Design Principles and Observable Indicators for AI-Enabled Pedagogical Accompaniment: Evidence from the Amico Dual-Mode Prototype in Italy and China

Synthesis. Benedetti (2026) argues that the value of AI-enabled educational systems depends less on technological sophistication than on accountable pedagogical mediation: transparent limits, bounded interaction, and an explicit orientation toward human relationships. Adopting a "human-in-command" stance — rather than merely Human-in-the-Loop — the paper defines AI-enabled pedagogical accompaniment as a directional, temporary, and safeguarded mediation operationalized through the concept of a relational bridge: a sequence of micro-mediations that lowers the threshold of access to the educational relationship and facilitates transitions toward meaningful human contact. The Amico dual-mode prototype (AmicoMio for structured guidance, AmicoTuo for reflective support) maps five design principles to observable indicators, with exploratory Italy–China pilot evidence suggesting feasibility and perceived usefulness in vocational education.

Key Findings

1. Pedagogical value is a function of accountable mediation, not "human-likeness": participants recognized Amico as a bounded support tool rather than an autonomous tutor or relational substitute, and transparency mechanisms appeared to prevent anthropomorphic misinterpretation across contexts.
2. The dual-mode configuration is a design asset: AmicoMio was associated with clarity, structure, and task-oriented usefulness, while AmicoTuo was associated with reflective meaning-making and motivational support — perceived as complementary rather than competing resources.
3. Observable indicators (transition-to-human contact, micro-action completion, communicative autonomy) supported directional mediation, orienting learners toward action and human engagement rather than prolonged AI interaction.
4. The system operated within intended boundaries: sessions remained bounded in duration and frequency, dependency risk signals were rare and promptly redirected toward human contact, and no expectations of substitution or emotional dependency emerged.

Human-in-Command and Adult Responsibility

The framework rests on a pedagogical interpretation of AI as mediated educational infrastructure rather than an autonomous instructional or relational agent. It explicitly rejects technocentric and substitutionist models in which AI serves as an empathic surrogate or autonomous tutor, and instead positions AI as a temporary, directional mediation designed to lower barriers to participation. A central pillar is the principle of human-in-command, whereby adults — teachers, tutors, and institutions — retain responsibility for educational decisions, relational boundaries, and ethical oversight. This stance foregrounds AI Governance and Ethics in educational AI deployment and connects directly to debates on Teaching and Teacher AI Competency. For learners experiencing vulnerability around Self-Efficacy, Motivation, or belonging, embedding adult responsibility into system design helps ensure that AI acts as supportive mediation rather than an alternative relational pole.

The Relational Bridge as Educational Mediation

The conceptual core of the framework is the notion of a relational bridge: a sequence of micro-mediations that reduce the threshold of access to the educational relationship. It is characterized by three properties — temporality (interaction explicitly limited in scope and duration), directionality (interactions oriented toward concrete next steps with teachers, peers, or communities), and boundedness (behavior constrained by pedagogical purpose and ethical safeguards). This operationalization aligns closely with Scaffolding theory's emphasis on calibrating support to learner needs, while the bridge's exit points toward human actors connect to Sociocultural Learning and Collaborative Learning. The effectiveness of the bridge is assessed not by engagement with the AI itself but by observable transitions toward educational participation, Learner Agency, and responsibility.

The Amico Dual-Mode Prototype

Amico was conceived as an AI-mediated interaction device explicitly designed for non-clinical educational contexts, with particular attention to technical and vocational education settings. Its scope is deliberately limited: it provides no curricular instruction, psychological counseling, or affective companionship, and does not optimize for content mastery. Instead, it supports actionable next steps, lowers barriers to participation, and facilitates transitions toward human actors.

Its distinctive feature is a dual-mode configuration. AmicoMio orients toward technical clarity, structured guidance, step-by-step reasoning, and task-focused support. AmicoTuo orients toward reflective dialogue, maieutic questioning, reformulation, and meaning-making — supporting moments of uncertainty and motivational fatigue without adopting a therapeutic stance. The two modes are complementary configurations within a single pedagogical framework rather than competing alternatives. Age-sensitive configuration adapted both mode-specific instruction sets and sampling parameters (temperature and top-p) across age bands, balancing determinism for task-oriented guidance with greater linguistic variety for reflective prompts. The design's emphasis on supportive, bounded dialogue situates it within Conversational AI and Pedagogical Agent research on educational companion systems.

