AI in Risk Management: A Practical ERM Playbook

2026-10-16 · Tommaso Maria Ricci

In the third quarter of 2025, Gartner asked 184 senior enterprise risk executives to rank their top emerging risks. AI-related information governance risk jumped from fourth to second place, and shadow AI climbed from fifth to third. Read that again: the people whose job is to manage risk now see AI as one of the biggest risks they face. At the same time, most of them are under pressure to use AI in risk management itself, to monitor more, faster, with the same headcount.

That tension is the whole story of AI in risk management in 2026. Used well, AI lets a small risk team see signals that used to be invisible: a supplier quietly heading toward insolvency, a control that stopped working three months ago, a regulatory change buried in a 200 page consultation. Used badly, it adds a new layer of model risk on top of everything else, with outputs nobody can explain to the board.

This guide is for CFOs, CROs, heads of risk and founders who want the first outcome without the second. It covers where AI actually adds value across the risk cycle, a step by step method for how to use AI in enterprise risk management, the governance frameworks that matter, a self-assessment scorecard and a 90 day roadmap.

What AI in risk management actually means

The phrase covers two different problems, and confusing them is the first mistake most companies make.

AI for risk management means using machine learning, language models and automation to do risk work better: identifying risks earlier, assessing them with more data, monitoring them continuously and reporting them faster.

Risk management of AI means governing the risks that AI itself creates: inaccurate outputs, biased decisions, data leaks, regulatory exposure, unauthorized tools. I cover that side in depth in the guide to AI governance for business and in the playbook on shadow AI risk management.

This article is mainly about the first problem. But you cannot do the first without the second. The moment you use AI to score supplier risk or flag control failures, that AI becomes a model your risk function owns, with its own failure modes. Every section below keeps both sides in view.

Why now

Three things changed between 2023 and 2026.

1. Unstructured data became usable. Most risk signals live in text: contracts, emails, incident reports, news, regulatory updates, audit findings. Language models can now read, classify and summarize that text at a cost that makes sense for a mid-sized company.
2. Adoption moved faster than control. AI tools spread through organizations bottom up. Gartner's quarterly emerging risk surveys show risk leaders reacting to exactly that: by the fourth quarter of 2025, with 367 executives surveyed, the firm reported growing awareness of risks tied to the rapid growth of AI use inside and outside the organization.
3. Value is still concentrated. According to McKinsey's State of AI survey, 37% of respondents attribute at least some EBIT impact to AI, and the share of AI high performers, those attributing at least 5% of EBIT to AI, has stayed flat at about 6%. Most companies use AI. Few have turned it into measurable results.

Risk management is one of the functions where the gap between experimenting and producing results is easiest to close, because the work is structured, repetitive and already documented.

Where AI adds value across the risk cycle

Every enterprise risk framework, whether you use COSO, ISO 31000 or your own, follows roughly the same cycle: identify, assess, respond, monitor, report. Here is where AI earns its place in each step, and where it does not.

Identify: risk sensing from unstructured sources

The classic risk register is updated once or twice a year in workshops. By the time a risk is written down, it has often already happened.

AI changes the cadence. Language models can scan sources continuously and surface candidate risks for a human to review:

• news and media on suppliers, customers, competitors and key markets;
• regulatory publications and consultations relevant to your sectors and jurisdictions;
• internal text: incident reports, customer complaints, audit findings, whistleblowing summaries, project status reports;
• contracts, to flag unusual clauses, concentration and renewal exposure.

The output is not a new risk register. It is a shortlist of signals, each with its source, that the risk owner accepts, rejects or escalates. The value is breadth: a two person risk team cannot read thousands of documents a month, but it can review fifty well chosen alerts.

Assess: better estimates with more data

Assessment is where most registers are weakest. Likelihood and impact are often scored on a five point scale by gut feel, and the same risk gets a different score depending on who is in the room.

AI helps in three concrete ways:

1. Historical analysis. Machine learning models can estimate the frequency and severity of operational losses, payment delays or quality incidents from your own data, giving the scoring discussion a factual anchor.
2. Consistency checks. A language model can compare how similar risks are scored across business units and flag outliers for discussion.
3. Scenario generation. Models can draft plausible stress scenarios, for example a combined supplier failure and currency shock, which the team then refines and quantifies.

None of this replaces judgment. It makes judgment better informed and easier to challenge.

Respond: faster, documented decisions

Once a risk is assessed, someone decides to accept, reduce, transfer or avoid it. AI is useful here mostly as an assistant: drafting mitigation plans based on past responses, checking that each action has an owner and a deadline, and comparing the plan against policy requirements. The decision stays with the risk owner.

