AI and Risk Assessment: Partner or Problem?

AI can analyse inspection images, monitor equipment and identify patterns across large volumes of information. Used proportionately, it could make risk management more accessible. Used without sufficient scrutiny, it may introduce false confidence and new areas of exposure.

Across the industries regulated by the Health and Safety Executive (HSE), businesses have reported using computer vision to inspect difficult-to-reach locations, sensor data to inform maintenance and large language models to analyse accident reports or help produce risk assessments.

In 2025, HSE documented approximately 250 self-reported AI use cases being developed or deployed across the sectors it regulates. It did not evaluate whether those applications or their implications were appropriate or effective.

For commercial insurance, this raises a more useful question than whether AI is simply good or bad. Can it help make meaningful risk management available to more businesses without reducing the quality of assessment or weakening accountability?

In the Bank of England and Financial Conduct Authority’s 2024 survey, 95% of insurance-sector respondents reported using AI. The category included several insurance markets and service providers, so the finding does not show adoption among commercial brokers or risk-engineering teams specifically.

From seeing risk to influencing decisions

AI can enter the risk process at several points. It may help people see hazards through images and monitoring, anticipate problems by finding patterns in operational information, or influence decisions by recommending priorities and possible controls.

Remote equipment may reduce the need for people to enter hazardous locations. Continuous monitoring may highlight changes between scheduled inspections, while analysis of maintenance data may help businesses investigate warning signs before equipment fails.

The consequences of an error differ at each stage. A system that organises photographs for a surveyor has a different influence from one that determines whether a control is adequate or generates a risk requirement. As technology moves from gathering evidence towards interpreting it, the need for proportionate oversight grows.

Scale can widen access

Structured risk-management support remains uneven. Detailed site surveys are resource-intensive and tend to be directed towards larger, more complex or higher-consequence risks. Other businesses may face familiar property, liability and operational hazards without receiving the same level of guidance.

AI could form part of a wider digital, remote and hybrid approach to closing that gap. It may help organise information, provide accessible initial guidance and identify cases requiring professional attention.

RiskSTOP’s 2026 Risk Management Infographic illustrates the scale and recurring nature of the issues involved. More than 27,000 risk-improvement requirements were raised across 16,800 commercial assessments, with electrical safety, combustible storage, hot-work controls and fire risk assessments among the most frequent findings.

Those figures do not demonstrate what AI can detect or predict. They show that effective risk management often depends on recognising, prioritising and addressing practical weaknesses across large numbers of commercial properties and operations.

For brokers, the opportunity is to extend appropriate support to more clients while preserving routes to greater expertise. Lower-complexity risks may benefit from accessible guidance, while unusual findings and higher-consequence exposures may need remote professional interpretation or a site-based assessment.

That is the practical meaning behind RiskSTOP’s view that risk management should be for everyone. Wider access should provide exposure-appropriate support, rather than a lower standard of protection for smaller businesses.

When confidence outruns evidence

An AI output is limited by the information available to the system and the circumstances it was designed to recognise. Missing, poor-quality or outdated data can produce an answer that does not reflect the complete exposure.

An image-based system can only assess what its inputs capture. It may identify visible concerns while missing how work is organised outside the image. A predictive-maintenance model may also perform less reliably when equipment or operating conditions fall outside the situations represented in its data.

Industry respondents to HSE’s research identified over-dependence, deskilling, warning fatigue, missed hazards and inaccurate safety assessments as potential risks. These concerns show that the problem is partly organisational. A technically capable system can still be used poorly or trusted beyond its intended limits.

Third-party dependency adds another layer. The Bank and FCA found that one-third of reported AI use cases were third-party implementations. Among firms using or planning to use AI, 46% reported only a partial understanding of the technologies involved. Limited access to the underlying model or training data can make effective oversight more difficult.

Systems that use images, monitoring information or worker data can also introduce privacy, data-security and cyber risks.

Accountability remains important. HSE says existing health and safety duties continue to apply when AI affects workplace safety. Those duties do not disappear when an organisation introduces AI or relies on a third-party system.

The International Association of Insurance Supervisors has identified continuing legal uncertainty around AI-related exposures. It has also warned that reliance on a small number of widely used models could create concentrations in which one weakness affects many organisations.

What credible partnership looks like

Human involvement only provides meaningful oversight when the reviewer has sufficient knowledge, relevant information and the authority to challenge or escalate the output. IAIS guidance also highlights appropriate training, independence and clear accountability.

For brokers considering AI-assisted risk information, useful lines of enquiry include the source of the evidence, what the system was designed to identify, what may fall outside its scope and who reviews material findings. There should also be a route for dealing with incomplete evidence, unusual risks or disagreement with an automated conclusion.

These questions become particularly relevant when AI-assisted information affects underwriting appetite, policy conditions or risk-improvement requirements. They do not require brokers to audit algorithms. They provide a practical basis for judging the evidence and deciding whether further professional input may be appropriate.

Clear explanation, challenge and escalation can reduce the risk of an automated judgement being accepted without sufficient scrutiny. They also keep the focus on helping businesses understand and improve their risks.

Wider access needs proportionate judgement

AI could help extend structured risk management to businesses that have historically received little support. Its value depends on reliable evidence, appropriate limits and accountable human judgement.

Automated guidance, remote assessment and professional site surveys serve different purposes. Higher-complexity or higher-consequence risks may still require professional, site-based assessment.

AI is a useful partner when it broadens access, directs attention and escalates complexity to people. It becomes a problem when automation is treated as assurance or used as a blanket replacement for professional judgement.

AI is only one route towards more accessible risk management. Although neither uses AI, RiskSTOP’s Instant Risk Guidance provides immediate, personalised guidance for SME clients, while Rapid Risk Management combines data-led remote assessment with human risk expertise. Alongside professional site surveys and practical risk-improvement support, they show how different delivery models can widen access.

Explore how RiskSTOP can support clearer risk assessment and practical risk improvement.