Implementation

Rules or AI in risk management: when to use each

· 13 min read ·

In risk management, "using AI" is not a single decision. It is a task-by-task choice between three different tools: rules written by people, machine learning models trained on historical data, and large language models (LLMs) that read and write text. Each gets things right and wrong in its own way. This article compares the three on concrete risk and compliance tasks, sets out how each one fails and shows when a simple rule is still the best answer.

This article is not about how to choose a platform or what makes an architecture "AI-native". Those topics are covered in Enterprise AI platform for GRC: what to evaluate and AI-native platform: what it means for GRC. The question here is more operational: for each task, which mechanism makes sense and what can go wrong with it.

Glowing digital security padlock on a technology background

Three different mechanisms under one label

Much of what is sold as "AI" in GRC (governance, risk and compliance) is in fact one of the three approaches below, or a mix of them. Separating them helps you ask the right questions.

  • Rules. Conditions written by people: "if a treatment action is overdue, notify the owner"; "if the vulnerability is in CISA's KEV catalogue, put it at the top of the queue". Same input, same result, every time. Any auditor can read the logic.
  • Machine learning. Models that learn patterns from examples: which audit findings tend to be rated critical, what normal access behaviour looks like for a group of users. They return an estimate, not a certainty.
  • Large language models (LLMs). Models that work with free text: they read a policy, a supplier questionnaire or a report, then summarise, classify and draft. They produce fluent text even when they are wrong.

The NIST AI Risk Management Framework (AI RMF 1.0), a voluntary framework published in January 2023, devotes a whole appendix (Appendix B) to how AI risks differ from those of traditional software. The points it lists include training data that becomes stale relative to the deployment context, greater difficulty in predicting failure modes of large pre-trained models and the need for more frequent maintenance because of data, model or concept drift. A rule has none of these problems. It has others.

Task-by-task comparison

The table summarises where each approach tends to work best on recurring risk tasks. The detail for each row follows.

TaskA rule handles wellWhere ML or an LLM helpsMain caution
Deadlines, expiries and SLAsEverything: expired evidence, overdue action, pending annual reviewAlmost nothingAI here only adds uncertainty
Classifying findingsStructured fields: source, severity, assetFree text from audits, penetration tests and incidents classified into a taxonomyInconsistent classes for similar cases
Suggesting control mappingsMappings already approved, applied againAn LLM proposes matches between a control and requirements in other frameworksSimilar wording is not equivalence; a cited requirement may not exist
Anomaly detectionKnown thresholds: privileged access out of hours, volume above a ceilingML learns the normal pattern and flags deviations nobody wrote downFalse positives and drift
Vulnerability prioritisationFixed triggers: listed in KEV, exposes a critical internet-facing assetOrdering the rest of the queue by combining many signalsOrder hard to explain without a recorded rationale
Drafting reports and summariesReport templates filled from the databaseAn LLM drafts the executive summary from the recordsConfabulation: invented figures, facts and citations
Reading supplier documentsPresence checks: the questionnaire arrived, the report is in dateAn LLM extracts answers and points out gaps in long documentsDocument content manipulating the model (prompt injection)

Classifying findings

A rule classifies well what already arrives structured: a critical penetration test finding on an asset flagged as critical goes to the top category, no debate. The problem is free text. Internal audit findings, incident reports and notes from supplier assessments rarely follow a pattern, and a keyword rule fails predictably ("password" appears in a policy finding and in a data leak alike).

Here an ML model trained on past classifications, or an LLM guided by the organisation's taxonomy, suggests the category, the affected control and the severity. The caution is consistency: two almost identical findings can receive different classes, and the model will not warn you. The practice that works is to treat classification as a suggestion with human approval and to review a sample of accepted classifications periodically to measure the error rate.

Suggesting control mappings

Mapping an internal control to requirements in several frameworks is comparative reading, and LLMs are good at it as a starting point: given the control "data at rest is encrypted", the model proposes the requirements in ISO/IEC 27001, the NIST CSF or an internal policy that deal with the same topic. That shortens the work for anyone starting from scratch.