Monitor: continuous control and KRI tracking

This is where AI delivers the most measurable value in most companies.

• Key risk indicators (KRIs) can be tracked automatically from source systems, with anomaly detection flagging unusual movements instead of waiting for a monthly threshold breach.
• Control monitoring can move from sample testing to full population testing for transactional controls: segregation of duties conflicts, duplicate payments, approvals outside policy, unusual journal entries.
• Third party monitoring can combine financial signals, news and delivery performance into an early warning score for critical suppliers.

The logic is the same one internal audit is moving toward, and I describe it in the guide to AI in internal audit. Risk and audit should share the same monitoring data rather than build two separate systems.

Report: from data to board narrative

Risk reports take a surprising amount of time to produce, and the board often reads only the first page. Language models can draft the narrative from structured data: what changed since last quarter, which KRIs moved, which risks were escalated and why. The risk team edits, verifies every figure and signs off.

Two rules apply. Every number in a draft must trace back to a source system. And the model should never be the last reader before the board.

Where AI does not help much yet

• Strategic and reputational risk judgment. AI can gather signals, but weighing a risk against strategy is a leadership conversation.
• Low frequency, high impact events. Models trained on history are weak on events that have rarely or never happened.
• Anything without data. If your incidents are not logged and your controls are not documented, AI has nothing to learn from.

AI risk management use cases by business function

The risk cycle is the same everywhere, but the data and the decisions differ by function. These are the use cases I see working in mid-sized companies, grouped by who owns the risk.

Finance and credit

Finance owns some of the most data rich risks in any company, which makes it a natural starting point.

• Customer credit risk. Payment behavior, order patterns and external signals can feed an early warning score that tells the credit team which accounts to call this week, before an invoice becomes a write-off.
• Liquidity risk. Forecasts of cash positions under different scenarios help treasury see a squeeze coming weeks in advance.
• Fraud and payment controls. Anomaly detection on payments and vendor master data catches duplicate invoices, changed bank details and unusual approval paths.

The common thread is that the data already exists in the ERP. The work is connecting it and deciding who acts on each signal.

Operations and supply chain

Supply chain risk is where unstructured data matters most. A supplier rarely announces trouble in a structured field.

• Supplier early warning combines delivery performance, quality incidents, financial filings and news into one view of critical suppliers.
• Concentration analysis maps which products depend on a single supplier, plant or region, so that a disruption scenario can be quantified rather than guessed.
• Logistics disruption monitoring tracks events such as port closures, strikes or extreme weather and matches them to your open orders.

Technology and cyber

Security teams have used machine learning for years in threat detection. The newer opportunity is connecting cyber signals to enterprise risk: translating a spike in phishing attempts or a critical vulnerability into business impact that the board understands. Language models help write those translations, as long as a security professional validates them.

Compliance and legal

Regulatory change management is one of the clearest wins. Models can read new regulation and guidance, summarize what changed and suggest which policies and teams are affected. Contract review for risk clauses, such as liability caps, termination rights and data processing terms, is another. In both cases a lawyer or compliance officer confirms the conclusion before anything changes.

People and HR

People risk includes key person dependency, turnover in critical roles and health and safety incidents. AI can analyze incident reports and near misses to find recurring causes, and can flag teams where workload or turnover suggests rising risk. This is also the area where caution matters most: any AI that influences decisions about individual employees deserves careful legal review, and in Europe it may fall into the high-risk category of the AI Act.

Choosing your first function

The best first function is usually the one with the cleanest data and the most frequent decisions. For most companies that means finance or procurement, not strategy or reputation. Start where you can measure results within a quarter, then use that credibility to expand.

How to use AI in enterprise risk management: a six step method

This is the method I use with leadership teams. It works for a 50 person company and for a group with several subsidiaries; only the scale changes.

Step 1: Start from a decision, not a tool

Pick one risk decision that is made regularly and badly today. Good candidates:

• which suppliers need a deeper review this quarter;
• which control failures need escalation;
• which regulatory changes affect which business unit;
• which customers are drifting toward payment default.

Write the decision down, who makes it, how often, with what information. If you cannot describe the decision in two sentences, you are not ready to automate any part of it.

Step 2: Map the data behind the decision

For the chosen decision, list every data source involved: systems, spreadsheets, documents, people's inboxes. Then answer three questions for each source:

1. Is the data complete and reasonably clean?
2. Who owns it and can grant access?
3. Does it contain personal or confidential data that limits how it can be processed?

Most AI projects in risk fail at this step, quietly, months later. A model that scores supplier risk on incomplete purchasing data gives confident answers that are wrong.

