How much control should you give AI in education?

AI can compare two policies and flag inconsistencies. It can draft communications, organise administrative work and help staff find information. At the other end of the spectrum, it can become a virtual adviser that interacts directly with students and provides personalised support.

All of these fall under AI in education, but they should not be treated as carrying the same level of responsibility or risk. Effective AI governance in education starts by recognising a simple principle: the more responsibility you give AI, the stronger the controls around it need to become.

For education leaders, the question is no longer only, "Can AI do this?"

It is also: "How much control are we comfortable giving it?"

A tool that summarises internal meeting notes is fundamentally different from a system giving guidance directly to a student.

The first creates an output that a member of staff can review before anything happens. The second can influence someone's experience before a person has had an opportunity to intervene.

Yet approaches to AI governance for schools can fall into two extremes.

On one side, AI experimentation happens independently across teams with limited consistency around data, permissions or accountability. On the other, concern about risk leads to restrictions that prevent even relatively low-risk applications from being explored.

Neither approach reflects how varied AI use cases actually are.

A more practical model is risk-proportionate governance:

More autonomy + greater potential consequence = stronger governance

That does not mean low-autonomy AI needs no controls. It means the level of oversight should reflect what the AI can access, what it is allowed to do and what happens if it gets something wrong.

A useful way to make that distinction is to think about AI adoption as an autonomy ladder.

As AI moves up each level, it goes from supporting a person to taking increasingly independent actions. Human oversight of AI should evolve with it.

Level 1: AI assists

At the first level, AI helps someone complete a task but has no authority over the outcome.

It could:

• Draft a routine parent communication
• Summarise meeting notes
• Reformat existing teaching materials
• Organise information for staff

The control model is straightforward:

AI creates → Human reviews → Human acts

If the draft is inaccurate, inappropriate or simply not useful, a person can correct it before anything happens.

This makes assistive applications a useful starting point for organisations beginning to explore AI. Our guide to 7 education AI quick wins leaders can use without losing control explores several of these bounded opportunities in more detail.

Level 2: AI recommends

At Level 2, AI begins influencing a decision.

It might compare two policies and flag inconsistencies, identify an unusual pattern in operational information or suggest which administrative processes deserve further investigation.

The model becomes:

AI analyses → AI recommends → Human decides

The human remains accountable for the outcome, but AI oversight in education becomes more important because the system's output can shape what that person thinks or does next.

Leaders should therefore understand what information informed the recommendation and ensure staff know that an AI suggestion is evidence to consider, not automatically the correct answer.

Level 3: AI acts within defined rules

At Level 3, AI moves from recommending an action to completing one.

For example, it could categorise incoming enquiries and route them to the relevant department, organise documents according to predefined rules or trigger a low-risk administrative workflow.

The relationship changes:

Human defines boundaries → AI acts → Human monitors

Nobody needs to approve every individual action. Instead, people decide what the AI is authorised to do, establish the boundaries and monitor whether it continues operating as intended.

Permissions, exception handling and monitoring therefore become much more important.

Level 4: AI interacts

Now the AI communicates directly with students, parents or staff.

An institutional knowledge assistant might answer staff questions. A student support assistant could respond to routine enquiries. An admissions assistant could explain processes and direct applicants towards relevant information.

The major difference is that an incorrect output can now reach someone before a member of staff reviews it.

Responsible AI in education at this level requires clear answers to questions such as:

• Which information can the system access?
• Which questions is it authorised to answer?
• What topics are outside its boundaries?
• When should it escalate to a person?
• How are interactions monitored?
• Can users clearly understand when they are interacting with AI?

Direct interaction makes the quality of the system's guardrails as important as the quality of its answers.

Level 5: AI operates with greater autonomy

At the highest level, AI can potentially manage more complex, multi-step activities.

A sophisticated virtual adviser might understand a student's question, retrieve information from different approved sources, maintain context across an interaction and guide the student through several steps.

The model now looks more like:

AI understands → AI recommends/reasons → AI acts within boundaries → Human oversees exceptions

This is where AI risk management in education becomes particularly important.

The question is no longer whether the technology can produce a convincing answer.

It is:

What happens if it confidently produces the wrong one?

Moving higher on the ladder does not mean removing people. It changes where people sit within the process.

There is an important caveat.

Autonomy alone does not determine risk.

A Level 2 tool recommending something based on highly sensitive student information may need stronger controls than a Level 3 system automatically organising non-sensitive internal documents.

Education leaders therefore need to assess three things together:

Autonomy + Data + Consequence

How independently can the AI act? What information does it need? What happens to a person or the institution if the output is wrong?

