AI for support: what to delegate and what not to delegate

A replicable framework for using artificial intelligence inclusively, effectively, and responsibly.

Artificial intelligence is entering schools with a concrete promise: reducing workload, increasing personalization, and making materials and activities more accessible.

For those who work in support, however, the question is not "can I use it?", but "what makes sense to delegate and what not, without losing educational quality, relationships and protection of the children? "

This article proposes a practical framework, designed for everyday use: a delegation matrix , operating procedures, (reusable) prompt examples, and a risk box to manage bias, data, and dependency.

This is not a toolkit, but a decision-making guide for using AI without delegating what cannot be delegated.

1) A guiding principle: AI is not a substitute for instructional design

In support, the goal is not to produce materials "faster," but to lower barriers and increase participation, autonomy, and expertise . Particularly when the work is repetitive, structured, or requires many variations, AI can be an excellent "copilot" for:

• adapt the form (language, channel, structure);

• multiply variants (levels, parallel exercises);

• provide feedback consistent with predefined criteria.

Instead, it shouldn't become the "authoritative" source for educational objectives, assessments, or decisions. Those require context, knowledge of the child, and professional responsibility.

If AI speeds up work but weakens educational intentionality, the cost is too high.

2) The delegation matrix: what to delegate, what to co-pilot, what to keep human

I propose a simple rule: the more standardizable and reversible the activity, the more delegable it is . The more contextual, identity-based, or high-impact it is, the more human it must be kept... and the greater the teacher's professional responsibility.

A. Delegable (AI “does”, you control)

Low-risk and easily verifiable activities . Here, AI can "do" things autonomously, provided there's a final check.

1. Linguistic simplification

2. Rewrite texts in simpler language (short sentences, frequent vocabulary).

3. Create “easy to read” versions without changing key content.

4. Structured synthesis

5. Point summaries, text maps, glossaries, guiding questions.

6. Extract “key ideas”, “difficult words”, “sequence of events”.

7. Generation of parallel exercises

8. Exercises with the same target skill but different examples (generalization).

9. Gradual difficulty versions (levels 1–3).

10. Preparation of textual visual aids

11. Step-by-step procedures, checklists, simplified rubrics, study prompts.

    
    

Minimum mandatory checks: accuracy of content, appropriateness of language, absence of stereotypes.

B. Co-pilot (AI suggests, you decide and refine)

Educationally sensitive activities , where the value lies in your career choice. Here, AI accelerates analysis, but doesn't make decisions.

1. Formative Feedback AI can generate feedback, but only if you provide: objective, criteria, expected level, and examples of appropriate responses. You then validate the tone, priorities, and next steps.

2. Planning micro-goals and AI strategies is useful for proposing alternatives (compensation strategies, scaffolding, mediators), but the decision must be based on observation and IEP.

3. Differentiation and UDL AI can suggest representation/action/expression options, but you choose what is realistic for the class, resources, and functional profile.

4. Designing accessible assessments AI can generate alternative formats (multiple choice with audio, fill-in-the-blank, guided responses), but you must ensure consistency with the objective and evaluative validity.

   
   

Required monitoring: consistency with the IEP, cognitive load, accessibility, fairness of assessment.

C. Non-delegable (full human responsibility)

Here, AI can at most “help you think,” but it must not determine final decisions or content.

1. High-impact educational and evaluation decisions

2. definition of PEI objectives, results, certifications, assessments.

3. clinical interpretation or diagnosis.

4. Relationship and behavioral management

5. emotional dynamics, conflicts, communication with family and team.

6. Strategies for crisis, anxiety, and dysregulation: direct observation and contextual consistency are required.

7. Sensitive and identity-related content

8. health issues, trauma, sexuality, critical family situations.

9. do not use AI as a “consultant” for specific cases of this type.

10. Personal data and documents with identifying data

11. PEI names, reports, diagnoses, minutes: should not be pasted into unauthorized devices.

3) The 5-step replicable framework (operating procedure)

This is a routine you can apply whenever you want to use AI. You don't need to use them all the time: even applying just a few improves the quality of your choices.

Step 1 — Define the observable goal (not the tool)

Example: “understand the key concepts of a science text”Not: “make a summary with AI”.

Step 2 — Identify the primary barrier

• linguistics (lexicon, syntax)

• attentional/organizational (sequences, load)

• motor/written production

• abstract comprehension / working memory

• communication (AAC, multimodality)

Step 3 — Choose the type of delegation (A/B/C)

Delegable, co-driven or non-delegable.

Step 4 — Generate 2–3 alternatives (not just one)

The typical mistake is to take the first version. With 2–3 variations, choose the best one and reduce bias/oversimplification.

Step 5 — Check with a quick checklist (60 seconds)

• Is it accurate?

