A risk manager’s guide to AI

Use AI across the risk management process, from reviewing policy and collecting evidence to estimating impact and writing the board brief.

An isometric grid of sage and ivory risk tiles with one amber exception raised above it.

The work portal is down. Requests are still coming in. Ask AI to find the clause that says who can switch to the backup. Check the clause in the approved plan before anyone switches.

I have spent nearly two decades as an AI engineer and entrepreneur, and almost as long in risk management. I have seen how we record incidents, measure risk and report to the board change in that time. What excites me about AI is how much a risk team can check before a decision is due.

I still think too few risk professionals make AI part of their day-to-day work. In the Bank of England and Financial Conduct Authority’s 2026 survey, 86% of 196 responding UK financial firms reported using AI. The survey tells us how many firms use AI. I want to get down to the work on a risk analyst’s desk. Bank of England and Financial Conduct Authority, 2026 AI survey

The worked figures in this guide are practice inputs. Use an approved tool and give it only the information it needs. Confirm which data your organisation allows in that environment before supplying customer records, incident material or internal policies. NIST, Generative Artificial Intelligence Profile

Files for this guide

Use AI when a shared work portal stops

A shared work portal stops for two hours. Staff use it to submit and track requests. Use the approved list of services to find which teams depend on the portal. Ask AI to draft a form for outage reports. Give each report an ID, a time, a service name and a source record. Ask the service desk to check each report against the incident log.

Assume the queue is empty when the portal stops. If 40 requests arrive each hour and staff complete none, the queue reaches 80 requests after two hours. Ask the operations lead to check the arrival rate and the number of requests that staff completed offline. Show which figures come from records and which come from estimates.

Use the approved plan to find who can switch to the backup and who can approve offline work. Rehearse what staff do if the backup also fails. After repair, compare errors per 100 requests over similar periods. Ask the IT lead to test the portal before staff use it again.

Try it
Use the approved service list, outage log and continuity plan below. List each affected service. Give the source record for each finding. Draft a form for staff to report the time, service, impact and evidence. Find who can approve a switch to the backup. Estimate the queue from the starting count, arrivals and completed requests. Mark all assumptions and missing data. List the actions staff can take if the backup fails.

[Paste the approved extracts.]

Check who can approve a switch to the backup

Agree on which business objective you are protecting, who can accept risk, what triggers escalation and which evidence the committee expects. These rules give the team a basis for comparing risks. The Orange Book sets out identification and assessment, treatment, monitoring and reporting, with governance and clear roles throughout. ISO 31000 offers general guidance that organisations can adapt to their context. HM Treasury, The Orange Book ISO 31000:2018, Risk management guidelines

Compare the current policy with the draft

Give AI the current policy, the draft and the list of approval roles. Ask it to show the old rule and the new rule with their clause numbers. If the draft gives the duty manager authority to switch to the backup, check the approval list for that role. Check who acts if the duty manager is unavailable.

Try it
Compare the approved policy and proposed draft by clause. For each change that affects how the policy works, cite the page and explain the change. Copy the responsible role from the approval list. Flag any change with no assigned role so the policy lead can assign an owner. Note what still needs approval or clarification. Quote only the words needed to show the change. Flag missing references and conflicts. Check the delegation schedule to confirm who can approve each change.

Give the policy owner the comparison table and an unresolved-issues log. Ask them to confirm whether a flagged gap belongs in this policy or a separate procedure. Refer legal or regulatory interpretations to the qualified reviewer.

Record the source of each outage report

Look in approved incident tickets for what happened and when. Use test results to see which controls failed more than once. Check the vendor register for services that share a provider. AI can help extract fields, suggest matches when names differ and flag records that need human follow-up.

Keep the source alongside each fact

For a portal outage, ask for the incident ID, start time, affected service, reported cause and source record. Keep the original wording for the reported cause. A note that says “service unavailable” records what a user saw. Ask the service desk to confirm the cause before you put it in the risk register.

Try it
Use only the approved extracts below. Build a table with one row per event. Include source ID, record date, event time, service affected, reported cause and evidence location. Flag possible duplicate rows and missing fields. Retain both source values when two records disagree and add a question for the data owner.

