AI Monitoring and Assurance: How Do You Know Your AI Controls Are Actually Working?
An AI use case has been approved.
The risk assessment is complete.
Policies have been followed.
Human oversight has been designed.
Permissions have been established.
The system goes live.
Six months later, leadership asks:
Is it still operating the way we approved it?
That question is harder than it sounds.
The data may have changed.
Employees may be using the AI differently.
A vendor may have updated the technology.
An AI agent may have gained additional permissions.
Exception rates may have increased.
Human reviewers may be approving outputs without meaningful challenge.
A control that looked effective during implementation may no longer work as intended.
This exposes an important weakness in many AI governance models.
They concentrate heavily on approval.
But approval tells leadership what was considered acceptable at a particular point in time.
It does not prove that AI remains acceptable afterwards.
That requires monitoring and assurance.
What is AI monitoring?
AI monitoring is the ongoing observation of an AI system, use case or agent to determine whether it continues to operate within expected performance, risk and authority boundaries.
Depending on the use case, monitoring might examine:
performance;
quality;
errors;
exceptions;
human overrides;
data;
permissions;
agent actions;
business outcomes;
incidents;
cost;
usage.
Monitoring answers:
What is happening?
What is AI assurance?
AI assurance goes a step further.
It asks:
Can we demonstrate that the controls and governance surrounding AI are operating effectively?
Assurance might involve:
control testing;
evidence review;
audits;
ownership reviews;
permission reviews;
policy compliance;
risk reassessment;
independent challenge.
So:
MONITORING = What is happening?
ASSURANCE = Can we demonstrate that the system and controls are working as intended?
Both matter.
For more on the wider governance capabilities leadership should expect around enterprise AI, read AI Governance Checklist: 15 Questions Every Leadership Team Should Ask.
Approval is a point in time
Imagine an AI use case is approved in January.
At approval:
it has 100 users;
uses internal data;
cannot take external actions;
requires human approval;
operates in one country.
By September:
2,000 employees use it;
customer data has been added;
it sends communications;
some decisions are automated;
three countries use it.
The original approval may still exist.
But the governance proposition has changed.
This gives us a core principle:
AI approval should not be treated as permanent permission under permanently unchanged conditions.
Material change should trigger reassessment.
What should organisations monitor?
There is no universal monitoring dashboard.
Monitoring should reflect the characteristics and consequences of the use case.
A low-risk internal assistant may require relatively limited monitoring.
A high-impact customer-facing agent may require considerably more.
The objective is proportionate observability.
The Stratify AI Monitoring & Assurance Framework
Structure monitoring and assurance across seven dimensions:
PERFORMANCE → BEHAVIOUR → AUTHORITY → RISK → OUTCOMES → CONTROLS → EVIDENCE
Together, they answer:
Is the AI working, operating within its boundaries and producing the outcomes leadership believes it is producing?
1. PERFORMANCE — Is the AI still working sufficiently well?
Performance may include:
accuracy;
quality;
reliability;
response time;
failure rate;
completion rate;
technical availability.
The appropriate measures depend on the use case.
The important point is that organisations should define:
expected performance
and
acceptable tolerance.
Otherwise monitoring becomes observation without a basis for intervention.
The Performance Boundary
For material AI, define:
EXPECTED
Normal performance.
WARNING
Performance requires investigation.
BREACH
Performance requires intervention.
This turns monitoring into a management mechanism.
2. BEHAVIOUR — Is the AI behaving as expected?
Technical performance alone may not reveal emerging problems.
Leadership may also need visibility into:
exceptions;
unexpected outputs;
repeated human corrections;
unusual usage;
changes in response patterns;
agent actions.
An AI system may remain technically available while operational behaviour deteriorates.
For example, an agent could successfully complete 99% of transactions while increasingly choosing an inappropriate route.
The system is functioning.
The behaviour is changing.
Monitoring therefore needs to examine how AI behaves, not merely whether it is available.
3. AUTHORITY — Is AI staying within its approved boundaries?
This becomes increasingly important with AI agents.