Step 3: Choose the simplest technique that works

Not every risk problem needs a language model. A rough guide:

Start with the simplest option. Complexity is itself a risk.

Step 4: Design human review into the workflow

Every AI output in risk management should have a named human reviewer, a defined action and a record. In practice:

• alerts go to a queue owned by a specific person, not to a shared inbox;
• each alert is accepted, rejected or escalated, with a short reason;
• the reasons are reviewed monthly to tune the system and spot drift.

This also protects against automation bias, the tendency to trust a machine output because it looks precise. If reviewers accept 99% of alerts without comment, either the system is excellent or nobody is really reviewing.

Step 5: Validate before you rely on it

Before an AI component influences a risk decision, test it against history. Take the last 12 to 24 months, run the model as if it had been live and compare what it would have flagged with what actually happened. Measure:

• how many real issues it caught;
• how many false alarms it raised;
• how much earlier it caught them than the existing process.

Document the result. That document is what you show the audit committee and, increasingly, regulators.

Step 6: Run, measure, extend

Run the pilot for one quarter with clear metrics, then decide to extend, adjust or stop. Only after the first use case is stable should you add a second. Companies that launch five AI risk use cases at once usually end up with five half working tools and a risk team that trusts none of them.

If you want a second pair of eyes on which risk decision to start with, you can reach me through the consultation request page on this site; the first conversation is usually enough to tell whether there is a project worth doing.

The new risks AI brings into the risk function

Using AI for risk management creates risks of its own. A credible program names them up front.

Model risk

Any model that influences decisions can be wrong in ways that are hard to see. Banks have managed model risk for years with independent validation, inventories and ongoing monitoring. Non-financial companies now need a lighter version of the same discipline: a list of every model in use, who owns it, what it is used for, how it was validated and when it was last reviewed.

Inaccurate and invented outputs

Language models can produce fluent, confident text that is wrong. In risk work this shows up as a summary that misstates a regulation, a contract review that misses a clause or a report that cites a number that does not exist. The NIST Generative AI Profile, published in July 2024, lists confabulation among the 12 risks that are unique to or made worse by generative AI. The mitigation is procedural: source links on every output, mandatory verification of figures and a reviewer who signs off.

Data protection and confidentiality

Risk data is among the most sensitive a company has: incidents, investigations, legal exposure, personal data of employees and customers. Before any of it goes into an AI tool, you need to know where it is processed, whether it is retained and whether the vendor can use it for training. The answer for most consumer AI tools is not acceptable.

Explainability

If a model flags a supplier as high risk, the procurement team will ask why, and so will the supplier. Choose techniques whose outputs can be explained in business terms, especially for decisions that affect people or contracts.

Concentration and vendor risk

Many AI features now come from a small number of model providers. If your monitoring, reporting and contract review all depend on the same provider, an outage or a pricing change becomes a single point of failure. Treat AI vendors as critical third parties, with the same due diligence as any other.

Shadow AI in the risk team itself

Risk professionals are not immune. Analysts pasting incident details into a public chatbot to draft a summary is exactly the behavior the risk function is supposed to prevent elsewhere. Provide approved tools first, then enforce the policy.

Governance frameworks that matter

You do not need a new framework for AI in risk management. You need to connect the ones you already use.

COSO ERM and ISO 31000

COSO's Enterprise Risk Management framework, updated in 2017 as "Integrating with Strategy and Performance", organizes ERM into five components and twenty principles. ISO 31000:2018 provides a shorter set of principles and a process. Both are technology neutral. AI fits inside them as a set of tools for identification, assessment and monitoring, and as a new category of risk to manage.

NIST AI Risk Management Framework

The NIST AI Risk Management Framework, released in January 2023, is voluntary and organized around four functions: Govern, Map, Measure and Manage. It is the most practical public reference for managing the risks of AI systems themselves. The July 2024 Generative AI Profile adds suggested actions specific to generative models. Even outside the US, many companies use it as a checklist because it is detailed and free.

The EU AI Act, if you operate in Europe

The EU AI Act entered into force on 1 August 2024, and its timeline was amended by Regulation (EU) 2026/1744, in force since 27 July 2026. For risk teams, the key dates are:

• 2 February 2025: prohibited practices and the AI literacy provision (Article 4), which asks providers and deployers to take measures that support the AI literacy of their staff, without setting a specific level;
• 2 August 2026: transparency obligations under Article 50, for example telling people when they interact with an AI system;
• 2 December 2027: obligations for high-risk systems listed in Annex III, which include AI used for creditworthiness assessment of individuals, for risk assessment and pricing in life and health insurance, and for employment decisions;
• 2 August 2028: obligations for high-risk AI embedded in products covered by EU product legislation (Annex I).

