Event digest: How AI is changing the risk function

Event digest: How AI is changing the risk function

Whether you attended the IRM Midlands Regional Group event or couldn’t join us, this digest brings together the discussion’s main arguments, answers to the audience’s key questions, and additional practical guidance. Risk can shape AI adoption and strengthen its own work AI is changing more than the tools organisations use. It is changing how work…

Whether you attended the IRM Midlands Regional Group event or couldn’t join us, this digest brings together the discussion’s main arguments, answers to the audience’s key questions, and additional practical guidance.

Risk can shape AI adoption and strengthen its own work

AI is changing more than the tools organisations use. It is changing how work gets done, how decisions are supported, and how expertise is applied. Risk management must evolve alongside it. The webinar explored two connected opportunities: Risk helping organisations turn useful AI experiments into dependable operational capability, and AI enabling risk teams to work more effectively themselves.

The central message was that risk professionals can become architects of safe adoption. That means helping shape valuable uses of AI, their ownership, controls and evidence from the outset, rather than only assessing a solution after its important design decisions have been made.

1. Start with valuable work, not a general ambition to “do AI”

An impressive demonstration is not the same as a worthwhile business improvement.

Greg emphasised the importance of starting with an operational problem: which workflow needs to improve, what happens today, and what outcome would make a change valuable? AI may reduce administrative effort, support complex analysis, or help specialists work through more information. The starting point should be the work and its intended impact, not simply the availability of a new tool. This also changes how success is measured. Producing a draft in two minutes says little if it then requires several rounds of correction. The comparison needs to include the whole workflow: preparation, review, rework, and the effort required to produce an outcome people can actually use.

Risk can contribute by challenging the proposed benefit as well as the potential harm. Is the problem meaningful? Is the baseline credible? What evidence would show that the change has improved the work?

2. Useful experiments need an operating model to become a dependable capability

Both speakers distinguished something that helps an individual from something an organisation can rely on. Tarquin described the gap between local experimentation and wider operational use. A proof of concept may work well, while questions about ownership, data, architecture, training, controls and measurement remain unresolved. These foundations determine whether the organisation can sustain and expand the value.

Greg illustrated the distinction through a transport analogy: building a useful agent is like building a car. Running a dependable transport system also requires roads, signals, permissions and visibility of what is happening. For AI, the equivalent is a clear role, approved information, permitted actions, human responsibilities and arrangements for review and intervention. Shared foundations help organisations manage capability across workflows rather than recreate these arrangements around every new application. Experimentation remains valuable, and the challenge is connecting it to a credible route into everyday work.

3. Risk’s opportunity is earlier design influence alongside independent challenge

Becoming an architect of safe adoption does not mean Risk takes over every AI initiative. It means contributing while the design can still be influenced. Tarquin explained how existing risk expertise provides a useful starting point. Understanding data quality, representativeness, controls, and acceptable outcomes can help shape AI-supported processes. However, familiar techniques cannot simply be transferred unchanged, particularly when systems use unstructured information or respond dynamically to different contexts.

Greg connected this to Risk’s relationship with the business. Early involvement allows Risk to help establish the conditions for commercial value, rather than inherit unresolved questions at the point of launch. Leadership and culture matter here. A reporting line or approval right alone does not ensure that Risk is included early enough or can influence a decision. The organisation must make that involvement practical.

Risk can therefore collaborate in design while retaining its ability to challenge whether the proposed use is appropriate, sufficiently evidenced and ready to progress.

4. Accountability requires purpose, visibility and authority to intervene

A named owner is necessary, but accountability must extend beyond a name in a register. Tarquin placed purpose and ownership at the centre of assessing a use case. The system needs a clear job and an outcome against which its performance can be judged. The owner needs to understand what it does and the evidence supporting its use.

Authority is equally important. Drafting information for review, recommending a decision, and acting on someone’s behalf are different activities. Controls should reflect those differences alongside potential impact and reversibility. The discussion also highlighted the ability to reconstruct what happened: which information was used, what the AI produced or did, and where a person reviewed or intervened.

