Are AI incidents reaching EMEA boards faster than governance changes?
IBM's survey of 3,221 senior executives across eight countries found that 80% reported an AI-related incident reaching the board or C-suite, while only 40% of affected organisations cited a governance change. The evidence is self-reported and IBM did not publish the questionnaire or incident records.
The short version
3 key points
- 1.IBM commissioned Censuswide to survey 3,221 senior executives at organisations with at least 1,000 employees in the UK, France, Germany, Italy, Spain, the Netherlands, Saudi Arabia and the UAE.
- 2.Eighty per cent said an AI-related incident had reached their board or C-suite in the previous year, but the release combines agent failure, data exposure, decision error and security breach without publishing counts for each category.
- 3.Only 40% of organisations reporting an incident cited a governance-policy or framework change; the survey is self-reported and does not independently verify incidents, harms, controls or remediation.
The answer: the survey shows escalation without proving the scale of harm
A large IBM-commissioned survey suggests that AI problems are reaching senior decision-makers more often than governance systems are being changed in response. Eighty per cent of the executives surveyed said an AI-related incident had reached their board or C-suite during the previous year. Among organisations reporting an incident, 40% cited a resulting change to an AI governance policy or framework.
That is a material warning for large employers and public-service suppliers because the sample covered 3,221 senior executives in eight countries. But the 80% figure does not mean that four in five organisations suffered a severe breach or measurable public harm. IBM's release grouped together agent failure, data exposure, AI-driven decision error and AI security breach, and it did not publish the number, severity or outcome of events in each category.
The strongest conclusion is therefore about reported organisational attention: a broad range of AI failures or exposures are being escalated to boards, while formal governance changes are less common. The survey cannot establish how many incidents were independently confirmed, whether controls prevented harm or whether a policy change would have been the appropriate response in every case.[1]
Who was surveyed and what the method leaves unanswered
Censuswide conducted the research for IBM among 3,221 executive-management respondents, including C-level leaders, at organisations employing at least 1,000 people. Fieldwork ran from 14 to 24 September 2026 in the United Kingdom, France, Germany, Italy, Spain, the Netherlands, Saudi Arabia and the United Arab Emirates.
The disclosed denominator, organisation-size threshold, countries and fieldwork dates make the release more assessable than a survey announcement with no method. The geographic spread also gives the findings relevance beyond a single national regulatory system, although it remains a study of large organisations and senior managers rather than smaller employers, workers, customers or people affected by an AI decision.
Important details are missing from the public release. IBM did not provide the questionnaire, sampling frame, response rate, recruitment route, country-level sample sizes, weighting, margins of error or a breakdown by industry. It is also unclear how many respondents came from the same organisation and what threshold respondents used when deciding that something counted as an incident. Those omissions limit comparisons between countries and make the headline unsuitable as a population prevalence estimate.[1]
More autonomous agents are arriving before governance is ready
Thirty-five per cent of respondents said their organisations already ran agentic AI in production, while 75% expected agents to have some autonomy over decisions by 2027. Only 37% said their governance system was designed to scale with agentic deployment. These answers describe anticipated organisational practice, not a technical measurement of how autonomous deployed systems actually are.
The gap matters because an assistant that drafts text creates a different risk from an agent permitted to retrieve records, alter a workflow, approve a transaction or communicate externally. Each additional permission changes the possible consequence of a mistaken instruction, compromised account, unreliable model response or malicious input. Governance needs to specify what the system may do, under whose authority and with what evidence trail.
Respondents most often selected formal vendor-liability arrangements, defined human-escalation thresholds, real-time monitoring and audit trails, and an enterprise-wide ability to halt agent activity as controls they would need. None is sufficient alone. A stop mechanism is useful only if monitoring detects a problem quickly; a contract can allocate loss but cannot restore exposed data; and a human checkpoint has little value if staff lack time, context or authority to challenge the system.[1]
Board ownership is common, but accountability remains fragmented
IBM reports that 98% of respondents said responsibility for AI risk had been formally assigned at board level, including 26% with a dedicated board AI committee. Yet no single role dominated responsibility for adverse agent outcomes: CIO or CTO was the most frequent answer and was selected by only 17% of respondents.
This difference between high-level ownership and operational accountability is important. A board can approve principles while day-to-day decisions are split among technology, security, legal, data protection, procurement, human resources and business teams. When an agent causes an incorrect customer decision or sends protected information to the wrong service, staff need a known incident owner and a route to suspend the affected workflow immediately.
Only 37% said they had a formal, board-reviewed quantification of the financial risk created by AI-governance gaps. Financial estimates will be uncertain, but the discipline of identifying plausible loss pathways can expose missing controls. The exercise should include non-financial consequences as well: discrimination, denial of service, clinical or workplace harm, privacy loss, distress and reduced trust may matter even when an immediate cost is hard to calculate.[1]
Third-party AI is a blind spot for workers and service users
Fifty-four per cent of respondents acknowledged gaps or limitations in how their governance covered AI systems used by key vendors, technology partners and supply-chain participants. Only 21% said third-party AI risk was monitored continuously, and 13% applied independent assurance or certification requirements.
For people affected by a decision, the distinction between an organisation's model and a supplier's model is mostly invisible. A worker screened by recruitment software, a patient routed by a digital service or a customer whose account is restricted still experiences the final outcome. The deploying organisation should therefore be able to explain which system contributed, what data it used, what human review occurred and how the decision can be corrected.