Most internal risk management uses, such as monitoring controls or reading regulation, are not high-risk under the Act. But if your AI touches credit decisions on individuals, insurance pricing or HR decisions, it may be. I cover the compliance side in the guide to AI for compliance.

What the board should see

A one page view, updated quarterly:

1. the list of AI use cases in risk management, with owner and status;
2. validation results for each use case that influences decisions;
3. incidents and near misses involving AI outputs;
4. the top AI-related risks across the company, including shadow AI;
5. regulatory dates coming up in the next 12 months.

Self-assessment: is your risk function ready for AI?

Score each statement from 0 (not true) to 2 (fully true). Add up the points.

Data

1. Our risk register is digital, current and owned by named people.
2. Incidents and losses are logged consistently in one system.
3. Key risk indicators are defined with thresholds and data sources.
4. We know where our sensitive risk data lives and who can access it.

Process

5. Risk assessments use a consistent methodology across business units.
6. Controls are documented with owners and testing frequency.
7. Escalation paths are clear and actually used.

People

8. Someone in the risk team has time to own a pilot for a quarter.
9. The team has basic literacy on what AI can and cannot do.
10. Business owners trust the risk function enough to act on its alerts.

Governance

11. We have an inventory of AI tools used in the company, including unofficial ones.
12. We have a policy on which data can be used in which AI tools.

How to read your score (maximum 24)

• 0 to 8: Foundation first. Fix the register, incident logging and data ownership. AI on top of messy data produces confident noise.
• 9 to 16: Ready for one pilot. Pick a monitoring use case with clean transactional data, such as control testing or supplier early warning.
• 17 to 24: Ready to scale. Run two use cases in parallel, build a model inventory and bring validation results to the audit committee.

A 30/60/90 day roadmap

This roadmap assumes a mid-sized company with a small risk team and one pilot use case.

Days 1 to 30: decide and prepare

• choose one risk decision using Step 1 of the method;
• map data sources and resolve access and privacy questions;
• define success metrics: detection rate, false alarm rate, time to detection, hours saved;
• agree on the human review workflow and name the reviewer;
• check AI vendors for data processing terms before any data moves.

Days 31 to 60: build and backtest

• build the first version with the simplest technique that could work;
• backtest on 12 to 24 months of history and document the results;
• run the system in parallel with the existing process, without acting on it yet;
• hold a weekly review of alerts with the business owner.

Days 61 to 90: go live and decide

• switch on the alerts in the real workflow, with human review on every one;
• measure the agreed metrics and compare with the baseline;
• present results to the CFO or audit committee with the validation document;
• decide: extend to a second unit, adjust, or stop.

After 90 days you should have one working use case, a validation record, a review workflow and the data to decide what comes next. That is more than most companies have after a year of AI strategy workshops.

Common mistakes to avoid

The same errors appear in almost every company that struggles with AI in risk management.

1. Buying a platform before choosing a decision. A tool without a defined decision becomes an expensive dashboard nobody opens.
2. Skipping the backtest. Without testing on history, you cannot tell the audit committee whether the system works, and you cannot defend it when it misses something.
3. Sending alerts to everyone. Alerts without a named owner are ignored within weeks. One queue, one owner, one documented outcome per alert.
4. Treating AI output as evidence. A model summary is a pointer to evidence, not evidence itself. Auditors and regulators will ask for the source.
5. Forgetting the risk of the tool. The AI you deploy for risk management must itself appear in your model inventory and your AI risk register.
6. Measuring activity instead of outcomes. The number of alerts generated says nothing. The number of issues caught earlier, and the losses avoided, say everything.

KPIs that show whether it works

Measure the outcome, not the activity. Useful indicators:

Detection

• share of material issues flagged before they caused a loss;
• average time between the first signal and escalation;
• issues found by AI that the existing process missed.

Efficiency

• hours per month spent on monitoring and reporting;
• share of controls tested on the full population rather than on samples;
• time to produce the quarterly risk report.

Quality

• false alarm rate and its trend;
• share of alerts closed with a documented reason;
• errors found in AI-drafted reports before sign-off.

Trust

• share of alerts the business acts on;
• audit committee and external auditor feedback on the evidence produced.

If false alarms grow and business action falls, the system is losing credibility. Fix it before adding anything new.

Build, buy or extend what you have

Three paths exist, and the right one depends more on your data than on your ambition.

Extend existing platforms. Your GRC suite, ERP and analytics tools increasingly include AI features. If they cover your first use case and your data already lives there, this is the fastest and lowest risk start.

