- Modelling shows that AI failures are commonplace and could put up to 3.6% of annual market revenue at risk in the future
- Research finds many organizations lack the transparency and governance needed to learn from failures
20 Oct 2026: Three in four businesses (75%) have experienced an AI failure in the past year, according to new research from BSI, creating additional risks for organizations chasing efficiencies and growth. 92% of respondents in a poll of 1,287 business leaders across 9 countries reported multiple incidents, with 9% reporting up to 100, while 30% of organizations affected said the problem had led to a security breach.
The global study, The cost of AI failure, explored the outcomes of situations where businesses have found that the AI they are using or developing has deviated from its intended goal or produced harmful results. It found that failures are already widespread, with 88% of those who were developing their own AI tools or products reporting a failure. For most, the consequences were operational, 44% said it had cost their company time, 29% said it had led to regulatory or compliance issues, 28% said it had had safety implications, and a fifth said it had led to reputational damage (21%). Despite this, under half of all businesses surveyed had a process in place for managing AI incidents and ensuring a timely response (47%).
While more than a quarter (27%) of business have already reported financial losses from AI failures, economic modelling by Alma Economics2, commissioned by BSI, finds that failures could put as much as 3.6% of market revenue at risk per year. Market revenue refers to the total annual value of economic activity in a market, measured by how much money is spent or earned each year. Alma researchers used this figure to estimate the potential financial impact of AI failure. This does not account for potential fines levied by regulators3. With 78% of businesses expecting their AI investment to increase, and deployment moving steadily from pilots to scale, exposure across every failure category is likely only to grow.
Today, the direct financial consequences of failures remain relatively contained. Most incidents centre on operational disruption rather than financial loss, with only 13% of organizations currently deploying agentic AI at scale. Yet the risks are growing.
The modelling shows that financial risk scales as the complexity of AI systems grows. Basic, but commonplace, chatbot failures pose a comparatively modest risk to market revenue of 0.1%. In fields like legal services and management consulting, hallucinated AI outputs were modelled as putting 2.5% of sector revenue on the line. The costliest risks were attributed to deployments of agentic tools which actively engage with enterprise systems, putting 3.6% of market revenue at risk for certain sectors.
How organizations are deploying AI
Organizations were using AI in different ways, but some agentic AI use was already widespread, with nine in ten exploring it in some form. Alma’s analysis identified agentic AI as the highest risk category, followed by hallucinations and fabricated outputs. As AI systems gain greater autonomy and use increases, failures could become more difficult to detect, more common and potentially more costly, especially as fines from regulatory authorities come into play.
Of those surveyed, 27% were piloting several use cases for agentic AI. 26% were deploying for specific use cases, 24% were exploring potential use cases, and 13% were deploying at scale. Asked about their concerns around agentic AI, 53% cited security fears, while 37% raised the lack of AI governance and a third (33%) worried about the absence of national or global regulation.
How corporate AI governance is evolving
Research conducted by BSI a year ago found businesses were at high risk of an AI governance gap, where AI tools are deployed without appropriate governance programmes in place. This year’s findings suggest progress is being made. 38% of business leaders said their organization has established an AI governance programme, up from 24%. Half had clear processes governing the use of confidential data for training AI, up from 40%, and 49% now have a risk assessment approach, up from 30%.
Yet questions remain as to whether organizations are fostering a culture where AI failures can be openly identified, reported and learned from. The findings point to a persistent confidence-action gap. While AI incidents are widespread, many organizations still lack the governance, reporting processes and transparency needed to turn failures into opportunities for improvement.
For example, two thirds (65%) express confidence that their organization is implementing safe AI policies aligned with standards, yet only 55% say these are in place. 68% say their organization is transparent about AI data and algorithms but only 39% understand what data is used to train or deploy AI systems. And only 27% have restricted employees from using unauthorised AI tools. 69% provide critical-use training, but only 48% have structured AI learning programmes.
Susan Taylor Martin, Chief Executive, BSI said: "As AI reshapes our businesses, economies and society, business leaders need to consider what happens when AI doesn’t work as intended. What’s clear is that AI failure is becoming a regular occurrence, bringing with it significant costs. We need to encourage leaders to do everything they can to prevent failures while being comfortable sharing what they have learnt when an issue arises.
Realising the benefits of AI requires taking responsibility for how it is used. Business has been here before, responding to growing cybersecurity risks over the past two decades by embedding safeguards proactively and rigorously investigating and implementing learnings from AI incidents. Now, we must prioritize putting in place robust governance and guardrails to enable safe innovation that protects us all. Acting early is likely to help reduce exposure to financial and legal liabilities, while also helping to build trust.”
With ongoing debate about the need for international oversight of frontier AI, business leaders were split over who should take responsibility. While 61% of business leaders expect tech companies to manage their infrastructure, more than a quarter explicitly call for an international non-governmental body to provide independent governance, confirming that enterprise decision-makers see international standards as an anchor for market trust.
How AI affects an organization’s risk profile
With the modelling suggesting that unmanaged AI failures can create material financial exposure, particularly as agentic AI and generative AI become embedded in core business processes, BSI is advising organizations to view AI governance as a strategic capability, much as they currently treat cybersecurity, privacy and compliance.
Despite the likelihood of failures, enthusiasm for AI remains strong, with 78% expecting their company’s AI investment to increase and 81% saying AI has delivered tangible benefits. Three quarters say AI projects are delivering ROI.
However, 40% also said that AI had been a source of risk or area of weakness for the business, while more than half (51%) felt AI investment had taken resources that could have been used on other projects. 68% said the cost of AI had increased for their organization year-on-year, and 71% expect AI-related costs to rise further. With AI adoption gathering pace, 54% said their organization has a standardized way to assess if an AI tool delivers as intended (up from 35%).
Of organizations that had experienced failure, a quarter said this was during experimentation or pilots (26%), while 32% had had problems with an internal tool. 44% of respondents said they had experienced issues with both internal and external tools. Failure was more prevalent amongst larger businesses, (83% versus 57% of SMEs), and skewed towards the technology (82%), financial services (82%) and medical device (94%) sectors.
Read the full report here.
About the research
BSI commissioned Alma Economics to assess the potential economic cost of AI failure and provide evidence of the scale of financial risk failures could create for businesses and markets.
The study explored instances whereby AI systems make an error, malfunction, degrade in performance, or produce a biased output, where this behaviour is generally unexpected. We also focused on failures occurring in real-world use, rather than during development or testing. This is distinct from failures of frontier models, which were not in scope.
The intention was not to calculate a single global cost of AI failure, but to estimate the order of magnitude of financial exposure associated with different types of failures.
The economic study used known data from real-world AI failures to estimate wider market-level financial exposure. Each estimate was built using four factors:
- Severity of the failure
- Size of the relevant market
- Extent of AI adoption/exposure
- Likelihood of the failure occurring (where applicable)
Alongside this, BSI commissioned FocalData to survey business leaders, replaying questions asked in summer 2025 to build a year-on-year picture. The survey covered the views of 1,287 business leaders (manager to C-suite) across nine countries (Netherlands, Japan, India, Australia, US, UK, France, China, Germany) and seven sectors (Healthcare and Life Sciences, Consumer Goods & Retail, Technology, Financial Services, Transportation & Logistics, Built Environment / property, Professional Services). Research was conducted in August 2026.