Two risks need attention. The first is mistaking similar vocabulary for equivalent requirements: two requirements can both mention encryption and demand different things. The second is confabulation, where the model cites a requirement identifier that does not exist or attributes to a standard something it does not say. Once reviewed and approved by someone who knows the framework, the mapping stops being a suggestion and becomes data. From then on, applying it again is rule work, not AI work.

Anomaly detection

Threshold rules detect well what you already know to look for: administrative access at 3 a.m., an export above a set volume, a configuration change outside a change window. As we show in Continuous internal audit: how to automate it, ML comes in for what nobody wrote down beforehand: a spending, access or approval pattern that departs from that user's or process's history.

The price is false positives, and it is higher than it looks when the real event is rare. A hypothetical example: 10,000 events a day, of which 10 are genuine problems. A model that catches 90% of real cases with a 2% false positive rate flags 9 true cases and around 200 false ones. More than 95% of alerts would be noise, even with a model that looks good on paper. That is why the metric that matters to the team is the share of alerts that are confirmed, not the advertised accuracy.

The second price is drift. The model learnt what "normal" looked like before a cloud migration, a reorganisation or a new system, and starts getting things wrong without any signal. The AI RMF addresses this in subcategories MEASURE 2.4 (monitoring the system's behaviour in production) and MANAGE 4.1 (post-deployment monitoring plans, with mechanisms for appeal, override and change management).

Data analysis dashboard with charts on a laptop screen

Vulnerability prioritisation

Some priority decisions should not depend on any model. If a vulnerability is listed in CISA's Known Exploited Vulnerabilities (KEV) catalogue and affects a critical internet-facing asset, a rule puts it at the top, predictably and with an easy justification. AI is useful for ordering the rest of the queue, combining signals such as exposure, asset criticality, compensating controls and the data processed. The case for prioritising by exploitation and context, not just the CVSS score, is in 48,000 CVEs a year: where to start.

Drafting reports and summaries

Drafting is where LLMs save the most time and where errors are most dangerous, because well-written text looks true. The NIST Generative AI Profile (NIST AI 600-1, July 2024) defines confabulation as the production of confidently stated but erroneous or false content, colloquially known as "hallucination". It explains that this is a natural result of how generative models are designed, predicting the next word from statistical patterns, and warns that outputs may include confabulated logic or citations that purport to justify the answer.

In a risk report, that means a figure that is not in the database, a control described as effective without evidence or an invented regulatory reference. Four safeguards reduce the problem:

  1. Limit the source. The model drafts from the records in your own database (risks, controls, evidence), not from its general knowledge.
  2. Require the origin. Every statement in the draft points to the record it came from; anything without an origin is cut.
  3. Check with a rule. Figures, dates and requirement identifiers cited in the text are checked automatically against the database. Here the rule watches the AI.
  4. Review before sending. No generated text goes to the committee or the board without being read by the person accountable for it.

Reading supplier documents

Extracting answers from long questionnaires, assurance reports and third-party policies is a good task for LLMs. There is, however, a risk specific to it: the document being read may contain instructions that change the model's behaviour. OWASP classifies this as prompt injection (LLM01:2025) and describes the indirect form, which occurs when the model takes in content from external sources such as websites or files. In practice, text extracted from a supplier document should be treated as untrusted data and checked by an analyst before it changes that third party's risk assessment.

Failure modes: how each approach goes wrong

Choosing between a rule and AI is largely choosing which kind of error you would rather manage.

  • Rules age silently. A rule written for the environment of two years ago keeps running without ever firing, and nobody notices. Test each critical rule against a known case from time to time. Another symptom is exception sprawl: when a rule needs dozens of "unless" clauses, the problem probably calls for a different approach.
  • ML loses fit and creates noise. Drift and false positives are the two most common failure modes. Track the rate of confirmed alerts and the rate of rejected suggestions over time; a gradual decline is a sign of drift.
  • LLMs are wrong with confidence. Confabulation, different answers to the same question and manipulation by external content. The defence is structural: a limited source, a mandatory origin, automatic checks and human review.