Buy a specialized tool. Third party risk monitoring, regulatory change management and transaction monitoring have mature specialist vendors. Choose this path when the specialist's data or models are clearly better than what you can build, and check integration and data terms carefully.

Build with your team. Makes sense when the decision is specific to your business and your data is your advantage, for example a supplier risk model built on your own delivery and quality history. It requires someone who can maintain the model after the pilot.

Whatever the path, ask every vendor the same questions:

• where is our data processed and stored, and is it used for training?
• how was the model validated, and can we see the results?
• can every output be traced to its source?
• what happens to our data and configurations if we leave?
• who is accountable when the system misses something?

Connect this work with finance. Many early warning signals, such as customer payment behavior and liquidity pressure, are also inputs for treasury, as I explain in the guide to AI cash flow forecasting.

What I have learned from AI projects that worked

I am a founder, not a theorist, and the AI projects I have led share a pattern that applies directly to risk.

At a sports distribution company, AI driven marketing contributed to a 30% increase in sales. At a hotel, revenue grew from 9 to 10 million. A medical center increased capacity by 20% with the same staff, and an agritourism business doubled its guests. Different sectors, same lesson: the results came from making an existing process disciplined and measurable, then using AI to run it at a scale people alone could not.

Risk management is the same. The companies that get value from AI in risk do not start with the most advanced model. They start with a clear decision, clean data, a named reviewer and a way to measure results. The technology is the easy part.

The hard part is organizational: getting business owners to act on alerts, getting the board to read a risk report, and making sure the risk function does not become the department that creates new risks with its own tools. A structured adoption approach helps, and the enterprise AI adoption framework describes how to sequence it across functions.

If you are deciding where AI fits in your risk function and want an outside view built on what has worked in practice, use the consultation request page on this site to describe your situation. I will tell you honestly whether there is a project worth doing now.

FAQ

How do you use AI in enterprise risk management?

Start from one risk decision that is made regularly and poorly today, such as which suppliers need review or which control failures need escalation. Map the data behind it, choose the simplest technique that works, design human review into the workflow and backtest the system on 12 to 24 months of history before relying on it. Run a one quarter pilot with clear metrics, then extend, adjust or stop. Add a second use case only after the first is stable.

What are the main use cases of AI in risk management?

The most valuable uses are continuous monitoring of key risk indicators and controls, early warning on suppliers and customers, scanning news and regulation for emerging risks, reading contracts for unusual clauses and drafting risk reports from structured data. AI is less useful for strategic judgment and for rare, high impact events that have little history. In every use case, a named person should review AI outputs before they influence a decision.

Will AI replace risk managers?

No. AI takes over the high volume work: reading documents, tracking indicators, testing transactions and drafting reports. Risk managers keep the parts that require judgment, such as weighing risks against strategy, challenging business owners and deciding what reaches the board. In practice, AI lets a small risk team cover far more ground, but it also adds model risk that someone with risk expertise must govern. The role shifts toward oversight, validation and decision support.

What risks does AI itself create for the risk function?

The main ones are model risk, inaccurate or invented outputs, leaks of confidential data, poor explainability, dependence on a few AI vendors and unauthorized tool use by staff. NIST's Generative AI Profile from July 2024 lists 12 risks unique to or made worse by generative AI, including confabulation. Mitigations are mostly procedural: a model inventory, validation before use, source links on every output, mandatory human review and clear rules on which data can enter which tools.

Is AI used in risk management high-risk under the EU AI Act?

Most internal uses, such as monitoring controls, scanning regulation or drafting reports, are not high-risk under the EU AI Act. AI used to assess the creditworthiness of individuals, to price life and health insurance or to make employment decisions falls under Annex III, with obligations applying from 2 December 2027 after the amendments of Regulation (EU) 2026/1744. Article 4 on AI literacy has applied since 2 February 2025 and asks companies to take measures supporting staff literacy.

Which framework should we use to govern AI in risk management?

Keep your existing ERM framework, whether COSO ERM or ISO 31000, and treat AI both as a set of tools inside it and as a new risk category. For the risks of AI systems themselves, the NIST AI Risk Management Framework, released in January 2023 and organized around Govern, Map, Measure and Manage, is the most practical public reference. Companies operating in Europe should map their use cases against the EU AI Act timeline.

How long does it take to see results from AI in risk management?

A focused pilot can show measurable results within 90 days: one month to choose the decision and prepare data, one month to build and backtest, one month to run live with human review. Results depend more on data quality than on the technology. Companies with a digital risk register, consistent incident logging and defined key risk indicators move fastest, while those without that foundation should fix it first.