Those three questions provide a much stronger foundation for AI governance in education than simply classifying every AI tool as high or low risk.

Before allowing an AI system to move further up the autonomy ladder, leaders should test whether the controls around it are moving too.

1. What information can it access?

Start by identifying exactly what the system needs.

There is a substantial difference between an AI using public information, approved internal documentation and sensitive student or staff information.

Access should follow the use case rather than giving a system broad access simply because it is technically convenient.

2. Who is affected if it gets something wrong?

Consider the consequence of an error.

An inaccurate first draft of an internal email can be corrected during review. Incorrect advice affecting a student's education, wellbeing or access to support could have much greater consequences.

The higher the impact, the stronger the case for human intervention.

3. When can a human intervene?

Human oversight can happen at different points:

Before action → During action → After action

At Level 1, someone may approve every output.

At higher levels, that becomes less practical. The focus shifts towards monitoring, escalation and intervention when the system encounters something outside its boundaries.

4. Does the AI know when not to answer?

A well-designed education AI system should not need to answer every question.

Sometimes the correct response is:

"This needs a person."

Leaders should define which topics require escalation and ensure there is a clear route to the appropriate member of staff.

5. Can you see what the AI is doing?

Greater autonomy requires greater observability.

Depending on the use case, this could involve reviewing interactions, maintaining appropriate logs, monitoring error patterns, gathering user feedback and checking whether the AI continues operating within its intended scope.

You cannot effectively govern a system you cannot evaluate.

Moving up the autonomy ladder should never become an objective in itself.

More autonomy does not automatically mean greater AI maturity.

Some decisions carry consequences that make human accountability essential, particularly around:

• Safeguarding
• Significant academic judgements
• Disciplinary decisions
• Sensitive wellbeing situations
• High-impact student decisions
• Exceptional circumstances requiring context and empathy

The role of AI in these situations may still be to retrieve information, identify patterns or support the person making the decision.

But supporting judgement and replacing judgement are different things.

Our broader guide to how AI can be used in education explores the wider applications, privacy considerations and risks education leaders should consider when developing their AI approach.

The goal is therefore not maximum autonomy.

It is the right level of autonomy for the consequence involved.

Consider two potential applications.

The first compares approved institutional policies and flags inconsistencies for a member of staff.

The second is a virtual student adviser capable of answering questions and providing personalised guidance.

Both may use advanced AI technology. Their governance requirements are very different.

The difference is not simply technical sophistication.

It is the AI's role in the outcome.

As that role becomes more direct, education organisations need stronger boundaries around access, accuracy, escalation and accountability.

Advanced AI autonomy in education does not need to mean giving technology unrestricted control.

The London School of Innovation provides a useful example of how deeply AI can be incorporated into an education platform. Geeks developed an AI-native learning environment containing more than 150 integrated AI agents, including a 24/7 personalised virtual tutor.

That level of capability is very different from using AI to draft an email or summarise a document.

It demonstrates why education leaders need to think beyond whether an AI feature works technically. As AI becomes embedded more deeply into the student experience, decisions around boundaries, access, oversight and human responsibility become part of the product itself.

Education organisations do not need to begin with a virtual adviser or autonomous agent.

A bounded Level 1 or Level 2 application can provide something equally important: evidence.

Start with a defined problem. Establish the boundaries. Measure what changes. Understand where the AI struggles. Then decide whether giving it additional responsibility would create enough value to justify the additional risk.

The progression should look more like:

Test → Measure → Learn → Increase autonomy where justified

This is also why Geeks' AI Adoption Framework starts with data, goals and measurable outcomes, then considers regulation, risk, human oversight and stakeholder engagement as adoption progresses.

The appropriate destination may be Level 2 for one process and Level 5 for another.

There is no requirement to reach the top.

AI systems will continue becoming more capable.

Education leaders will increasingly encounter tools that can not only create and analyse information, but communicate, recommend and act.

Capability, however, is not the same as permission.

The leadership conversation needs to progress through four questions:

Can AI do it?
↓
Should AI do it?
↓
How much autonomy should it have?
↓
Where must a person remain responsible?

That is the foundation of effective AI governance in education.

Responsible adoption is not about requiring a human to approve every click. It is about keeping people accountable where the consequences matter, while giving AI enough freedom to create genuine value.

For education organisations still determining where those boundaries should sit, Geeks' AI Opportunity Discovery approach helps identify and prioritise AI opportunities against the organisation's actual processes, data, goals and constraints before deciding what should be built.