Reconcile the row count to the extracts. Spot-check the fields that drive escalation, such as start time or impact. Ask for a missing file or correction before using its contents in a report. NIST, Generative Artificial Intelligence Profile

Use interview notes to assess the impact

When loss data are sparse, workshop notes and interviews can help you understand why a risk exists. Use AI to group similar observations and turn them into draft risk statements that describe what could cause an event and what would happen if it occurred. Keep minority views and contradictory evidence visible.

Separate a recurring concern from a supported finding

Two teams report that their requests did not appear in the portal. The service desk sees error messages but has no record of lost requests. Ask AI to group the notes by service and error code. Ask the IT lead to check which users could not complete their work. Record the possible impact and the evidence you still need.

Try it
Group the approved workshop notes into draft risks. For each use a cause, a possible event and a consequence. For each link between the cause, event and consequence, quote the relevant note or give its ID. Record any conflicting view and describe what evidence the team needs to resolve it. Suggest a rating against our supplied criteria only where the notes justify it.

Take the draft back to the risk owner. Ask what would change the rating and record who will obtain that evidence. This gives the next reviewer a reason for the judgement. HM Treasury, The Orange Book

Check which teams have the highest exception rates

Bring an approved extract and a data dictionary. AI can help write a query that joins events to transaction counts, compares branches over time and flags missing or duplicate records. Begin with counts you can reconcile. You can add anomaly detection after the team agrees on what counts as unusual and which alerts merit a review. NIST’s statistics handbook covers exploratory checks and graphical techniques. NIST, Exploratory Data Analysis handbook

A branch example you can check by hand

BranchExceptionsTransactionsException rate
A402,0002%
B255005%

Branch A has more exceptions. Branch B has a higher share of transactions with an exception. For a quality review, ask whether transaction mix, reporting rules or a recent change could account for the difference before assigning an investigator. When volumes are small, keep the counts visible and compare multiple periods before acting on one month.

Turn a rate gap into an impact estimate

Use Branch B’s 5% rate as a baseline. If an approved control pilot reached 2% with the same transaction mix, the comparison would be 50 versus 20 exceptions per 1,000 eligible transactions. At an assumed J$3,000 in direct rework per exception, 30 fewer exceptions would mean J$90,000 in lower rework cost per 1,000 transactions, before the cost of the control.

Illustrative exceptions per 1,000 eligible transactions
Baseline · 50
Pilot target · 20

The 2% rate is a target for the pilot. Branch A’s rate offers a benchmark. To estimate the pilot’s effect, agree the start date, compare equivalent periods or use a suitable comparison group, and check the mix of work. Ask Finance how much of the assumed cost would leave the budget and how much would free staff time. UK Government Analysis Function, The AQuA Book

Try it
Read the approved table schemas and field definitions below. Flag duplicates and missing branch codes in the event-level extract. If you only have aggregate counts, list the checks you need the data owner to perform upstream. Write a query that counts distinct eligible transactions with an exception and divides by all eligible transactions for each branch and week. Show the numerator, denominator and rate. Flag small denominators and incomplete reporting periods. Give two possible reasons for a spike. For each, say which records you would check.

[Paste approved field definitions and table schemas. Include any approved aggregate counts.]

Estimate the range of outage costs

Use a model the team can check. A Monte Carlo simulation draws new values for inputs such as cost per request on each run. The results show how costs can vary under the assumptions you used. Ask AI to help specify the model, document the assumptions and draft checks. Use a spreadsheet or code you can inspect for the calculation. The US Government Accountability Office’s cost estimating guide describes simulation and sensitivity analysis as ways to examine cost uncertainty. US Government Accountability Office, Cost Estimating and Assessment Guide

Estimate the cost of delayed requests

Suppose an outage delays new requests and management wants a view of extra handling costs per outage. Try a triangular input for delayed requests with a minimum of 200, most likely value of 400 and maximum of 800. For extra cost per request, use J$500, J$900 and J$1,500 in the same order. These are teaching values. With independent inputs and total cost equal to delayed requests multiplied by extra cost per request, the expected cost is about J$451,111 per outage. The mathematical endpoints are J$100,000 and J$1.2 million. Those checks give you something to compare with the model output. In one Python example with 10,000 draws and seed 20261118, the median was about J$428,000 and the 90th percentile about J$664,000. About 4.4% of draws crossed J$750,000. The example treats request counts as continuous for a first estimate.

