from breached organizations, not a measure of risk at every company.

AI-related incidents averaged 5.33 million dollars, against 4.70 million for breaches with no AI involvement or where the organization was unsure. IBM also reported unapproved-AI incidents at 43 percent of organizations, compared with 20 percent in its prior benchmark. Regulatory fines were reported in 21 percent of those incidents.

These comparisons do not establish that a written policy prevented a breach, or that unapproved AI caused the difference in average cost.

What the incident data actually shows was missing

Among organizations that experienced an AI-related security incident, 92 percent lacked proper AI access controls at the time. That identifies a control gap in those incidents; it does not tell us which policies the companies held or how effective those policies were.

Only 19 percent of breached organizations reported meaningful coordination between governance and security teams. The report does not establish why that coordination is missing. Name who must resolve shared access, data, review, and incident decisions.

Agent credentials need named owners, restricted access, and a way to revoke them, just as employee credentials do. The connector article works through those access boundaries.

Governance as four surfaces, not one document

Turn the policy into specific controls that teams can inspect and test. These four surfaces provide a practical way to organize that work.

Access: what the system can reach and who approved it. Give the workflow the fields it needs; record what stays disconnected and why.

Review: can the reviewer inspect the evidence and stop a consequential action? A zero stop rate alone proves nothing about quality. Test a seeded bad recommendation, record the decision and reason, and inspect what happens downstream.

Logging: join the model identifier, instructions, retrieved context, output, review, and downstream action in one trace. Pinning a model helps identify what ran; it does not make a later LLM rerun deterministic.

Audit scope: identify whether the AI layer is inside the formal audit and what systems, data, people, and controls that includes. Vendor review supplies the data-destination and retention questions.

The regulatory floor is arriving whether the mechanism exists or not

Article 50 of the EU AI Act, the transparency obligations, took effect August 2, 2026, per the European Commission's own guidance. Providers of AI systems that interact directly with people must disclose that the interaction is with AI unless that is already obvious. Providers generating synthetic audio, image, video, or text must mark that content in a machine-readable, detectable format. Deployers using emotion-recognition or biometric-categorization systems must disclose that to the people affected, and deepfake content must be labeled clearly at first exposure. A limited grace period runs to December 2, 2026 for the content-marking obligation specifically, on systems already on the market. Penalties reach 15 million euros or 3 percent of global annual turnover, whichever is higher.

Article 14 sets a separate bar for human oversight on high-risk systems: a human overseer must be able to understand the system's capabilities and limits, stay alert to the tendency to over-trust its output, and be able to override or reverse a decision. Article 14(5) specifies separate verification by two qualified people for the remote biometric identification systems in Annex III point 1(a), subject to its stated exceptions. Other high-risk obligations and deadlines must be checked for the specific system.

These duties depend on the system, use, provider or deployer role, and applicable dates. Four operating surfaces do not establish compliance with the AI Act. Identify the obligations that apply to the deployment, then retain evidence that the required controls work.

Two frameworks, and why neither one is the whole answer

Two primary references serve different purposes. ISO/IEC 42001 specifies an AI management system that can be certified through an external assessment. NIST's AI Risk Management Framework is voluntary guidance for managing AI risks. Choose how to use them against the organization's obligations and actual systems.

Neither is a substitute for the four surfaces above. ISO 42001 can certify that a management system exists around AI. It does not, by itself, prove your access controls are enforced or that your review gate has ever stopped anything. The frameworks are the scaffolding a mature program hangs on. They are not the program.

Why marketing cannot outsource this to security

Commercial teams need to work with security and legal on the workflows they operate. IBM's 19 percent coordination finding identifies a reported gap; it does not prove that a split between functions caused incidents.

