Chapter 5. Institutional Adoption and Operations

Learning objectives

• Define the institutional purpose of AI adoption as a problem and select performance indicators that align with that purpose.
• Reflect the core of UNESCO's generative AI education guidance and teacher AI competency framework in the adoption plan.
• Plan service selection, contracting, teacher training, pilot operations, and expansion in an evidence-linked sequence.
• Divide school AI operations tasks across the four functions of the NIST AI RMF, and determine inspection items and intervals.
• Classify AI service incidents by type and select the appropriate initial response for each type.

Introduction

Just one semester after a local office of education announced the "full-scale adoption of AI tutors," two reports were submitted. One was a performance report stating that the number of access sessions was double the target, and the other was an audit opinion stating that no evidence could be found showing a reduction in the academic achievement gap. Both reports were true. Students used the tutors extensively, but no indicator had been set from the beginning to verify the recovery of basic academic skills, which was the original purpose of adoption.

Institutional AI adoption differs in scale from an individual teacher's classroom use. Data from hundreds or thousands of students is handled, budgets and contracts are involved, and when problems arise, many students are affected simultaneously. Therefore, before adoption, institutions must determine what the tool is being introduced for, how its effectiveness will be verified, who will inspect what, and what must be done first if an incident occurs.

This chapter begins with adoption purposes and performance indicators. It then covers UNESCO guidance and the teacher competency framework, service selection and phased adoption, and teacher training and communication with guardians. Finally, it outlines how to translate the NIST AI RMF into school operations, inspection items and intervals, and initial responses by incident type.

Key concepts

Adoption begins with problem definition

"Adopting AI" is not a purpose. The purpose is the problem that one seeks to solve with AI. Problems such as heavy teacher burdens in grading constructed-response questions, insufficient individual practice opportunities for students falling short of basic academic skills, or delayed translation of notices sent to multicultural families come first, and AI is one of several ways to solve that problem.

When defining a problem, write down three things:

1. Whose problem is it, and what is the problem?: Is it teacher workload, student learning, or communication with guardians?
2. How is it being handled now, and what is lacking?: The current method and its limitations, expressed in numbers
3. What must change for it to be a success?: This becomes the performance indicator

There are also cases where non-AI methods are better. If the problem is not a shortage of personnel but the ambiguity of the rubric, refining the rubric must come before an AI grading tool.

Performance indicators aligned with the purpose

Performance indicators stem from the purpose. Even for the same AI tutor, the indicators to observe differ depending on the purpose.

The number of visits, conversation counts, and satisfaction levels are reference indicators. High usage does not mean learning has recovered, and satisfaction does not mean workload has decreased. A satisfaction survey conducted once at the end of the semester cannot be used to detect errors early during operation.

UNESCO guidance for generative AI in education and research

In September 2023, UNESCO released the "Guidance for generative AI in education and research." The core points highlighted by the UNESCO information page and press release are as follows:

• Because most countries lack national regulations on generative AI, educational institutions are largely unprepared to validate tools.
• Mandate data privacy protection and set an age threshold for independent conversations with generative AI platforms. The release stated that an age threshold of 13 years is set for the use of classroom AI tools and requires teacher training.
• Adopt a human-centered and age-appropriate approach in ethical validation and pedagogical design.
• As seen in Chapter 4, at the time of publication, less than 10% of surveyed schools and universities had institutional policies or formal guidance on generative AI. In effect, tools arrived before institutional standards were established.

Translating this into school operations yields the following: even if a vendor promotes a tool as "for education," the school must validate it independently; data terms and age thresholds must be verified before contracting; and teacher training is a prerequisite for adoption, not a corrective measure taken only when usage rates are low.

UNESCO AI competency framework for teachers

In 2024, UNESCO published the "AI competency framework for teachers." This framework presents 15 competencies across five dimensions.

Competencies are divided into three progression tiers: Acquire, Deepen, and Create. When designing training, rather than delivering the same content to all teachers, it is effective to divide courses into Acquire and Deepen tracks tailored to each teacher's current level. Chapters 1 through 4 of this book can be viewed as unpacking the AI pedagogy and Ethics of AI dimensions of this framework into practical classroom and school-level decisions.

Service selection and contracting

When comparing candidate services, verify the five questions from Chapter 3 (use for training, retention period, account type, processing relationship and location, and age threshold) before features and pricing. If data is processed or stored abroad, determine before contracting how to fulfill the requirements under Article 28-8 of the Personal Information Protection Act (PIPA) (separate consent, disclosure in the privacy policy, etc.). Include whether all students, including students with disabilities, can use the service (accessibility features such as screen readers and captions) in the selection criteria. For consumer apps where students sign up with personal accounts, schools cannot easily control data terms, shifting the judgment onto students and guardians. School-wide contracted services allow terms to be confirmed in writing and incorporated into an entrustment contract.