• Does it retain the essential concepts?

• Is it accessible (structure, length, vocabulary)?

• Does it respect dignity and high expectations?

• Avoid personal data?

• Does it promote autonomy (not dependence)?

If a response does not pass this checklist, it is not ready for classroom use.

4) Practical examples: simplification, synthesis, exercises, feedback

Here you'll find ready-to-use prompts to reduce barriers without lowering your goals. Replace the parts in square brackets.

Always use a clear source text and check the output before using it in class.

A) Simplification (three levels)

Prompt

Rewrite this text for a student in [grade/age], keeping all the key concepts. Create 3 versions:

• Level 1: very short sentences, maximum 12 words, everyday vocabulary.

• Level 2: Short sentences with 3–5 keywords in bold (keywords must match essential disciplinary concepts).

• Level 3: Original text but with short paragraphs and headings. Add a mini-glossary (max 8 words) and 3 comprehension questions.

Typical use in advocacy: offering options without “lowering” the goal, but by changing the form.

B) Structured summary for guided study

Prompt

Transform text into:

1. bulleted list of 6 key concepts

2. timeline (if any) in 5 steps

3. 8 question/answer flashcards

4. a text map with headings and subheadings. Stay true to the content and flag any ambiguities.

C) Generation of parallel exercises (generalization)

Prompt

Objective: [e.g. distinguish cause and effect]. Create 12 exercises on three levels (4 per level), with simple instructions.

• basic level: choice between 2 options

• Intermediate level: choice between 4 options

• Advanced level: guided response with “why” Also provide the solution and typical error for each level.

D) “Criteria-anchored” formative feedback

Here it is essential to provide criteria, otherwise the AI invents “generic” feedback.

Prompt

You are an academic tutor. Please provide feedback on this paper based on the following criteria:

• C1: correctness of content (0–2)

• C2: clarity and order (0–2)

• C3: use of disciplinary vocabulary (0–2)Return:

• score per criterion with short justification

• 2 things done well (specifications)

• 1 priority for improvement

• a “next step” with a short exercise (3 minutes)Tone: respectful, encouraging, without infantilizing.

RISK BOX (to be kept always visible)

1) Biases and stereotypes

Risk: AI may produce examples or feedback that reinforce stereotypes (disability = inability, gender roles, low expectations). Practical mitigation:

• explicitly ask “avoid stereotypes”

• check examples, names, roles, contexts

• Prefer neutral, inclusive example datasets—if an example makes you uncomfortable, that's a sign: rewrite it or discard it.

  
  

2) Data and privacy

Risk: pasting identifying information (name, diagnosis, IEP, specific incidents) or sensitive materials into chat. Practical mitigation:

• use anonymous data and functional descriptions (e.g. “student with reading difficulties”)

• Do not enter diagnoses or clinical details

• If you have to work on a real text, remove personal references and recognizable context

  
  

3) Dependence (for teacher and student)

Risk: AI becomes a shortcut: less planning, less student autonomy, more "off-the-shelf solutions." Practical mitigation:

• use AI to create scaffolding , not replacements

• favors outputs that guide processes (steps, questions, criteria)

• define a “fading plan”: gradually reduce media

  
  

4) Hallucinations and inaccuracies

Risk: AI can invent data, definitions, and references. Practical mitigation:

• check disciplinary concepts and definitions

• ask “if you are unsure, state so; and always verify definitions, dates, and disciplinary concepts.”

• Please provide the source (source text) when possible

  
  

5) Fairness in evaluation

Risk: Inconsistent or penalizing feedback/evaluations for those who use compensatory tools. Practical mitigation:

• separates “competence” from “form” when applicable

• currency with clear rubrics

• documents adjustments and criteria shared within the team

5) A final mini-checklist for the support teacher

Before using an AI output in the classroom, ask yourself:

1. What barrier am I reducing?

2. Does the student remain the protagonist?

3. Does the output increase range or replace it?

4. Is it consistent with the objective and IEP?

5. Have I avoided personal data?

6. Have I checked for bias and accuracy?

This checklist is not meant to limit the use of AI, but to make it intentional.

If you answer “no” to even just one question, it’s not a “ban”: it’s a signal that more co-design, more control, or a different type of support is needed.

Closing: How to Turn AI into an Inclusive Lever

AI becomes truly useful in support when it's treated as an infrastructure for accessibility and differentiation , not as a "task generator." Its greatest value lies in multiplying options and reducing barriers, while your professional value remains irreplaceable: observing, prioritizing, building motivation and relationships, ensuring dignity and fairness.

---

Support ForAllWe

ForAllWe is an independent project on digital accessibility and inclusive technology. If you find our content useful, you can support us with a voluntary contribution. Your support helps us remain independent and involve people with disabilities as authors and testers.