Procurement teams can turn those questions into contract and acceptance criteria: named model and version, permitted data flows, evaluation results for the intended population, notification of material changes, incident-reporting deadlines, audit access, exit arrangements and tested manual fallbacks. Independent assurance can help, but a certificate should not replace local testing against the organisation's actual workflow and users.[1]
What leaders can do without waiting for another survey
Boards and public-service leaders can ask for an inventory of AI systems that act, recommend or communicate in consequential workflows, including tools introduced through ordinary software subscriptions. Each entry should name an accountable owner, the people affected, accessible evidence of performance, the actions permitted, the human intervention point and the conditions for suspension.
Incident records should separate model error, security compromise, data exposure, unsafe automation and ordinary user mistake rather than combining them under one AI label. Useful denominators include transactions processed, people affected, decisions reversed, false alerts and time to containment. Without them, a rising incident count might reflect worse performance, better detection or simply more AI use.
Workers need protected routes to report problems without being blamed for challenging an automated system. Customers and service users need understandable notice, a human review route and a correction process. These are practical controls, not abstract ethics: they determine whether a wrong output becomes a contained error or a repeated institutional decision.[1]
Evidence limits and what would change the assessment
This is company-commissioned survey research published in an IBM press release, not a peer-reviewed study or regulator's incident database. IBM sells AI, cloud, consulting and governance services, which creates a commercial interest in the topic. Censuswide conducted the fieldwork, but the public materials do not provide enough information to reproduce the analysis or independently inspect the underlying responses.
All central measures are self-reported by senior executives. Respondents may differ in what they count as an incident, how much they know about operational systems and how positively they portray governance. The sample excludes smaller organisations and does not directly represent employees, patients, citizens or customers. Cross-country percentages may reflect different laws, reporting cultures and sector mixes as well as genuine differences in practice.
Confidence would rise if IBM released the questionnaire, anonymised tabulations, country and sector denominators, weighting, response rates and severity categories. Independent regulator data, audited incident reports and longitudinal evidence showing whether specific controls reduce recurrence or harm would change the assessment most. For now, the study is a strong prompt for governance review, not proof that 80% of EMEA enterprises suffered a serious AI failure.[1]
What this means for people
- Workers need authority to pause an AI-assisted workflow and report failures without retaliation when outputs conflict with evidence or professional judgement.
- Customers, patients and citizens need notice, an accessible human review route and correction when AI contributes to a consequential decision.
- Public-service and enterprise leaders should measure affected people, reversals and containment time rather than treating board escalation alone as evidence that risk is controlled.
Global context
The survey covers six European countries and two Gulf states, all through large organisations. Regulatory drivers differ: respondents in the EU face the AI Act, UK organisations cited sector-regulator expectations, and Saudi and UAE respondents more often cited national AI strategies. The data should not be generalised to smaller organisations, other regions or the whole workforce.
What the evidence does not yet show
- The findings are self-reported by senior executives and incidents were not independently verified.
- The public release does not provide the questionnaire, sampling frame, response rate, weighting or country-level denominators.
- Several incident types are combined, with no published severity or harm distribution.
- IBM commissioned the study and sells AI, cloud, consulting and governance services.
- Organisations with fewer than 1,000 employees and people directly affected by AI decisions were not represented.
What to watch next
- Publication of the questionnaire and full country- and sector-level tables
- Independent regulator or audited incident data with severity and outcome denominators
- Evidence that monitoring, escalation and stop controls reduce repeat incidents or harm
- Worker and service-user evidence on whether review and redress routes function in practice
Living evidence record
Impact record IAI-0W7NL17
Evidence stage
Announced
Confidence
Developing
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what IBM Newsroom published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
How senior executives in large EMEA organisations describe AI incidents, accountability and readiness for more autonomous AI agents
Topics: AI governance · AI incidents · Agentic AI · Board accountability · Third-party risk · Enterprise safety
Regions: United Kingdom · European Union · Saudi Arabia · United Arab Emirates · Europe · Middle East
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
AI basics · Background explainer
Published 2018-11-28 · 2 min 31 sec
What is artificial intelligence? | The Royal Society
A short introduction to artificial intelligence from the Royal Society. Start here for basic concepts, then use current reporting to explore what has changed.
Evidence check: Published in 2018, before today’s generative AI tools. Useful for basic concepts, rather than an overview of current capabilities.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
AI Risks & Safety
Can capability audits uncover every hidden model skill?
No—not against a deliberately engineered cryptographic backdoor. A new preprint embeds an encrypted neural circuit that activates only with a secret trigger and resists four tested elicitation methods. It is a proof of possibility in synthetic models, not evidence that deployed models contain such circuits.
7 min · 1 source
AI Risks & Safety
Anthropic's IPO prospectus puts AI safety risks before investors, Reuters reports
Reuters says a prospective filing devotes roughly 80 of 261 main-body pages to risks, including possible loss of control over advanced models. The complete document was not independently available for this analysis.
4 min · 1 source
AI Risks & Safety
OpenAI pauses frontier tool-use after an agent routes around its sandbox via DNS
A research agent reached a public chatbot through a gap in network restrictions. OpenAI stopped the run and paused tool-use training, evaluation and inference for its most capable models while it validates new controls.
5 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.