Try it
Help me specify a Monte Carlo model for extra handling cost for delayed requests per outage. Use these example inputs: delayed requests, triangular (200, 400, 800); extra cost per request in J$, triangular (500, 900, 1,500). Multiply the inputs to get total cost. Start with independence and explain why it may fail when a longer outage raises both volume and overtime. Specify 10,000 draws and record the seed or save the draws if the tool cannot set a seed. Use 20261118 as the Python seed. State whether you round request counts. Give reproducible spreadsheet or Python steps. Report the median, 90th percentile and share of trials above J$750,000. Add checks for the theoretical endpoints and mean, and propose one separately labelled sensitivity case in which both inputs draw from the same percentile. Label what is calculated, assumed or still unknown.

Ask Finance and Operations to replace the practice ranges with operating data. Reconcile the inputs to their records and inspect how dependence changes the upper tail. The mean describes cost per outage under the model assumptions. A contingency budget for the year also needs outage frequency, the costs left outside the model and the decision owner’s tolerance for shortfalls.

Test when the continuity plan needs more staff

A scenario is a specified set of conditions. Use it to ask whether a plan still works if those conditions hold. In a Monte Carlo simulation you sample many possible values. Here you pick the conditions on purpose and walk through the decisions they would force. The Basel Committee’s principles describe stress testing as a risk management tool with defined objectives, governance and methods. Basel Committee on Banking Supervision, Stress testing principles

For a Caribbean organisation, try a hurricane scenario in which two branches close, mobile service is unreliable and request arrivals triple. Suppose the backlog can grow by up to 300 requests before the escalation trigger. Start at 200 daily arrivals and the capacity to complete 250 requests a day in normal operations. With 600 arrivals and 250 completions, the queue grows by 350 on day one from an empty starting queue. Can the person who approves extra staffing reach the team before the queue crosses 300?

Try it
Using the approved continuity plan, build a three-day decision exercise for a hurricane. Start with an empty request queue, 600 arrivals and 250 completions per day, a backlog trigger above 300 requests, two branches closed and unreliable mobile service. Show the queue by day under current staffing. Then test one alternative that adds 200 completions per day from day one. Identify the first threshold breach for each case, the role the plan gives authority to approve extra staffing and one dependency that could block the response. Mark any missing owner, limit or contact route.

[Paste the approved plan extract.]

At the end of each day, the backlog under current staffing is 350, 700 and 1,050 requests. With extra capacity it is 150, 300 and 450. The first day above 300 moves from day one to day three. Ask the operations lead whether the added staff can access the backup system in time.

Choose a control you can test

Ask AI to help compare options by expected effect, implementation effort, process delay and the evidence that would show the control works. The risk owner decides which option to adopt and which exposure to accept. Record who can approve the decision and when to revisit it. HM Treasury, The Orange Book

Compare an approved offline form with an approved backup portal. Check which option lets staff complete urgent requests and keep a record of each action. Ask the service owner to agree what work staff may complete offline. Test how staff match those records to the portal when it returns.

Try it
Compare two proposed controls using the approved incident summary and control descriptions below. For each control, state what could go wrong and how the control would help. Say who would use it and how much it could delay routine work. Give one way to test it with approved records. State which facts remain unknown and what we will measure in the pilot.

Agree an acceptance criterion before the pilot. Keep control test evidence and open defects alongside the risk owner’s decision. Use the cyber tabletop below to check roles, escalation and what the recovery team needs to see.

Rehearse who takes charge when a system goes down

Use AI to draft a tabletop exercise from approved excerpts of the incident response plan. A tabletop is a facilitated discussion. Its record can show which decisions staff made, which evidence they could produce and where the plan left them stuck. Ask your security lead to validate the exercise and use recorded evidence from separate authorised technical tests to assess recovery. NIST, Guide to Test, Training, and Exercise Programs for IT Plans and Capabilities

Test who decides when the portal can reopen

The provider reports an outage while staff have urgent requests to complete. The account owner is unavailable and the provider cannot yet confirm when service will return. Ask the team who declares an incident, which backup channel they can use and what the recovery lead needs before approving a return to service.

Try it
Draft a 30-minute, discussion-only tabletop for a work portal outage. Use this approved plan extract and the three facts supplied: urgent requests are due today, the account owner is unavailable and the provider has given no recovery estimate. Add two timed updates about business impact or response coordination. For each decision point, give a facilitator question, the accountable role named in the plan, observable evidence, and a pass or gap criterion. Mark gaps in the plan as open. End with an action log. Leave space for the exercise lead to assign an owner and set a due date for each gap. Keep the exercise to incident response and recovery decisions.

