AI Risk Classification: How to Classify AI Use Cases by Risk
Not every use of artificial intelligence creates the same level of risk.
An employee using AI to summarise a non-sensitive internal document is not equivalent to an AI system influencing recruitment decisions.
An AI assistant drafting a marketing headline is not equivalent to an AI agent authorised to issue customer refunds.
And an internal productivity tool should not necessarily require the same approval process as AI affecting financial, regulatory or safety critical decisions.
Yet organisations can easily fall into one of two extremes.
They either apply heavy governance to almost everything, slowing responsible AI adoption.
Or they allow AI use to expand without distinguishing between relatively harmless experimentation and applications capable of creating significant consequences.
Neither approach scales particularly well.
The alternative is risk based AI governance.
That begins with a deceptively simple question:
How risky is this particular AI use case?
Answering that consistently requires an AI risk classification framework.
What is AI risk classification?
AI risk classification is the process of assessing an AI use case according to the potential consequences of its use and assigning a level of governance proportionate to that risk.
The objective is not simply to label AI:
Low Risk
Medium Risk
High Risk
and move on.
Classification should determine what happens next.
For example:
Does the use case require approval?
Who needs to approve it?
What documentation is required?
Is human oversight necessary?
What data can the AI access?
How frequently should performance be reviewed?
What monitoring is required?
Who can stop the system?
In other words:
Classification should drive control.
That is what makes a risk model operational rather than administrative.
For the wider principles, read What Is AI Governance? A Practical Guide for Business Leaders
Why classify AI use cases rather than AI tools?
This distinction is important.
The same AI technology can create very different risks depending on how it is used.
Imagine the same generative AI model being used for two tasks.
Use case A
Summarising an internal brainstorming session.
Use case B
Recommending which employees should be selected for redundancy.
Same underlying technology.
Very different consequences.
If governance focuses only on:
Which AI platform are we using?
it can miss the more important question:
What is the AI doing, who is affected, and what happens if it gets it wrong?
That means effective AI governance increasingly needs to operate at the use case level.
The problem with treating all AI as high risk
When organisations first introduce AI governance, caution is understandable.
But if every use case requires:
Security review.
Legal review.
Risk approval.
Executive approval.
AI committee approval.
Extensive documentation.
Then low-risk AI becomes unnecessarily difficult to use.
Employees become frustrated.
Business teams bypass formal processes.
Governance teams become overwhelmed.
And shadow AI can increase.
Strong governance should therefore allow organisations to say:
This is sufficiently low risk to proceed within predefined boundaries.
That is just as important as identifying applications requiring greater scrutiny.
Good AI governance should enable faster low-risk adoption and stronger high-risk control.
The problem with treating all AI as low risk
The opposite extreme is equally problematic.
An organisation approves a generative AI platform.
Employees then assume:
“It's approved, so we can use it for anything.”
But approval of the technology does not automatically mean approval of every use case.
The same tool might be appropriate for:
internal drafting
but inappropriate without additional controls for:
customer decisions
employee decisions
sensitive information
regulated activity
autonomous actions.
This creates a useful governance principle:
Tool approval and use-case approval are different decisions.
The Stratify AI Risk Classification Model
For each material AI use case, I would assess six dimensions:
IMPACT
DATA
DECISION
AUTONOMY
REVERSIBILITY
EXPOSURE
Together, these help leadership determine the level of governance required.
1. IMPACT: Who or what could be affected?
Start with consequence.
Ask:
Who could be affected if the AI performs incorrectly?
Possible impacts include:
Employees
Recruitment.
Performance.
Promotion.
Work allocation.
Disciplinary processes.
Workforce decisions.
Customers
Pricing.
Service.
Eligibility.
Recommendations.
Complaints.
Financial outcomes.
The organisation
Financial loss.
Operational disruption.
Legal exposure.
Reputational damage.
Wider stakeholders
Suppliers.
Partners.
Regulators.
Members of the public.
The greater the potential consequence, the stronger the governance case.
2. DATA: What information can the AI access?
Data can materially change the risk profile.
Ask whether the AI uses:
public information
internal information
commercially confidential information
customer data
personal data
employee information
financial information
special-category or highly sensitive information.
Also ask:
Where is the data processed?
Is it retained?
Can the AI provider use it?
Can information move outside organisational boundaries?
Who can access the outputs?
A relatively simple AI task can become higher risk if it involves highly sensitive data.
3. DECISION: What role does AI play?
AI can interact with decisions in several ways.
INFORM
AI provides information.
RECOMMEND
AI suggests a decision.
INFLUENCE
AI materially shapes the human decision.
DECIDE
AI determines the outcome.
The closer AI moves towards determining consequential outcomes, the stronger the need for clear accountability and oversight.
This is particularly important because organisations can describe a system as merely “supporting” a decision even when employees rarely challenge its recommendation.
The formal process and the real process may be different.
4. AUTONOMY: Can the AI take action?
This becomes increasingly important with AI agents.
Ask:
Can the AI only generate information?
Can it prepare an action?
Can it communicate externally?
Can it modify business records?