Alongside “Can we?”, organisations need to ask “Should we?”. Technical capability does not settle whether a use fits the organisation’s purpose or preserves an important human relationship. These questions need to reach people who can authorise, restrict or stop use, not merely record a concern.

5. Oversight must continue as data, systems and use change

Tarquin explained that changes in incoming information, document formats or content can affect outcomes. As AI becomes more significant to a workflow, waiting for the next periodic review may leave the organisation unaware of an emerging problem. This makes monitoring and reassessment part of the operating design. Organisations need to know what acceptable performance looks like, recognise departures from it and have a route to intervene.

Existing risk registers, controls and reporting processes remain useful foundations. The task is to establish what can be extended and where different controls or specialist expertise are needed, rather than assume either that everything must be rebuilt or that existing arrangements are sufficient. Human review also needs to remain substantive. Convincing, comprehensive output can encourage over-reliance. Reviewers must retain the knowledge and critical judgement to recognise errors, challenge conclusions and escalate concerns.

6. AI can change Risk’s work, not merely make its reports faster

Tarquin described three potential waves of change: releasing capacity, improving how processes achieve their objectives, and eventually redesigning how the function operates. The first opportunity is to reduce effort spent gathering information and performing repeatable analysis. The second is to improve the work through broader information, better-supported analysis and more timely evidence. The larger opportunity is to reconsider the process itself, rather than only accelerate existing steps.

Greg connected this to earlier intervention. As organisations change more quickly, Risk needs the capacity to understand developments, investigate emerging issues and influence decisions. AI can help connect information and direct attention towards matters requiring human judgement. But faster information is not the same as better risk management. Benefits need to be demonstrated through usable work, quality, rework, response to issues and total human effort. Released capacity also needs a deliberate purpose.

The future described was not a risk function without people. It was one in which people spend more of their expertise on investigation, judgement under ambiguity and independent challenge. AI can strengthen Risk’s role as a critical friend to the organisation: helping it move forward while understanding and managing the consequences.

Questions raised by the audience

The audience’s questions centred on practical confidence: how to contribute to AI adoption, challenge it effectively and preserve the expertise that makes Risk valuable. These answers bring together the main themes, with additional written guidance beyond the live discussion.

1. Who should supply and maintain the knowledge an AI system uses, and how can risk professionals contribute to its ongoing improvement?

The operational owner should create the knowledge base a specific agent uses, drawing on approved policies, procedures, and operational expertise. This is where the complexity comes in. Risk and Compliance must ensure relevant policies are accurate and that regulatory or internal changes to policy and process are accurately reflected in agent behaviour, working with the operational owner and Technology where needed. Updating a document alone does not establish that this has happened. Affected behaviour needs testing, with material changes approved before release.

This is a practical opportunity for Risk to partner with Operations: keeping the agent useful, current and aligned with how the organisation should work.

2. What skills do risk professionals need to assess and challenge AI confidently, and how can they develop them while keeping pace with change?

Risk professionals do not all need to become AI engineers. However, they do need enough practical understanding to question the information a system uses, its permitted actions, its limitations, and the evidence supporting its performance. Data literacy, systems thinking and the ability to recognise convincing but unsupported conclusions are important foundations. Existing risk, modelling and validation expertise remains valuable. Working through a defined use case with Technology and Operations can build understanding more effectively than trying to follow every new tool or announcement.

3. As AI takes on more risk-management tasks, what should remain human, and how can risk teams demonstrate their continuing value to the board?

AI can take on substantial analysis, preparation and monitoring work. That does not make generated output equivalent to an accountable risk function. Judgement under ambiguity, independent challenge, risk acceptance and organisational accountability still require people with the expertise and authority to act. Risk’s true value is not simply in the black and white. It is in the grey: applying human judgement, helping the business say yes where appropriate, and finding workable mitigations. That is central to being a commercial enabler. Demonstrate that value through issues resolved, decisions improved and earlier intervention. Jobs will change, but effective human judgement becomes more important.

