IBM EMEA AI governance survey: eight in 10 executives say AI incidents reached the board or C-suite

Eight in 10 senior executives at large organizations in Europe and the Middle East say an AI-related incident reached their board or C-suite in the past year, according to IBM research released Oct. 14.

How those incidents were handled once they reached the top is the detail worth slowing down for. Of the executives whose incident reached the board or C-suite, IBM’s report says 43 percent said it was formally reviewed there, while 37 percent said it was communicated informally. Neither the EMEA report nor its accompanying release includes the tabulations behind those shares or accounts for the remaining responses. Still, on the report’s own framing, over a third of the executives reporting an escalation said the incident was communicated informally.

The report does not say what record a formal review or an informal communication produced. For records managers, privacy officers and litigation teams, that missing variable is the one that matters. What senior leaders knew about an AI incident, and when they knew it, is a fact a regulator or opposing counsel may later ask a company to document.

IBM’s examples of the incidents in question include an agent failure, data exposure, a flawed AI-driven decision and a security breach. Its report describes agentic systems, the subject of much of the study, as ones that “can plan, decide and act on behalf of organisations.” Censuswide conducted the survey for IBM among 3,221 executive management respondents at organizations with 1,000 or more employees in France, Germany, Italy, the Netherlands, Saudi Arabia, Spain, the United Arab Emirates and the United Kingdom. IBM labels the study EMEA, though no African market was included, and responses were collected between Sept. 14 and 24, 2026.

What changed after incidents reached the top

The survey also asked what happened after an incident reached senior leaders, and the answers sharpen that documentation question. Among those organizations, 40 percent cited “a change to AI governance policy as a primary outcome,” the report said. Similar shares reported changes to investment or deployment plans (38 percent), engagement of external advisers (37 percent) and notification of regulators (35 percent).

A summary line atop IBM’s press release described the same 40 percent as organizations “changing AI governance policy or frameworks in response.” The report’s wording is narrower: respondents cited a policy change as a primary outcome. Neither document says whether respondents could choose several outcomes, so the remaining respondents simply did not report a policy change as a primary outcome, which is not the same as having changed nothing.

Board responsibility without agreement on agents

The policy figure sits against near-universal ownership on paper. Ninety-eight percent of organizations surveyed have formally assigned responsibility for AI risk at board level, and 26 percent have a dedicated board-level AI committee, according to the report.

Agreement on who answers for an individual agent is far thinner. Asked which role is accountable for an adverse outcome caused by an agent, respondents named no single role over 17 percent of the time. The chief information officer or chief technology officer led at 17 percent, followed by shared C-suite accountability at 14.5 percent, the chief risk officer or chief compliance officer at 14 percent and the chief AI officer at 13 percent. Smaller shares named the deploying team, the AI vendor, a business-line leader or the chief executive. IBM’s release calls that “no clear consensus,” and that is the right reading: the figures show answers spread across respondents, not that any one organization lacks an owner.

Liability drew a similar split. One in five respondents said legal liability should be shared between the deployer and the vendor. Nineteen percent called the issue too complex for a general answer, and another 19 percent said “the necessary regulatory framework does not yet adequately exist,” in the report’s words. Fifteen percent pointed to an individual accountable executive, and 14 percent assigned liability to the AI vendor.

IBM’s own executive framed ownership as the point. Ana Paula Assis, IBM senior vice president and chair for Europe, the Middle East and Africa and for Asia-Pacific, said in the release that the AI debate “has moved from adoption to governance and accountability.” Governance “must be built into how agents operate, with clear boundaries on what they can do, clear ownership when things go wrong, and controls that work in real-time,” Assis said.

For information governance and discovery teams, the split has a practical edge wherever an organization has not settled ownership of its agents internally. There, it is harder to know in advance whose files, messages and system records will hold the account of an incident. Identifying custodians and the systems that hold relevant data is a common early step in a legal hold.

Controls executives say they will need

The capabilities executives expect to need as agent autonomy grows point back to that ownership question. Formal vendor-liability frameworks led at 33 percent, followed by defined thresholds for human escalation at 32 percent, real-time monitoring and audit trails at 30 percent, and the ability to pause agent activity across the enterprise at 29 percent.

That list arrives as autonomy expectations climb. IBM said 35 percent of respondents run agentic AI in production today, and the report said 75 percent “expect agents to act without human approval of every action by 2027.” Only 37 percent said their governance is designed to scale with agentic AI, while another 45 percent said it could cope but needs strengthening. When deployment outpaces governance, 56 percent of organizations keep deploying within agreed risk limits as their frameworks adapt, and a third wait until governance is ready.

Gartner has flagged a related risk from outside IBM’s data. In June 2025, the research firm predicted that over 40 percent of agentic AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. “Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” said Anushree Verma, a senior director analyst at Gartner, in the firm’s announcement. That forecast was issued about 15 months before IBM’s fieldwork and is not a finding about IBM’s respondents.

Where the EU’s logging rules land

Where autonomy forecasts stay uncertain, regulation gives part of the record question a fixed date. In the European Union, 41 percent of executives surveyed said their organizations are accelerating governance investment ahead of the AI Act’s high-risk obligations, which IBM’s report said are expected in December 2027. That timing reflects the Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on July 24, 2026. For systems classified as high-risk under Annex III, it set Dec. 2, 2027, as the date the act’s high-risk requirements in Sections 1, 2 and 3 of Chapter III begin to apply. The one exception in that date provision, Article 6(5), does not touch the logging rules. The date for systems classified under Annex I is Aug. 2, 2028, though Article 2(2) keeps the logging and retention rules described below away from products listed in Section B of that annex.