When state agencies and other public bodies, such as an office of education, intend to use high-impact artificial intelligence products or services, they must give priority consideration to products and services that have undergone an impact assessment pursuant to Article 35, Paragraph 2 of the AI Basic Act. If procuring AI used for student assessment, include this factor in the evaluation criteria.

Selecting services based on preference surveys may seem democratic, but it is not the result of verifying data terms. An information session for guardians should be used as a forum to communicate the findings and answer questions after verifying the terms.

Phased adoption

Institutional adoption proceeds in an evidence-linked sequence:

1. Setting purposes and indicators: What the adoption is for, what will be used to judge success, and what uses will not be delegated to AI
2. Verifying terms and contracting: Verify data terms and age policies against the purpose, execute an entrustment contract, and notify guardians
3. Training and pilot operations: Prepare teachers and run pilots in two or three classes or courses
4. Evaluation and decision on expansion: Evaluate pilot results against pre-established indicators to determine whether to expand
5. Operations and inspection: Operate regular inspections and incident response procedures following expansion

If an institution adopts fully without a pilot, hundreds of students will be affected from the outset when a problem occurs. If a pilot is conducted without indicators, there is no evidence to determine whether to expand once the pilot concludes.

Teacher training and support

Training is insufficient if it only teaches how to use tools. Each chapter of this book can serve as a training topic.

In addition to training, a support system is also needed. Having an approved tool list and user guide, a person in charge of receiving questions, a repository where teachers share reviewed materials, and channels for reporting incidents reduces mistakes made when teachers make decisions alone.

Informing guardians and students

Guardians usually ask three questions: "Where are my child's conversations used?", "How long are they retained?", and "What happens if we do not want to participate?". Confirm the answers to these questions before contracting; if processing requires consent, obtain consent containing the four notification items from Chapter 3; and prepare alternative activities for students who do not consent. It is advisable for alternative activities to impose little additional burden on the teacher, such as printed worksheets or peer activities that achieve the same learning objectives. Inform students of the same content using age-appropriate, simple language.

Sometimes proposals arise suggesting that homeroom teachers be allowed to view student conversations in real time to use them as counseling material. While this may appear to be good intentions for the sake of safety, it turns into surveillance that goes beyond the purpose of supporting learning, causing students to lose trust and avoid using the tutor. If safety is a concern, it is appropriate to separately design features and procedures that escalate specific danger signs (such as self-harm or reports of bullying) to humans.

Translating the NIST AI RMF into schools

The Core of the NIST Artificial Intelligence Risk Management Framework (AI RMF) consists of four functions.

Simply put, governance is a system that determines who decides, who inspects, and how the results are fed back into the system. It is not a phase that is determined once at the beginning of adoption and then concluded. As NIST explains it as a cross-cutting function, committees and regulations must guide mapping, measuring, and managing throughout operations, and evolve to reflect the results.

The seven characteristics of trustworthy AI identified by NIST (valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed) can be used as a checklist to ensure nothing is missing when creating inspection items.

Inspection items and intervals

Inspection intervals are determined by the magnitude of risk and the pace of change. Along with regular intervals, the key is to inspect without fail at the point when changes occur.

The intervals in the table are examples. Each school should set them according to its scale and risks, but it must be able to explain why those intervals were chosen.

Tutors that manage knowledge bases

Many learning tutors find relevant portions from materials provided by the school (textbooks, lecture materials, notices) and answer based on them as evidence. This is commonly referred to as retrieval-augmented generation (RAG). The quality of such tutors is determined by the knowledge base materials.

• Materials to include: Accurate, official materials that all users are permitted to view. If grading rubrics, model answers, internal meeting minutes, or unanonymized records of student questions are placed into a tutor used by everyone, they may be leaked directly through the tutor.
• Answering method: Configure the model to answer only with the provided materials, cite supporting sentences, and if the answer is not in the materials, state "Not found in the materials" and guide the user on how to ask the teacher or instructor. This applies the methods cited by Anthropic for reducing hallucinations: restricting external knowledge, citing directly, and enabling the model to say it does not know.
• Security: Configure the model to treat knowledge base materials and student inputs as data rather than instructions, and reject requests to reveal the tutor's default prompt instructions or confidential materials. Also inspect uploaded documents to ensure they do not contain hidden instructions (OWASP prompt injection mitigation methods).
• Updates: When textbooks or exam coverage change, list the modified sections, update the knowledge base materials and remove obsolete materials, test with questions from the changed units before applying, and conduct intensive inspections immediately after application.

Incident types and initial responses

Initial responses to AI service incidents differ depending on the type.