Can it initiate transactions?
Can it use other systems?
Can it trigger additional agents?
Can it take actions without prior human approval?
As autonomy increases, potential consequences can increase with it.
That does not automatically make autonomous AI inappropriate.
It means the boundaries need to become clearer.
For more, read Is Your Organisation Ready for Agentic AI?
5. REVERSIBILITY: Can we undo the outcome?
This dimension is often overlooked.
Suppose an AI makes an incorrect decision.
How easy is it to correct?
An AI-generated internal summary can simply be rewritten.
An incorrect customer email may require an apology.
An incorrect financial transaction may be recoverable but expensive.
An unfair employment decision may create much more serious consequences.
So ask:
If this goes wrong, how easily can we put things back?
Low reversibility generally increases the need for stronger controls before the action occurs.
6. EXPOSURE: How widely can failure spread?
Finally, consider scale.
An AI system used occasionally by one employee creates a different exposure from a system making 100,000 decisions every day.
Ask:
How many people could be affected?
How frequently does the AI operate?
How quickly could an error propagate?
Would failure remain local or become systemic?
AI can create enormous efficiency partly because it operates at scale.
That same characteristic can also scale mistakes.
The Stratify AI Risk Matrix
Once those six dimensions have been assessed, the use case can be placed into a practical governance tier.
TIER 1 — LOW RISK
Typical characteristics:
Limited consequence.
Non-sensitive data.
Internal use.
AI assists rather than decides.
Outputs easily reversible.
Limited scale.
Examples might include:
Internal brainstorming.
Formatting.
Drafting non-sensitive internal material.
Summarising low-risk information.
Governance approach
Approved tools.
Basic acceptable-use rules.
Employee training.
Limited formal approval.
Normal monitoring.
The objective is to allow these uses to move quickly.
TIER 2 — MODERATE RISK
Typical characteristics:
Some customer or operational impact.
Internal or moderately sensitive information.
AI recommendations influence work.
Errors create manageable consequences.
Human review remains practical.
Examples might include:
Draft customer communications.
Internal analysis.
Operational recommendations.
Knowledge assistants accessing internal information.
Governance approach
Named business owner.
Use-case registration.
Defined data requirements.
Appropriate human review.
Periodic performance checks.
Relevant specialist review where necessary.
TIER 3 — HIGH RISK
Typical characteristics:
Significant impact on customers, employees or finances.
Sensitive data.
AI materially influences important decisions.
Errors could create substantial harm or regulatory exposure.
Outcomes may be difficult to reverse.
Examples might include AI influencing:
Recruitment.
Employee decisions.
Credit.
Pricing.
Material financial decisions.
Regulated processes.
Governance approach
Formal approval.
Named accountable executive/business owner.
Documented risk assessment.
Strong human oversight.
Testing and validation.
Ongoing monitoring.
Clear incident escalation.
Regular review.
TIER 4 — CRITICAL / AUTONOMOUS
Some organisations may benefit from a fourth tier.
This is particularly relevant as AI agents gain greater authority.
Characteristics might include:
High-consequence decisions.
Significant autonomous action.
Access to sensitive enterprise systems.
Ability to initiate material transactions.
Potential for systemic impact.
Safety-critical or highly regulated activity.
Governance approach
Executive-level approval.
Strict permissions.
Explicit decision rights.
Defined autonomy limits.
Continuous or enhanced monitoring.
Strong auditability.
Human escalation.
Stop authority.
Formal incident procedures.
Regular assurance.
The terminology can vary by organisation.
The important point is that governance intensity increases with consequence.
The Stratify Governance Escalator
This gives us a core model for the future AI Governance Machine.
LOW RISK
→ Enable
MODERATE RISK
→ Register & Review
HIGH RISK
→ Assess & Approve
CRITICAL
→ Control & Assure
Or more simply:
ENABLE → REVIEW → APPROVE → ASSURE
This is the relationship between classification and governance.
A risk score without a governance response has limited value.
Risk classification should determine human oversight
Once a use case is classified, the organisation can determine the appropriate human role.
For example:
Low risk
Human uses the AI normally within policy.
Moderate risk
Human reviews important outputs.
High risk
Human approval may be mandatory before consequential decisions or actions.
Critical
Human supervision, defined escalation and strict autonomy limits may be required.
The appropriate model depends on the specific use case.
For a deeper framework, read Human Oversight in AI: When Does AI Need a Human in the Loop?
Risk classification should determine decision rights
Classification should also influence who can approve the use case.
For example:
Tier 1
Business function can proceed using approved tools within policy.
Tier 2
Business owner plus relevant specialist approval.
Tier 3
Formal cross-functional governance approval.
Tier 4
Executive approval and potentially Board visibility depending on materiality.
That creates proportionate escalation.
Low-risk AI does not continually travel upwards.
High-consequence AI cannot quietly appear inside a department.
This connects directly to AI Decision Rights: Who Should Be Allowed to Approve, Override or Stop AI?
Risk classification should determine monitoring
Monitoring should also increase with risk.
A Tier 1 tool might require little more than normal technology and policy monitoring.