In every case, NIST's AI RMF notes that accuracy measures should consider false positive and false negative rates and be taken on test sets representative of real conditions of use. A demonstration on sample data measures none of this.

When a simple rule is the better choice

AI is not the "advanced" option and the rule the "basic" one. For many risk tasks, the rule is simply the right tool. It tends to win when:

  1. The criterion is explicit and binary. A deadline has passed, a piece of evidence exists, MFA is enabled, a critical supplier has no assessment. There is nothing to estimate.
  2. The result must be reproduced identically. If an auditor or regulator asks why something was classified a certain way, "because condition X is true" is a stronger answer than a probability.
  3. There is no history to learn from. An ML model without enough labelled examples only reproduces noise. Start with rules and accumulate human decisions; they become the training history later.
  4. The cost of error is known and asymmetric. If missing a case is unacceptable, a conservative rule, even a noisy one, is usually easier to defend than a model that sometimes fails to flag.
  5. The volume is low. If the team can review everything, AI saves little and adds another system to govern.

A quick test: if the rule fits in one sentence and almost never has exceptions, write the rule.

Combining them: the rule as guard rail, AI as suggestion

Designs that work rarely pick just one side. Three patterns come up often:

  • Rule first, AI second. Fixed triggers handle the obvious cases (KEV, overdue deadline, exposed critical asset) and AI orders or classifies what is left.
  • AI proposes, approval becomes a rule. A control mapping or classification suggested by AI, once approved, becomes data in the database and is reapplied deterministically.
  • The rule watches the AI. Automatic checks confirm that what the model cited exists: the requirement, the figure, the source record.

In all of them, the decision to accept a risk, approve an exception or take an issue to the board stays with a person. In Futture's Eagle, AI is used for assisted classification, identifying relationships, contextual prioritisation and spotting inconsistencies, and the user reviews, approves, rejects or adjusts every relevant result. The platform was designed so that each recommendation keeps its context, data source, rationale, confidence level and decision history, which are precisely the elements needed to manage the failure modes described above.

Next step

List the ten recurring tasks that take up most of your risk or compliance team's time and put each one in one of three columns: rule, ML or LLM. For each task that lands in an AI column, decide before any pilot who reviews the output, which metric shows the model has got worse (confirmed alerts, rejected suggestions, errors found in review) and which simple rule acts as a safeguard. The order of implementation is covered in AI compliance automation: where to start.

Further reading: AI-native platform: what it means for GRC, Continuous internal audit: how to automate it and AI Governance Software: What It Is and How to Choose It.

Frequently asked questions

What is the difference between rule-based automation and AI in risk management?

A rule executes a condition someone wrote (if X, then Y) and always gives the same result for the same input. Machine learning models learn patterns from historical data and return an estimate, and LLMs work with text: they read, summarise, classify and draft. Rules are predictable and easy to audit; AI copes better with what nobody anticipated, but it fails in less obvious ways.

What is AI hallucination and how can you reduce the risk in risk reports?

It is what NIST calls confabulation: the model produces erroneous or false content stated with confidence, including invented justifications and citations. To reduce the risk, restrict the model to records in your own database, require every statement to point to its source record, check figures and identifiers automatically and keep human review before any text leaves the team.

When is a simple rule better than AI?

When the criterion is explicit and binary (a deadline has passed, a piece of evidence exists, MFA is enabled), when the result must be reproduced identically for an auditor or regulator, when there is no labelled history to train a model, or when the volume is low enough for manual review. If the rule fits in one sentence and almost never has exceptions, write the rule.

What is model drift and why does it matter in risk?

Drift is the loss of fit between the model and reality when the data or the phenomenon itself changes: a cloud migration, a reorganisation or a new attack pattern. The NIST AI RMF notes that AI systems may require more frequent maintenance because of data, model or concept drift. In risk work the effect is silent: the model keeps answering, only worse.

Official references

About the author

Founder, Futture.ai | CISSP

Cybersecurity, governance and risk management specialist with over 25 years of industry experience. CISSP-certified, founder of Network Secure and Futture.ai.

Read next

← Back to the blog