Series Overview
Designing with Friction: AI, Learning, and K–12 Leadership
Generative AI is changing more than the tools students and educators use. It is changing the conditions under which work is produced, learning is demonstrated, data is remembered, and institutional decisions are made. This five-part series examines what K–12 leaders need to preserve, redesign, and govern as AI becomes embedded in classrooms, platforms, procurement, policy, and school operations.
The central argument is that schools should not respond to AI only through bans, detection tools, or broad acceptable-use language. They need a clearer framework for protecting the human work of learning: thinking, explaining, revising, questioning, judging, and taking responsibility for ideas. That requires distinguishing productive friction, which supports learning, from exclusionary friction, which blocks access without educational value.
In this series:
Part 1: The Illusion of Completion examines why polished student work may no longer be reliable evidence of learning.
Part 2: Writing Policy for the Learning Sequence argues that AI policy should address when AI enters the learning process, not only whether a tool is allowed.
Part 3: From Assistant to Actor explores why agentic AI creates new procurement and governance challenges for school districts.
Part 4: The Compliance Ghost in the Machine focuses on persistent AI memory, student data privacy, and the risks of invisible personalization.
Part 5: Tiered Decision Rights offers a framework for deciding what AI may generate, recommend, automate, and what must remain human.
You are reading Part 5: This post closes the series with a decision-rights framework for governing autonomous AI actions in school settings.
Tiered Decision Rights: Governing the Actions of Autonomous Systems
In the first, an AI tool drafts an email to a parent about missing assignments. The message appears in the teacher’s inbox. The teacher reviews it, adjusts the tone, adds context, and decides whether to send it.
In the second, an agentic system monitors the student information system, notices a pattern of missing work, generates a message, selects the parent contact, and sends the email automatically.
Both scenarios involve AI. Both may save time. Both may use similar language on a vendor website.
But they are not the same governance problem.
The first scenario supports human judgment. The second scenario replaces it at the point of action.
As AI systems become more autonomous, school districts need tiered decision rights. A tiered decision-rights framework defines which actions an AI system may take independently, which actions require human approval, and which actions must remain outside AI authority altogether.
This is not only a technical control. It is an educational values statement.
Schools are human institutions. They serve children, families, and communities. Many school decisions depend on context that does not live neatly in a database. A family may be navigating housing instability. A student may be grieving. A missing assignment pattern may reflect confusion, avoidance, disability-related need, language barriers, disengagement, or a temporary crisis. A human educator may know that the “efficient” message is not the right message.
An AI system does not know that in the same way.
It may infer patterns. It may optimize workflows. It may generate fluent communication. It may even be useful. But it does not carry professional responsibility. It does not understand the relationship between a school and a family. It does not live with the consequences of a message that damages trust.
That is why autonomy must be governed.
A practical framework can divide AI actions into four levels.
Level 1: AI may generate, but not act.
At this level, AI produces drafts, summaries, templates, checklists, or suggestions. A human must review and decide what happens next. This is appropriate for many low-risk productivity tasks, such as drafting a newsletter paragraph, summarizing meeting notes, creating a first version of a rubric, or suggesting language for a lesson plan.
Level 2: AI may recommend, but human approval is required.
At this level, AI identifies possible actions but cannot complete them. For example, a system may flag students who might need follow-up, recommend a family contact, suggest intervention resources, or identify missing data. A staff member must review the recommendation, apply professional judgment, and approve any action.
Level 3: AI may automate low-stakes actions within defined limits.
Some actions may be safe to automate if they are narrow, reversible, transparent, and auditable. Examples might include sending a reminder to a staff member, organizing non-sensitive records, scheduling a meeting based on already-approved availability, or routing a help ticket to the right department. Even here, the district should require logs, opt-outs, role-based permissions, and clear escalation pathways.
Level 4: AI may not act.
Some decisions should remain human. These include actions related to discipline, grading, special education eligibility or services, student placement, threat assessment, mandated reporting, formal intervention assignment, high-stakes family communication, or any decision that meaningfully affects a student’s rights, record, access, or educational trajectory.
The line between these levels should be written into policy, procurement, and implementation plans. It should not be left to vendor defaults.
This matters because automation bias is real in practice even when people retain formal authority. If a system recommends an action with confidence, staff may accept it because it appears efficient, objective, or data-driven. Over time, the human review step can become symbolic unless districts intentionally protect it.
Human-in-the-loop is not enough if the human is rushed, unclear about their role, or unable to see how the recommendation was produced.
A stronger approach is meaningful human control. That means the human reviewer has enough information, authority, time, and institutional permission to disagree with the system.
District leaders should ask vendors direct questions.
Can the district disable autonomous actions?
Can different roles have different approval rights?
Can high-stakes categories be blocked entirely?
Are recommendations explainable enough for staff review?
Are all actions logged?
Can the district audit what the system did, when it did it, what data it used, and who approved it?
Can users appeal, correct, or reverse system-supported decisions?
If those controls are not available, the tool is not ready for high-trust school environments.
The central question is not whether AI can make schools more efficient. It can.
The better question is where efficiency belongs.
Efficiency is welcome when it reduces clerical burden, improves access, or gives educators more time for human work. Efficiency becomes dangerous when it quietly replaces professional judgment in decisions that require context, empathy, accountability, and care.
Autonomous systems will keep advancing. District policy needs to advance before those systems become embedded in everyday school operations.
The rule should be clear: the higher the stakes, the stronger the human authority.
Schools can delegate tasks.
They should not delegate responsibility.