A Tier 3 system may require:
Performance metrics.
Error rates.
Override rates.
Complaints.
Bias/fairness checks where relevant.
Exceptions.
Incidents.
Changes in model behaviour.
A Tier 4 agent may additionally require detailed audit trails of:
What it accessed.
What it decided.
What it changed.
Which actions it took.
Which systems it used.
When humans intervened.
The principle remains the same:
The greater the consequence, the greater the evidence leadership should require that the system remains under control.
Who should classify an AI use case?
Classification should not necessarily sit with one central AI team.
The business owner usually understands the use case.
Technology understands the system.
Security understands technical exposure.
Data/Privacy understands information risk.
Risk, Legal or Compliance may understand relevant regulatory consequences.
The classification process should therefore bring together the expertise required without turning every use case into a committee meeting.
For lower-risk uses, a structured self-assessment may be sufficient.
Higher-risk applications can trigger specialist review automatically.
This creates another useful principle:
Governance should escalate expertise when the risk requires it.
When should an AI use case be reclassified?
Risk classification should not be permanent.
A use case may change.
For example:
AI initially drafts emails.
Later it begins sending them.
The risk has changed.
An internal assistant later gains access to customer data.
The risk has changed.
An agent initially operates on 100 transactions.
Later it handles 100,000.
The exposure has changed.
A new regulation affects the process.
The context has changed.
Reclassification triggers might therefore include:
new data
new users
new decisions
greater autonomy
greater scale
new integrations
significant incidents
material model changes
regulatory change
This makes classification a lifecycle control rather than a one-time approval exercise.
The AI Risk Classification Test
Before classifying a material AI use case, leadership should answer:
Who could be affected?
What is the potential consequence if the AI is wrong?
What data does it access?
Does AI inform, recommend, influence or decide?
Can it take actions?
How autonomous is it?
Can an incorrect outcome be reversed?
How widely could failure spread?
What human oversight is practical?
What level of approval, monitoring and assurance is proportionate?
Those ten questions provide a much stronger foundation than simply asking:
“Is this AI high risk?”
Risk is not a reason to stop using AI
This is important.
The purpose of classification is not to create a growing list of things the organisation cannot do.
It is to determine how the organisation can do them appropriately.
A high-value AI opportunity may also be high risk.
That doesn't automatically mean:
Don't proceed.
It may mean:
Proceed with stronger controls.
Likewise, a low-risk opportunity should not spend three months moving through an approval process designed for consequential AI.
The objective is balance:
Value
and
Control.
Good AI governance enables both.
Frequently Asked Questions About AI Risk Classification
What is AI risk classification?
AI risk classification is the process of assessing an AI use case according to its potential impact, data, decision-making role, autonomy, reversibility and exposure, then assigning governance proportionate to that risk.
How do you classify AI risk?
Assess the specific use case rather than the AI tool alone. Consider who may be affected, what data is involved, what decisions AI influences, what actions it can take, how easily mistakes can be reversed and how widely failure could spread.
What is a high-risk AI use case?
A high-risk AI use case is one where errors or inappropriate use could create significant consequences for individuals, customers, employees, finances, regulatory obligations, safety or the organisation.
Are all generative AI tools high risk?
No. Risk depends primarily on how the technology is used. The same generative AI model could support a low-risk internal drafting task or a much higher-risk consequential decision.
Should every AI use case require approval?
Not necessarily. Lower-risk use can often operate within predefined policies and approved tools, while higher-risk or more consequential applications should trigger stronger assessment and approval.
How does AI autonomy affect risk?
Greater autonomy can increase risk because AI may move from generating information to making decisions or taking actions. Governance should therefore consider permissions, limits, human oversight, monitoring and stop authority.
How often should AI risk be reviewed?
Review should reflect risk and change. Material changes in data, autonomy, scale, integrations, model behaviour, regulation or use should trigger reconsideration of the classification.
From AI risk registers to AI decision intelligence
There is also a bigger opportunity here.
Traditional risk management often produces registers.
AI governance needs to produce decisions.
Leadership should eventually be able to see:
Which AI use cases exist?
What risk tier are they?
Who owns them?
What controls apply?
Which require human oversight?
Which have exceptions?
Which need review?
Which have increased in autonomy?
Which should be escalated?
That moves AI risk management from static documentation towards decision intelligence.
And that is exactly where governance becomes useful to leadership rather than merely compliant.
Because the objective isn't to create a perfect AI risk register.
It is to make sure the organisation knows:
Which AI needs freedom, which AI needs scrutiny, and which AI needs leadership attention.
How Strong Is Your AI Risk Governance?
Do you know which AI use cases create the greatest risk across your organisation?
And do your strongest controls actually sit around those applications?
Stratify's Executive AI Governance & Risk Assessment helps leadership teams examine whether AI governance is proportionate, visible and operational.
It assesses areas including:
AI governance foundations
Risk management
Human oversight and decision rights
Data, security and access
Monitoring
Incident response
Test Your AI Governance
Because effective AI governance doesn't mean controlling everything equally.
It means applying the right level of control to the things that matter most.
Comments