An agent may have permission to:
read certain data;
update certain fields;
send certain communications;
transact below defined limits.
Monitoring should help identify:
permission breaches;
unexpected actions;
changes in access;
repeated escalation;
delegation to other systems;
attempts to operate outside approved boundaries.
For more on defining the authority an AI agent should receive, read AI Agent Permissions: What Should an AI Agent Be Allowed to Access and Do?
Permission drift
Permissions can accumulate over time.
An agent begins with:
read-only access.
Then gains:
update access.
Then:
communication authority.
Then:
transaction authority.
Each change may appear individually reasonable.
Collectively, the agent has become significantly more powerful.
This is permission drift.
Monitoring should therefore examine not only:
What permissions does the agent have?
but:
How have those permissions changed?
4. RISK — Has the risk profile changed?
Risk classification should not be a one-time activity.
Risk may change because:
usage increases;
new data is introduced;
new users gain access;
autonomy increases;
the AI becomes customer-facing;
business dependency increases;
regulatory conditions change;
incidents occur.
A material change should be capable of triggering:
reclassification;
additional controls;
new approval;
reduced authority;
human review.
For more on establishing a proportionate risk classification model, read AI Risk Classification: How to Classify AI Use Cases by Risk.
The Material Change Trigger
A governance review should potentially occur when there is significant change in:
PURPOSE
What is the AI being used for?
SCALE
How widely?
DATA
What information does it use?
AUTONOMY
What can it decide or do?
IMPACT
Who can it affect?
DEPENDENCY
How important has it become?
This prevents an AI use case quietly becoming materially different from the one originally approved.
5. OUTCOMES — Is the AI creating the expected business result?
Monitoring should not become exclusively technical.
Leadership needs to understand whether AI is creating the outcome that justified the investment.
For example:
AI reduces handling time.
But:
Does customer satisfaction fall?
AI generates more sales activity.
But:
Does conversion improve?
AI automates processing.
But:
Does error-related rework increase?
The organisation therefore needs to monitor both:
AI PERFORMANCE
and
BUSINESS PERFORMANCE.
For more on tracing AI improvements through to measurable business value, read AI Benefits Realisation: How to Turn AI Productivity Into Measurable Financial Value.
Good monitoring can reveal false success
Suppose an AI tool saves employees 20% of task time.
The technology dashboard looks excellent.
But employees spend the saved time fixing AI errors.
Or customer complaints increase.
Or managers do not redeploy the capacity.
The AI is producing activity metrics.
The business is not receiving equivalent value.
That is why monitoring needs to connect technology to outcomes.
6. CONTROLS — Are the safeguards actually operating?
This is where assurance becomes important.
An organisation may have documented controls such as:
human approval;
restricted data access;
transaction limits;
mandatory escalation;
periodic review.
But are they happening?
For example:
Control design
Human reviews every high-impact output.
Operational reality
Employees approve outputs automatically because volume is too high.
The control exists.
But it may no longer be effective.
Assurance therefore asks:
Is the control operating as designed — and is the design still sufficient?
The Stratify Control Assurance Test
For each material control, ask:
PURPOSE
What risk is this control intended to address?
OWNER
Who owns the control?
OPERATION
How should it work?
EVIDENCE
What proves it happened?
EXCEPTION
What happens when it fails?
TEST
How do we know it remains effective?
REVIEW
When should it be reconsidered?
This turns governance controls into testable mechanisms rather than policy statements.
Human oversight needs assurance too
Human oversight is often presented as a control:
A human reviews the AI.
But leadership should ask:
What does the human review?
Do they have enough information?
Do they have enough time?
Can they challenge the AI?
Do they understand their accountability?
How often do they override it?
What happens after an override?
A human being present does not prove meaningful oversight.
For more on designing human involvement that actually changes outcomes, read Human Oversight in AI: When Does AI Need a Human in the Loop?
7. EVIDENCE — Can leadership prove any of this?
This is the final layer.
If leadership says:
“Our AI is appropriately governed,”
what supports that statement?
Evidence might include:
monitoring records;
risk reviews;
permission reviews;
decision logs;
incident records;
control tests;
human overrides;
training records;
benefit validation;
ownership reviews.