From Design Principles to Observable Indicators

A critical gap the paper addresses is the distance between normative principles and empirical evaluation. The framework translates five design principles — transparency of system identity and limits, scaffolding toward human contact, maieutic questioning, prevention of dependency dynamics, and data minimization — into observable, measurable indicators. These indicators are treated as signals supporting professional judgment, not exhaustive KPIs, deliberately combining light quantitative traces with qualitative rubrics and facilitator observation. Sources include interaction logs, structured rubrics, pre/post self-report measures, and post-interaction questionnaires. This principle-to-indicator mapping functions as an accountability device, connecting to Educational Measurement and making the intended pedagogical function testable through concrete traces. Transparency regarding system identity avoids anthropomorphic misrepresentation and builds Trust; data minimization and purpose limitation align with Privacy concerns in educational deployment.

Methods and the Italy–China Pilot

The study adopts an exploratory, mixed-methods research design appropriate to early-stage educational innovation, enabling iterative refinement while linking theory, design, and empirical observation. Pilot activities were conducted in Italy (with Edulife) and China (with Zhejiang Yizhong), involving N = 30 participants (12 Italian, 18 Chinese) across 20 bounded sessions (10 per context), spanning three broad age groups and including learners, teachers, and parents. The primary focus was technical and vocational education, where learners often face challenges related to orientation, Self-Efficacy, and continuity of engagement. Each participant engaged with both interaction modes in varied order to reduce sequence effects, with educators maintaining an oversight role consistent with the human-in-command principle. This design reflects Mixed-Methods Research and Design-Based Research methods. The author discloses positionality as co-founder of Edulife and deputy general manager at Zhejiang Yizhong, mitigating bias through facilitator oversight, triangulation across logs, rubrics, and surveys, and reporting evidence as exploratory rather than efficacy-based. The study also frames Italy–China as a cross-context stress test, connecting to Culturally Relevant Pedagogy and the principle that implementation must adapt to diverse K-12 and Adult Learners educational systems.

Initial Observations

Preliminary findings are reported descriptively and interpretatively, without claims of statistical generalization. Both interaction modes were perceived as useful but for different purposes: AmicoMio for clarity and task orientation, AmicoTuo for reflection and motivational support. Observable indicators showed that transition-to-human prompts occurred and micro-action completion rates were higher with structured prompts, while communicative autonomy improved between pre- and post-interaction phases, particularly after AmicoTuo sessions. Usage remained bounded with no clear patterns of excessive or exclusive use. Cross-context differences were context-sensitive rather than divergent: Italian participants emphasized the dialogical, meaning-making aspects of AmicoTuo, while Chinese participants emphasized its relational harmony and contextual sensitivity, with AmicoMio valued for technical reliability in work-oriented contexts. Notably, participants in both contexts recognized the system's orientation toward human mediation and did not report expectations of substitution or emotional dependency — evidence bearing on AI Misuse and Learning Harm and Reducing AI Misuse.

What this means for practice

• Instructors. Adopt human-in-command as a design requirement rather than a slogan: adults retain the educational decisions, relational boundaries, and ethical oversight while the AI supplies only temporary mediation.
• Keep the tool's role bounded and directional: cap session duration and frequency, and design each interaction to end in a next step with a teacher, peer, or community instead of more conversation with the AI.
• Separate task support from reflective support in the interface. The two modes were valued for different purposes — structured guidance for technical clarity, maieutic questioning for reflection and motivational support — and learners treated them as complementary.
• Evaluate by whether learners transition to human contact and act, not by engagement with the AI: the framework's indicators are transition-to-human prompts, micro-action completion rates, and pre/post communicative autonomy.
• Instrument the safeguards before scaling: flag repeated or exclusive use patterns, solicit no sensitive data, and prompt human contact when usage thresholds are reached.

Limitations

• The pilot covers N = 30 participants (12 in Italy, 18 in China) across 20 bounded sessions, and the preliminary findings are reported descriptively and interpretatively, with the author explicitly making no claim of statistical generalization.
• The author holds leadership roles in both partner organizations — co-founder of Edulife and Deputy General Manager at Zhejiang Yizhong Intelligent Technology Co., Ltd. — disclosing this as positionality and mitigating it only through facilitator oversight and triangulation across logs, rubrics, and surveys.
• Participants include learners, teachers, and parents across three broad age groups, so the pre/post self-report measures and rubrics mix roles and ages rather than isolating learner effects.
• There is no comparison condition: the study is exploratory design-based evidence on feasibility and perceived usefulness, not an efficacy evaluation, and larger-scale validation and longitudinal tracking of learner trajectories remain future work.

Citation

Pier Paolo Benedetti (2026). Design Principles and Observable Indicators for AI-Enabled Pedagogical Accompaniment: Evidence from the Amico Dual-Mode Prototype in Italy and China. Accepted at ICAIE 2026.