DataCanvas-EDU: An Agentic Framework for Instructor-Guided Synthetic Data Generation in Business Analytics Education

Synthesis: An, Hamdani & Fox (2026) identify a practical bottleneck in business analytics teaching: authentic datasets are hard to obtain, adapt poorly to a given course, and — when they are popular — have already been absorbed into LLM training data together with their worked solutions. Their response, DataCanvas-EDU, is an agentic framework in which an instructor states teaching goals in conversation and an AI agent writes generation code, checks the resulting data, and prepares assignments, reference analyses and rubrics across four phases: Plan, Create, Verify/Test Analysis and Evaluate. The paper illustrates the approach with a 15,000-order food-delivery case containing nine deliberately designed patterns, and packages the framework as a reusable agent skill.

Key Findings

1. Data contamination is the pedagogical problem, not just a data problem. When models have seen a dataset and its published analyses, students receive explanations drawn from existing work instead of practicing investigation of unfamiliar data.
2. Instructor intent is the specification. Teaching goals and intended patterns are expressed conversationally, and the agent's generation code is derived from them, keeping the design decision with the teacher.
3. Four phases organize the workflow — Plan, Create, Verify/Test Analysis, Evaluate — with instructor review and revision built into each rather than appended at the end.
4. Verification is internal to the pipeline: the agent checks the generated data against the intended patterns before materials are produced, addressing the failure mode where synthetic data silently lacks the structure a case is meant to teach.
5. A worked case demonstrates scale and control — 15,000 orders and nine designed patterns — and the framework ships as a reusable AI Agent Skill for compatible agent environments.

What the framework changes about case preparation

Preparing a data-analytics case traditionally consumes instructor time on pattern investigation, result verification, and writing both the assignment and its reference solution. Automating generation does not remove the instructor from that loop; it moves their work to specification and review. That division is what makes the approach credible for Assessment: rubrics and reference analyses are produced as first-class artifacts alongside the data, so the case arrives with its own grading apparatus.

Why synthetic data needs its own validity argument

A generated dataset can look realistic while failing to instantiate the pattern the case teaches, and can drift away from the distributions students will meet in practice. The Verify phase is the authors' answer, and it connects this work to the wider Assessment Validity and Benchmark questions that surround any AI-produced instructional material. It also raises the same reliability concerns that apply to other generative content pipelines — here, an incorrect generated pattern propagates into both the case and its rubric.

What this means for practice

• Instructors. State teaching goals and intended patterns conversationally, then review the agent's generation code and verification results before the case reaches students, so the design decision stays with you.
• Instructors. Distribute only the student materials and run reference analyses in a separate AI session, since a model that has already seen the case's solution has consumed the investigation you want students to perform.
• Instructional designers. Produce the rubric and reference solution as first-class outputs of the same pipeline that generates the data, so every case arrives with its grading apparatus attached.
• Instructional designers. Verify generated data against the intended patterns before packaging, and rerun the reference analysis whenever a pattern is strengthened - the authors' own consistency check found the verification program testing only Monday and Tuesday for the office-lunch rule while the generator and reference plot used all weekdays.

Limitations

• The framework is illustrated with a single WindowDash food-delivery case (15,000 orders, nine designed patterns) and evaluated on no students; no preparation-time, revision-effort or material-quality data is reported, and the authors list those as future studies.
• Validity of the synthetic distributions against real business data is untested, and the authors warn that generation and verification may share the same mistaken assumption, so an incorrect pattern can propagate into both the case and its rubric.
• One of the nine designed patterns, price-rating sensitivity for slow orders, rests on three observations above $100 and longer than 45 minutes; the authors concede these "provide limited support" and require the reference explanation to state that limitation alongside the finding.
• The agentic pipeline depends on an external agent environment (it ships as an AI Agent Skill), and generation quality is bound to the model and the code the agent writes, which the paper does not benchmark.

Connected Concepts

• Agentic AI
• Large Language Models (LLMs)
• Generative AI
• Learning Design
• Assessment
• Authentic Assessment
• Problem-Based Learning
• Experiential Learning
• Assessment Validity
• Hallucination Risk
• Educational Technology Developers

Connected Articles

• AgentSchool: An LLM-Powered Multi-Agent Simulation for Education — Multi-agent simulation in education
• Enacting Constructive Conflicts with AI Agents to Enhance Reconsideration among Novice Interaction Designers — Designing constructive conflict into educational agents
• Mapping the Emerging Curriculum for AI-Assisted Software Engineering via Syllabus Analysis — AI-assisted curriculum analysis
• The AI disruption in engineering education: an analysis of changing student norms through cultural historical activity — Disruption of engineering education by chat assistants

Citation

An, B., Hamdani, M., & Fox, J. (2026). DataCanvas-EDU: An Agentic Framework for Instructor-Guided Synthetic Data Generation in Business Analytics Education. arXiv:2609.19617.