This creates an important distinction:
GOVERNANCE CLAIM
versus
GOVERNANCE EVIDENCE.
The stronger the potential consequence of the AI, the stronger the evidence leadership may reasonably require.
The Stratify AI Assurance Ladder
Governance confidence can develop through five stages:
DEFINED → IMPLEMENTED → OPERATING → EVIDENCED → ASSURED
DEFINED
The control exists on paper.
IMPLEMENTED
The control has been put into operation.
OPERATING
There is evidence it is actually being used.
EVIDENCED
Records demonstrate its operation and outcomes.
ASSURED
The organisation has tested whether the control remains effective.
This distinction is important.
A policy can be Defined without ever reaching Assured.
Monitoring should trigger action
Dashboards alone do not create governance.
Monitoring becomes useful when information leads to decisions.
For example:
Performance deteriorates
→ investigate.
Exception rate increases
→ review workflow.
Permission drift occurs
→ reassess authority.
Risk changes
→ reclassify.
Control fails
→ remediate.
Incident occurs
→ respond.
Value deteriorates
→ review investment.
This gives us another cycle:
MONITOR → DETECT → ASSESS → ACT → EVIDENCE → REVIEW
Monitoring and incident management should connect
Monitoring should help identify when normal variation becomes an incident.
That requires thresholds and escalation.
For example:
one unusual output
may be an exception.
Repeated harmful outputs
may become an incident.
A major unauthorised action
may immediately become critical.
For more on what should happen when AI crosses from normal operation into a material incident, read AI Incident Management: What Should Happen When AI Goes Wrong?
The AI Monitoring Record
PERFORMANCE METRICS
What should be measured?
BUSINESS METRICS
Which outcome matters?
RISK INDICATORS
What could signal increased exposure?
CONTROL INDICATORS
What demonstrates safeguards are operating?
AUTHORITY
What boundaries apply?
EXCEPTIONS
What unusual events occurred?
INCIDENTS
What material events occurred?
EVIDENCE
What supports governance claims?
REVIEW DATE
When is formal reassessment required?
OWNER
Who reviews the information?
Assurance should be proportionate
Not every AI use case requires formal audit.
That would be expensive and slow.
Assurance should reflect:
risk;
impact;
scale;
autonomy;
data;
regulatory context.
A low-risk productivity assistant may need:
light monitoring;
periodic ownership review;
basic policy compliance.
A high-impact autonomous agent may require:
continuous monitoring;
permission review;
action logs;
control testing;
formal assurance.
The principle is:
More consequential AI should require stronger evidence that governance is working.
Assurance should be independent where appropriate
For some material AI, the person responsible for operating the use case should not be the only person deciding whether controls are effective.
Independent challenge may come from:
Risk;
Internal Audit;
Security;
Compliance;
external assurance;
another governance function.
The appropriate model depends on the organisation and consequence.
The purpose is not bureaucracy.
It is avoiding:
“We checked our own system and concluded everything is fine.”
Assurance findings should become actions
Suppose assurance identifies:
human oversight is weak;
permission reviews are overdue;
monitoring thresholds are undefined;
risk classification is outdated.
Those findings should not disappear into an audit report.
They should become:
GAP
↓
ACTION
↓
OWNER
↓
PRIORITY
↓
DUE DATE
↓
EVIDENCE REQUIRED
↓
STATUS
↓
REVIEW
The AI Governance Evidence Register
For material AI, leadership could maintain an evidence view containing:
CONTROL
What should happen?
OWNER
Who owns it?
EVIDENCE
What proves it?
LAST TESTED
When?
RESULT
Effective / Partially Effective / Ineffective.
ACTION
What needs changing?
NEXT REVIEW
When?
Now governance becomes something leadership can demonstrate rather than merely describe.
Monitoring should connect to accountability
If a control fails, someone needs to act.
If performance deteriorates, someone needs to decide.
If value disappears, someone needs to explain why.
Monitoring without ownership creates information without response.