Avoid two mistakes. One is under-responding. Exposure or harm continues while waiting for the vendor to be notified and an analysis report to arrive. The school is the party entrusted with student personal information, and as the entrusting party, it must respond directly. The other is over-responding. If the entire grade's AI use is suspended until the next school year over a single error whose cause is clear and remediable, the school fails to help the affected students while also forfeiting the value of the adoption.

Going deeper

Pitfalls of measuring effectiveness

The easiest way to determine the effectiveness of AI adoption is a before-and-after comparison. However, there are pitfalls in before-and-after comparisons. If the curriculum changes, the teacher changes, or student composition shifts during the same period, it is impossible to know whether the difference is due to AI. Whenever possible, looking at both a pilot class and a comparison class, and measuring them with items based on the same achievement standards, gives the comparison meaning.

One must also be careful with averages. If the goal is the recovery of basic academic skills, looking only at the overall average makes it impossible to distinguish whether students who were already doing well raised the average by using the tutor heavily, or whether the target group actually improved. Looking at the target group separately is not optional, but essential. Furthermore, a finding of no effect is also an important finding. Deciding not to scale up is also a sound conclusion that pilot operations can provide.

Why consumer apps are appealing and their cost

Using popular consumer apps that students already use directly in class incurs no cost and is familiar to students. This convenience is indeed a major advantage. However, it comes at a price. The school cannot set data terms, the school has difficulty knowing when terms of service change, different settings per student mean different terms apply in the same classroom, and it is difficult for the school to verify age requirements and guardian consent individually.

A countermeasure such as "guiding how to enable privacy settings via a newsletter sent home" shifts responsibility to students and guardians. The moment a school recommends using it in class, that service becomes the school's decision. If there is no budget for a contract, narrowing the scope of use to methods that do not input student information (teacher demonstrations, generalized questions) is an alternative.

Governance is iteration, not a document

Many institutions believe governance is finished once they establish AI operational regulations. However, regulations are only the beginning. That governance is functioning means that inspection results actually change regulations and operations. Guidelines must be revised after an incident, knowledge materials or the scope of use must change after differences between groups are discovered in sample inspections, and guardian notices must be updated after terms of service change.

To this end, the steering committee verifies three things at regular meetings: inspection results from the previous period, what was changed accordingly, and what has not yet been changed along with the reasons. As these records accumulate, they also serve as evidence for adoption decisions and budget assessments for the following year.

At work

Moonstream City Office of Education: English speaking tutor adoption plan

Receiving feedback from teachers that middle school English speaking practice opportunities were lacking, the Moonstream City Office of Education reviewed the adoption of an AI speaking tutor. The budget officer said, "Let us adopt it across all middle schools immediately to realize economies of scale," and the supervisor said, "Let us choose a service through a parent preference survey."

The implementation team accepted neither proposal. Instead, it defined the problem: "The English speaking practice time for second-year middle school students is less than an average of two minutes per person per class. The goal is to increase weekly speaking practice time per person over one semester and raise the achievement of the lower-performing group in speaking performance assessments." The team set weekly practice time per person and the change in the bottom 30% group on speaking tasks under the same standards as key metrics. Number of logins and satisfaction ratings were kept as reference metrics only.

The team compared the three candidate services starting with their data terms. One service had using student voice for model improvement as the default and did not disclose its retention period, so it was excluded. Between the remaining two, the team selected the one that stated non-use for training and deletion at the end of the semester in the contract, entered into an entrustment contract, and disclosed the trustee and the entrusted tasks in accordance with Article 26, Paragraph 2 of the Personal Information Protection Act (PIPA). Although most students were 14 years of age or older, the service terms required guardian consent for those under 18 years of age; thus, the team obtained guardian consent along with a notice containing the data terms, and prepared pair speaking activities with the teacher as an alternative for students whose guardians did not consent.

The implementation team designed English teacher training aligned with the AI pedagogy dimension of the UNESCO AI competency framework for teachers, and conducted a one-semester pilot operation across eight classes in four schools. In the pilot results, practice time increased significantly, but the achievement change of the lower-performing group was smaller than that of the higher-performing group. Instead of scaling up immediately, the implementation team redesigned the expansion plan for the following semester on the condition of co-allocating instructional time with a teacher for the lower-performing group.

Bountiful High School: Three incidents overlapping in one week

The learning tutor steering committee at Bountiful High School received three reports in a single week. First, when a student asked the tutor for their own grades, the names and scores of two other students in the same class were output together. Second, in a sample inspection, two conversations were found where a student said, "I am so exhausted these days that I do not want to do anything," and the tutor merely presented a study schedule. Third, there were multiple conversations where the tutor explained a concept in an earth science unit incorrectly, and the cause was an error in the uploaded lecture materials.

