By some estimates, AI has yet to experience lift off in many UK organisations. Government figures from February claim that just one in six businesses use it, with most having no plans for adoption. Yet many larger companies are embracing it with gusto. The challenge for them is to harness the technology for competitive differentiation while managing the inherent risks that accompany it.

A new BSI report lays bare the impact of those risks, claiming AI failure could cost nearly 4% of annual revenue. Executives are confident they have a handle on things. But as always, there’s a gulf between what they think and the reality on the ground. AI governance programmes are still vanishingly rare. But ISO 42001 can provide the structure and discipline that many organisations are missing today.

The Scale of the Governance Challenge

AI failures are already making the news with growing frequency. Stories of rogue agents hacking companies, hijacking websites, and inserting malware into open source repositories are now almost too numerous to recall individually. But on a potentially much wider scale, AI systems are hallucinating, producing biased output, leaking corporate and customer information, and failing in other ways.

The size of the problem is still not fully understood given the proliferation of shadow AI in the enterprise. Over a third (35%) of UK and US infosec leaders say that using GenAI without permission is the most common employee mistake of the past 12 months, according to IO’s State of Information Security Report 2026.

That said, the BSI has had a pretty good stab at estimating the magnitude of the AI governance challenge. Its report is based on interviews with 1,287 business leaders in nine countries (Netherlands, Japan, India, Australia, US, UK, France, China, Germany) and seven sectors.

It claims that three-quarters (75%) of organisations have experienced an “AI failure” in the past year, with 92% of these respondents reporting multiple incidents and one in 10 (9%) citing up to 100. The impact is significant. Nearly a third (30%) say the failure led to a security breach. And a similar share (27%) cite financial losses. The BSI calculates that an AI failure incident could shave between 0.1% and 3.6% off global market revenue each year. Agentic incidents account for the larger figure, which makes sense considering the risks associated with AI working autonomously.

“AI failure is not a hypothetical, but a regular, occurrence, and one that is likely to bring with it significant costs, both to individual organisations and the wider markets in which they operate,” says BSI CEO Susan Taylor Martin. “Many, if not most, organisations will experience some form of AI failure, whether that is a hallucination, the wrong information being shared with a client or a chatbot responding in an inappropriate way.”

The governance gap highlighted by the report last year does appear to be closing. The share of respondents reporting an established AI governance programme has risen from a quarter (24%) to nearly two-fifths (38%) over that time. But it in no way matches the over-confidence of business executives, 65% of whom say they’re feel sure that their organisation is implementing safe AI policies aligned with standards.

Starting with ISO 42001

The gap between executive pronouncements and reality is often wide. But here it feels like the gulf is particularly vast. Only 55% of respondents say “safe AI policies” are even in place, for example. This is where a structured, formalised governance programme comes into its own.

Tim McGarr, global head of AI market development & partnerships at BSI, believes ISO 42001 can provide the foundations of that programme. He likens it to ISO 27001; a complementary standard which has become the foundation for best practice information security management.

“ISO/IEC 42001 gives organisations a structured framework for establishing, implementing, maintaining and continually improving AI governance. It helps them identify and manage AI risks, assign accountability, assess impacts, monitor performance, and demonstrate responsible AI management in a consistent and auditable way,” he tells IO (formerly ISMS.online).

“ISO/IEC 42001 helps organisations build trust, strengthen resilience, get ready for regulation and enable safe AI innovation at scale.”

Ronald Lewis, head of cybersecurity governance at Black Duck, adds that ISO 42001 helps to “institutionalise” AI governance.

“It gives organisations a measurable framework for accountability across the AI lifecycle and helps standardise how AI resources are selected, developed, deployed and monitored,” he tells IO. “Ultimately, good governance shouldn’t slow AI adoption. It should provide the structure and discipline organisations need to deploy AI at speed and scale while understanding and managing the risks they’re taking.”

Joining the Dots

However, AI governance can’t operate in isolation. An AI failure such as a data leak can be both a privacy and a security risk, for example. In fact, three-quarters (74%) of UK and US security professionals say they manage security, privacy and AI as one connected system, according to the IO report.

Lewis believes this is the right approach.

“You can’t effectively govern AI without governing the data it consumes, how that data is protected and the security risks the technology introduces. AI governance should therefore build on the policies, practices and risk-management disciplines already operating across the enterprise, not become another siloed compliance programme,” he explains.

“Done correctly, it fits seamlessly into the enterprise and enables responsible AI adoption. Any other approach risks becoming disruptive rather than enabling.”

The BSI’s McGarr agrees, arguing that a more holistic approach to risk management strengthens business resilience. “ISO 42001 shares a common structure with ISO 27001 (information security) and ISO 27701 (privacy), making it easier to incorporate good AI governance,” he adds. “Organisations can draw on proven governance disciplines from cybersecurity and privacy while applying them to AI, helping to reduce both the likelihood and impact of failures while building stakeholder trust.”

Maintaining Momentum

These frameworks also preach a message of ongoing monitoring of controls and continuous improvement. It’s a message that GRC teams must absorb if they are to effectively manage risk across dynamic AI environments, says McGarr.

“Models, data, user behaviour, regulatory requirements and operating environments can all change over time, creating new risks after an initial assessment has been completed,” he explains. “Organisations need ongoing monitoring and measurement to identify emerging issues early, respond quickly to incidents, and continuously improve controls.”

Black Duck’s Lewis takes a similar view, noting that “the world has moved on” from compliance based around periodic, point-in-time snapshots.

“The attack surface is evolving, threat actors are reprioritising targets, and adversaries are constantly looking for new weaknesses,” he concludes. “Periodic audits still provide valuable assurance, but they can’t be the primary mechanism for governing something that changes continuously.”

Expand Your Knowledge

Webinar: Lessons from One of the World’s First ISO 42001 Certifications

Blog: The Revolut Breach and the Need for Joined-Up Governance

Podcast: Phishing for Trouble S2 E11: Beyond the Headlines: AI, Regulation and the Reality of 2026