4. How should education and early-career development change so risk professionals build practical judgement alongside AI, and what role should professional bodies such as IRM play?

If AI performs tasks through which junior professionals previously gained experience, development must become more deliberate. Reviewing AI output alone is not a substitute for learning how to investigate, weigh conflicting evidence and explain a decision. Supervised casework, independent analysis before comparing with AI, and discussion of mistakes and exceptions can preserve that learning. Development should build both AI fluency and the judgement to know when to challenge it. IRM needs to recognise this shift and support members through practical learning, highlighting the key concepts and emerging trends they should consider as their work changes.

5. What practical controls should a smaller institution put in place first to protect personal data, verify vendor security assurances and maintain data integrity when using AI?

Start with three controls: approved tools and clear data permissions; evidence of the supplier’s storage, encryption, retention, model-training and subprocessor arrangements; and checks linking AI outputs back to source records. At Vigilant AI.ai, we’ve seen this as a real challenge for smaller firms, which often have to trade off operational value for safety. We have purposefully designed our platform to combine defined AI Teammate roles and approved knowledge with access controls, configured policy checks on incoming requests and files, and records supporting auditability and controlled improvement. This gives Operations, Risk and Compliance a shared foundation for governing and maintaining useful AI work whilst keeping the sought-after operational impact central to the design.

Privacy and security review, and appropriate output checking, remain essential.


Continue your learning and the conversation

Institute of Risk Management and the Midlands Regional Group

The Institute of Risk Management brings together professionals who want to strengthen their risk-management knowledge, judgement and practice. Membership offers a way to continue learning and connect with others navigating the changes discussed in this webinar.

The Midlands Regional Group creates opportunities to exchange experience, explore emerging issues and learn from practitioners through events such as this one. Whether you attended live or are reading this digest afterwards, we encourage you to stay involved.

Explore the Midlands Regional Group and its previous events

Join the Midlands Regional Group on LinkedIn

Vigilant AI.ai

Greg Coleshill is Vigilant AI.ai’s Chief Commercial and Operating Officer. His contribution to the webinar focused on turning useful AI into dependable operational capacity and the opportunity for Risk to help shape that transition. Vigilant AI.ai provides the Agentic Workforce Operating System for regulated businesses.

Connect with Greg on LinkedIn

Further reading: AI Teammates – the future of work: from individual AI agents to a dependable transport system. Greg’s article develops the transport analogy discussed in the webinar: why useful individual agents need a wider operating system before organisations can rely on them.

4most

Tarquin Clarke, a Director at 4most, brought a risk, data and modelling perspective to the discussion. He explored how Risk’s mandate and expertise need to evolve, why dependable AI requires more than technical capability, and how oversight can be judged by its effect rather than its volume of activity. 4most is a specialist analytics consultancy supporting financial institutions with data-driven decisions. Its work spans areas including credit risk, modelling, stress testing and regulatory change.

Connect with Tarquin on LinkedIn

Further reading: The Risk Function of Tomorrow by Tarquin Clarke.





← Previous

Underestimated business risks: insights from Risk Expo 2026
Underestimated business risks were the focus of IRM’s first roundtable at Risk Expo 2026, where…


24 NOVEMBER 2026

Risk management software: trends, insights and best practice

This webinar explores how risk management software enables organisations to identify, assess, monitor, and mitigate risk more effectively.

Find out more

26 NOVEMBER 2026

Managing the risk of modern slavery: Guidance for the prevention, identification and response to human trafficking and forced labour

The IRM Environmental & Social Governance Special Interest Group invites you to a knowledge sharing event ahead of the launch of the new ISO Standard, ISO 37200, 'Managing the risk of modern slavery – Guidance for the prevention, identification and response to human trafficking and forced labour'.

Find out more

2 DECEMBER 2026

Building a risk-aware organisation: Lessons from the GFSC

Drawing on the GFSC’s own experience, the session will examine how risk management is embedded across the organisation, the role of risk culture and ownership, and the practical challenges of ensuring that risk management is more than a periodic reporting exercise.

Find out more

Advertisement