Spencer Stuart's December 2025 CMO survey, covered in depth in The Last CMO, found AI strategy and implementation most often led by the CTO or CIO, at 44 percent of companies, with the CMO at 32 percent and the CEO at 19 percent. Those figures describe leadership assignments, not measured coordination between governance and security. Whether a single seat should hold both the commercial and the AI mandate is the argument in who should own AI: CAIO vs CTO vs CIO. The role decision still needs an agreed mandate; these operating controls need named owners whichever title leads the work.

The series

All six published articles:

AI Governance for Commercial Teams: the operating system, not the policy PDF. This piece.
What Actually Enters SOC 2 Scope When You Automate With AI: the audit-scope surface. Published September 8.
The Human Gate That Isn't a Rubber Stamp: the review surface. Published September 10.
What to Connect and What to Refuse: the access surface, MCP connectors. Published September 15.
The AI Audit Log: the logging surface, reconstructing why the model said that. Published September 17.
AI Vendor Review: where your data actually goes. This closes the series.

The discipline underneath all six is the same one behind AI Commercialization: the complete guide. A functioning model and an enforced control answer different questions. Use the articles to test the control in the workflow.

Frequently asked questions

What is AI governance for a commercial team?

AI governance connects written rules with enforced controls. For a commercial team, inspect data and model access, reviewer authority and evidence, decision traces, and formal audit scope. Name the owners and test those controls against the actions the workflow can take.

Why do AI governance policies fail to prevent incidents?

A policy can define a control without ensuring it is implemented. IBM found 92 percent of organizations with AI-related security incidents lacked proper AI access controls at the time. The report also found only 19 percent of breached organizations coordinated AI governance and security. It does not establish the preventive effect of policies or compare policy ownership across incident groups.

What is shadow AI and how big a problem is it in 2026?

Shadow AI is employees using tools the organization has not approved. IBM reported unapproved-AI incidents at 43 percent of breached organizations, compared with 20 percent in the prior benchmark. Regulatory fines were reported in 21 percent of those incidents. Make approved tools usable, restrict unauthorized access, and investigate shadow use.

What does the EU AI Act require starting August 2026?

Article 50's transparency obligations took effect August 2, 2026. Providers of AI systems interacting directly with people must disclose the interaction is with AI unless obvious, and providers generating synthetic content must mark it in a machine-readable, detectable format. Deployers must disclose emotion-recognition or biometric-categorization use and label deepfakes clearly at first exposure. A grace period runs to December 2, 2026 for the content-marking obligation on systems already on the market. Penalties reach 15 million euros or 3 percent of global turnover. Article 14 separately requires human oversight on high-risk systems, with a specific two-person verification rule for remote biometric identification in Annex III point 1(a), subject to Article 14(5) exceptions. Applicability and deadlines depend on the system.

Who should own AI governance, the CMO, the CTO, or security?

Commercial, technology, security, and legal owners need agreed responsibilities and an escalation path. IBM reports only 19 percent of breached organizations coordinated AI governance and security. That finding does not choose a title. Name the owner for access, review, logging, and audit scope and require them to resolve shared decisions.

Sources

Breach-study sample and dates: IBM's July 29, 2026 study announcement (602 breached organizations globally; March 2025 to February 2026). AI-incident costs, unapproved AI, access controls, and governance/security coordination: IBM, "Every AI agent followed the rules, and the data still leaked" (checked October 6, 2026). EU AI Act Article 50 transparency obligations and penalties: European Commission, Digital Strategy FAQ (accessed August 2026). EU AI Act Article 14 human oversight requirements: Article 14, EU AI Act. ISO 42001 versus NIST AI RMF certification and adoption comparison: ISO/IEC 42001 and NIST AI RMF. CMO AI-ownership split: Spencer Stuart, "The AI Reckoning" (December 2025; no published sample size or fieldwork dates, treat as directional).

About the author

Jeff Brokaw is a sitting CMO and Certified Chief AI Officer who ships AI in production, not slideware. He has been building AI systems commercially since 2016. He built the commercial engine behind $185M in new-business revenue for a defense manufacturer, and authored the go-to-market behind a $114M institutional raise that came together in under 30 days.