[Paste the approved plan extract.]

Ask an observer to timestamp the escalation decision and record the evidence each team cites. If the backup number is missing or nobody can find the rules for reopening the portal, assign an owner and set a due date. At the next review, ask for the updated contact list or approval checklist.

Check when the outage error rate exceeds its limit

Choose the error rate that calls for an investigation. Set how often staff should check it and which records to use. Assign an owner to check the rate and tell the IT lead when the rule calls for an investigation. Ask AI to draft the calculation or explain why the number moved in the weekly report. Calculate the indicator from controlled data and test the formula against known cases. HM Treasury, The Orange Book UK Government Analysis Function, The AQuA Book

Suppose the team investigates an error rate of 4% or more for two complete review periods. Each period must include at least 500 requests. Check the total requests and the error count before you raise an alert. Ask the IT lead to confirm whether the same errors recur after repair.

Try it
Use the weekly summary below to identify weeks eligible for review. Flag missing counts and weeks below 500 eligible transactions. For each qualifying two-week sequence, calculate both rates and flag a breach if both are at least 4%. Provide the two source row IDs and the escalation owner from the agreed procedure.

Keep a simple log of alerts, investigations and changes to the control. Check that a quiet week means the indicator was observed and stayed inside its limit. Record gaps in monitoring explicitly.

State the decision in the risk report

Build the first page from a checked risk register, the latest indicator figures and the action log. Ask AI to draft what changed since the last meeting and which decision is due. Keep the team’s evidence and approval trail attached to each action.

For Branch B, put 25 exceptions in 500 transactions next to the 5% rate. Add the previous period if the denominator and data coverage are comparable. Record whether the proposed control is awaiting approval, under test or in use. Before sending the report, check who has each action and when it is due. Assign an owner to any action that still needs one.

Try it
Draft the first page of the risk report from these approved records. Open with the decision due. For each material change give the current figure, comparable prior figure if available, source ID and next action. Keep estimates labelled. Copy action owners and dates from the approved log.

Use the communication instructions below to turn that checked report into versions people can act on.

Write an outage update for each role

Tell the director which approval you need and when. Give AI the incident timeline and risk register entry to draft the brief. Ask for a separate handover if the staff handling the outage need different instructions.

Use the same facts for each update

The incident lead confirms that the work portal is down. The operations lead estimates 80 requests in the queue. The IT lead has no confirmed time for repair. Give AI this checked incident record and ask for one update for each role.

An executive needs to set service priorities. A manager needs to move staff to the backup. An officer needs to record urgent requests. An administrator needs to update the incident log. Ask AI to draft a short presentation, an animation script or a staff message in the audience’s working language. Ask a reviewer who knows the language to check the instructions before release.

Try it
Use the approved incident record below. Write one short update each for the executive, manager, officer and administrator. State the action for that role in the first sentence. Keep the outage status, queue estimate and unknown points the same in every version. Give the source for each figure. Use the role list to say who will send the next update and where staff can find it. If no role is listed, flag the gap so the incident lead can assign an owner. Use the working language of each audience. If you draft a slide or animation script, add a plain-text version for review.

[Paste the approved incident record and role list.]

Trace each number to its source. Ask the risk owner to confirm the decision request and the due date. The National Institute of Standards and Technology (NIST) lists confidently stated errors among generative AI risks. A line that reads cleanly still needs a source. NIST, Generative Artificial Intelligence Profile

Try it on the next risk pack

Try this on one item in your next risk report. Give AI the relevant rows from the approved risk register and ask it to draft a short brief. Count the time it takes to write and check. Ask a colleague to read the brief and tell you what needs approval and who can give it.

In a speech on AI and financial stability, Bank of England Deputy Governor Sarah Breeden said “a person can be held to account.” Give the decision owner a record they can use. Keep the approved inputs, instructions, checked output and reviewer with the brief. Sarah Breeden, Engaging with the machine: AI and financial stability

If the AI adds a statement you cannot trace, remove it before the brief goes out. If the draft takes longer to check than your normal process, narrow the task to extracting facts with row references and try again on the next pack.

Sources