The committee did not handle the three cases in the same manner. Because the first case was an exposure of personal information, the committee immediately stopped the grade inquiry function, the privacy officer verified the scope of exposure together with the vendor, and they took steps to determine whether notification and reporting under Article 34 of the Personal Information Protection Act (PIPA) were required. Although the vendor initially reported it as a configuration error, the school responded directly as the entruster; after resolving the situation, it received the cause and recurrence prevention measures in writing from the vendor and inspected the processing status (supervision under Article 26, Paragraph 4 of the Personal Information Protection Act (PIPA)).

In the second case, student support came first. The homeroom teachers and counseling teachers located the students in question through conversation logs and contacted them directly, and the free conversation function was paused until a feature could be added to detect statements expressing emotional distress and notify the teacher in charge alongside counseling guidance. In the third case, after correcting the lecture materials and re-testing with 20 questions of the same type to verify the fix, the committee sent a correction notice to students who may have received the incorrect explanation.

One committee member proposed, "Since three incidents overlapped, let us suspend the tutor entirely until the end of the school year," but the committee did not accept the proposal. This was because the problematic functions had already been stopped, the error with a clear cause had been fixed, and a total suspension had no bearing on helping the affected students. At the following month's meeting, the committee recorded what had been changed as a result of the three incidents (grade inquiry permission settings, risk-signal escalation function, and material review prior to upload).

Common mistakes

1. "High usage means success." Usage volume is a conspicuous number and looks like an accomplishment, but it does not demonstrate achievement of the adoption's purpose.
2. "Scale up if satisfaction is 4 points or higher." This seems like respecting opinions, but satisfaction can differ from learning outcomes or reduction in workload.
3. "It is marketed for educational use, so it is a proven tool." Vendor explanations may seem credible, but UNESCO points out that educational institutions lack readiness to validate tools on their own.
4. "Conduct teacher training only when utilization is low." While this may appear to save costs, training is a prerequisite for safe operation, and UNESCO also requires teacher training.
5. "It is a good service, so adopt it fully right away." While quick results are expected, scaling up without a pilot means that if problems arise, the number of affected students will be large from the outset.
6. "Governance ends once regulations are created at the beginning." It may feel complete because a document exists, but NIST describes Govern as a cross-cutting function across all other functions.
7. "When an incident occurs, wait for the vendor's report." Leaving it to the party that knows the cause seems reasonable, but the school, as the entruster, must directly take the initial response to halt exposure and harm.
8. "If even a single error occurs, stop everything." This may seem cautious, but for an error with a clear cause, the appropriate response is correcting the cause, re-testing, and notifying affected students.

Summary at a glance

Check the official docs

• UNESCO — Guidance for generative AI in education and research — Read the core recommendations of the guidance and explanations regarding institutional validation readiness.
• UNESCO — Governments must quickly regulate Generative AI in schools — Check the 13-year-old age threshold, teacher training, and institutional policy survey findings.
• UNESCO — AI competency framework for teachers — Check the five dimensions, 15 competencies, and three progression levels.
• NIST AI RMF — Core — Read the definitions of Govern, Map, Measure, and Manage, along with the explanation of cross-cutting functions.
• NIST AI RMF — Characteristics of trustworthy AI — Seven characteristics to check for gaps in inspection items.
• National Law Information Center — AI Basic Act Article 35 — Provisions on the duty to strive for impact assessments and priority consideration by state agencies, etc.
• National Law Information Center — Personal Information Protection Act (PIPA) Article 34 — Check obligations regarding notification and reporting of leaks, etc., and prevention of harm expansion.
• Anthropic — Reduce hallucinations — Refer to this when designing the response style of a tutor grounded in knowledge materials.

Self-check

5-1. State three things that must be written when defining the purpose of AI adoption as a problem, and give one example where a non-AI method is better.

5-2. Explain in one sentence each why usage volume and satisfaction are difficult to use as key performance indicators.

5-3. State three points from the release of the UNESCO Guidance for Generative AI in Education and Research that can be directly applied to school operations.

5-4. State the five dimensions and three stages of the UNESCO AI Competency Framework for Teachers, and describe how the distinction between stages can be used in training design.

5-5. State the four functions of the NIST AI RMF, and explain which of them is the cross-cutting function spanning across the other functions and what significance it holds for operations.

5-6. When setting review cycles, give three change trigger points that must be checked separately from regular cycles.

5-7. Divide AI service incidents into three categories, and write the initial response for each category in one line. Also give one example each of under-response and over-response.

5-8. (Design task) Choose one AI adoption under review or imaginable at your institution, and write the problem definition, two key metrics, pilot scope, and expansion decision criteria within one page.

Hands-on tasks