Timing also depends on when a system reached the market or went into service. Under Article 111(2), high-risk systems placed on the market or put into service before the relevant application date are covered only if, from that date, they are subject to “significant changes in their designs,” in the provision’s words. Providers and deployers of high-risk systems intended for use by public authorities face an outer deadline of Aug. 2, 2030, and Article 111(1) sets a separate timetable for AI components of certain large-scale EU IT systems.

Chapter III’s Sections 2 and 3 carry the act’s record-keeping rules. Article 12 requires high-risk AI systems to technically allow for the automatic recording of events, or logs, over the system’s lifetime. Article 26(6) requires deployers to keep those automatically generated logs for a period appropriate to the system’s intended purpose, of at least six months, unless other Union or national law provides otherwise. That duty reaches the logs only to the extent they are under the deployer’s control.

Article 26(5) separately requires a deployer that identifies a serious incident to notify the provider immediately, followed by the importer or distributor and the market surveillance authorities concerned. The act defines a serious incident as one that directly or indirectly leads to death or serious harm to health, serious and irreversible disruption of critical infrastructure, an infringement of obligations under Union law intended to protect fundamental rights, or serious harm to property or the environment. IBM’s examples are broader, and the report does not say how many escalated incidents would meet that definition.

All of these duties attach only to systems that meet the act’s high-risk classification. IBM’s report does not say how many of its respondents’ agents fall into that category.

Article 26(6)’s control qualifier deserves attention when an agent runs on someone else’s platform. Article 19 places a parallel retention duty on providers for logs under their control, so which party holds a given agent’s logs is a factual question for each deployment. IBM’s third-party figures bear on that question without answering it. Fifty-four percent of executives acknowledged gaps or limitations in governing AI that vendors, partners and supply-chain participants use. Where controls exist, the report said, 43 percent rely on contractual provisions and 39 percent on audit rights. Separately, 21 percent reported continuous monitoring of third-party AI, and 13 percent require independent assurance or certification. The report does not say whether the contract and audit shares are of all respondents or only those with controls in place, so those two cannot be ranked against the other two.

Reading a vendor’s governance survey

Those third-party figures, like every number in the study, come from research commissioned by a company with a commercial stake in the answer. IBM sells AI governance software, and the report should be read with that in view. Its five case studies, featuring Banco Santander, Bertolotti, Etisalat, A1 Telekom Austria and Celerity, describe deployments built on IBM products, including watsonx.governance and watsonx Orchestrate. Neither the EMEA report nor the EMEA release publishes question wording, a margin of error or per-country sample sizes.

IBM’s country releases fill in part of that gap. Its U.K. release says the study included 504 U.K. respondents, and its Italian release puts the Italian sample at 510. The report calls the United Kingdom “the most advanced market for agentic AI in the study,” with 57 percent running it in production, and sector regulators top the reported drivers of U.K. governance investment, at 27 percent. In Saudi Arabia and the United Arab Emirates, national AI strategy requirements lead at 61 percent. The EMEA report and release publish no sample size for the Gulf group or the EU group, so those comparisons carry more uncertainty than the headline figures.

Settling the record before the next escalation

Whatever the regional picture, the practical work starts with incidents already on the books. If an AI incident reached the board or C-suite this year, the useful exercise is to find its record. That means the paper, minutes or message that carried it, who sent it, and whether any decision that followed is written down where a reviewer could find it. An informal communication is not necessarily an undocumented one, but only a search will tell.

IBM’s own recommendations point the same direction. The report says “every agent should have a defined identity, owner and scope,” and it calls for monitoring and audit trails that verify controls were applied. Those attributes are also what a discovery team needs to scope a collection: which agent, owned by whom, acting on which systems. ComplexDiscovery examined that problem in its August 2026 tutorial on AI agents as evidence.

Vendor contracts deserve a second look as well. Audit rights already feature in IBM’s data, but the narrower question is whether a contract also says who keeps an agent’s logs, for how long, and how quickly they can be preserved and produced once a dispute or inquiry begins. For organizations deploying Annex III high-risk systems in the EU, Dec. 2, 2027, is as much a records date as a compliance date. The logging and retention rules apply from then, subject to the transitional provisions of Article 111.

IBM’s survey has eight in 10 executives saying an AI incident reached their board or C-suite in the past year, and over a third of those saying it was communicated informally. If a regulator asked tomorrow for the record of the last AI incident your senior leaders heard about, would you hand over minutes, an email thread or a recollection?

News sources

Assisted by GAI and LLM Technologies

Source: HaystackID published with permission from ComplexDiscovery OÜ

Advisor’s Note: The IBM survey underscores a challenge many organizations face as AI adoption accelerates: governance must extend beyond policy into operational execution. As AI incidents increasingly reach executive leadership, organizations need repeatable oversight, validation, and accountability processes that produce defensible evidence of responsible AI management.

HaystackID® AI Governance Services help organizations operationalize AI governance by embedding controls, validation workflows, decision rights, and monitoring into day-to-day delivery and risk management. By translating governance intent into measurable operational practices, organizations can scale AI adoption while strengthening compliance readiness, supporting stakeholder trust, and maintaining defensible documentation as regulatory and business requirements evolve. Learn more at AI Governance Services | HaystackID.