For more on connecting outcomes, authority, evidence and review into an operational accountability model, read AI Accountability in Practice: How to Build an Ownership Model That Actually Works.
The Fractional CAIO Monitoring & Assurance Review
This could become another recurring element of the Fractional CAIO cadence.
For material AI:
PERFORMANCE
Is it working?
BEHAVIOUR
Is anything changing?
AUTHORITY
Is it operating within boundaries?
RISK
Has exposure changed?
OUTCOME
Is value emerging?
CONTROL
Are safeguards working?
EVIDENCE
Can we prove it?
ACTION
What needs attention?
DECISION
Continue / Change / Restrict / Pause / Stop?
The Fractional CAIO does not replace:
Risk;
Security;
Internal Audit;
Compliance.
The role ensures evidence from those functions connects into the wider enterprise AI management system.
10 Questions to Test AI Monitoring and Assurance
Does every material AI use case have defined performance and business-outcome measures?
Can the organisation detect material changes in AI behaviour?
Are AI agents monitored against their approved permissions and authority?
Do material changes in purpose, scale, data or autonomy trigger governance reassessment?
Are risk indicators reviewed after AI enters production?
Can leadership demonstrate that documented AI controls are actually operating?
Is human oversight tested for effectiveness rather than assumed to work?
Does material AI have sufficient evidence to support governance claims?
Do monitoring and assurance findings become owned actions with deadlines and evidence requirements?
Can leadership identify which AI systems currently require intervention, reassessment or additional assurance?
If several answers are unclear, the organisation may have AI governance structures without sufficient evidence that those structures are working.
AI governance does not end at approval
Approval matters.
Policy matters.
Risk classification matters.
Human oversight matters.
But none of them should be assumed to remain effective forever.
AI changes.
Data changes.
People change.
Technology changes.
Permissions change.
Business dependency changes.
Governance therefore needs to move from:
“We approved it.”
to:
“We can demonstrate that it continues to operate within the boundaries and outcomes we expect.”
That is the difference between governance as documentation and governance as an operating capability.
Frequently Asked Questions About AI Monitoring and Assurance
What is AI monitoring?
AI monitoring is the ongoing observation of an AI system, use case or agent to determine whether performance, behaviour, risk and authority remain within expected boundaries.
What is AI assurance?
AI assurance is the process of evaluating whether AI governance, controls and oversight are appropriately designed and operating effectively.
What is the difference between AI monitoring and AI assurance?
Monitoring observes what is happening. Assurance evaluates whether the controls and governance surrounding that activity are working as intended.
What should organisations monitor in AI systems?
Depending on the use case, monitoring may include performance, quality, exceptions, human overrides, permissions, agent actions, risk indicators, business outcomes, incidents and operating cost.
How often should AI be reviewed?
Review frequency should reflect risk, impact, scale and autonomy. Higher-consequence AI may require continuous monitoring and more frequent formal reviews.
What is AI permission drift?
Permission drift occurs when an AI system or agent gradually accumulates greater data, system or action authority over time without the overall change receiving sufficient governance attention.
How do you know whether an AI control is working?
Define what the control should achieve, identify its owner, determine what evidence demonstrates operation and periodically test whether it remains effective.
Is human oversight an AI control?
Yes, but simply involving a human does not prove the control is effective. Organisations should evaluate whether the human has sufficient information, authority, capability and time to intervene meaningfully.
Should every AI system receive formal assurance?
Not necessarily. Assurance should be proportionate to the risk and consequence of the use case. Lower-risk applications may require lighter assurance than high-impact or highly autonomous AI.
What happens when AI monitoring identifies a problem?
The organisation should assess the issue and determine whether it requires investigation, remediation, escalation, risk reassessment, permission changes or a Continue / Change / Restrict / Pause / Stop decision.
How Strong Is Your AI Governance?
Effective AI governance needs to continue after implementation. Leadership needs confidence that AI remains within approved boundaries, controls operate as intended and material changes are identified.
Take our free 5-minute Executive AI Governance & Risk Assessment →
Discover where your organisation is strong, where governance gaps exist and what to prioritise next.
Comments