One way to do this is through learning loops, which is a practice borrowed from product development. It begins with an explicit hypothesis: We believe X will happen because Y. This will feel familiar to universities. It mirrors the logic of hypothesis testing that underlies academic research itself. The difference is, the object of inquiry is not a disciplinary question, but an institutional choice. 

‍

A learning loop has four elements: an explicit hypothesis, indicators to monitor, a defined observation period, and decision criteria about what happens next.

Articulating a hypothesis forces clarity about what you are actually testing and separates institutional learning from vendor claims. Vendors arrive with their own narratives about impact. Without an institutional hypothesis, adoption easily becomes an implicit endorsement of someone else's vision.

‍

When defining indicators or signals to monitor, students should be included as participants in the process, not merely subjects of study. They are experiencing AI’s effects across multiple contexts simultaneously. Their observations about how AI is shaping their learning, their workflows, and their relationship to intellectual work are data that institutions need and should collect systematically.

‍

These observations are not substitutes for controlled research. They do not establish causation. They are indicators of trajectory; evidence about whether institutional choices appear to be moving in the intended direction. This kind of monitoring is not a substitute for rigorous research on AI and learning. It is a governance tool for institutions that must make decisions while that research is still developing.

‍

In a university context, there are two broad signals worth attending to, in addition to what institutions will define locally.

‍

The first is displacement. Every efficiency gain from AI automation displaces something. Sometimes what is displaced is acceptable, even desirable. Sometimes it is not, and not all displacement is equivalent. Shedding administrative burden is categorically different from eroding the relational and intellectual core of teaching and learning: the informal moments of connection during advising, the experience of receiving feedback from professors, the slow reading that builds expertise in research. In some cases, AI shifts not just workflows but what counts as knowledge or intellectual labor. Every hypothesis in a learning loop should identify the human interaction at risk and define which displacements are tolerable and which are stopping conditions.

‍

The second signal worth attending to is adoption itself. If faculty, staff and students are bypassing institutional tools for personal AI accounts, that is not merely a policy issue. It is feedback about whether your infrastructure choices are serving the people who use them. That information should shape decisions about vendors, investment, and whether to build rather than buy. Institutions often have more capacity to build than they recognize, including technical talent and deep contextual knowledge.

‍

To break endless data-gathering, learning loops are timeboxed. A semester, a quarter, a defined observation period. If possible, before observation begins, decision criteria are established. Under what conditions do we continue? Under what conditions do we revise or stop? This prevents initiatives from drifting indefinitely and reduces the temptation to rationalize ambiguous results.

‍

A teaching and learning center is piloting an AI tool to help faculty develop course materials, including learning objectives, assessments, and scaffolded activities aligned to outcomes. A learning loop for this pilot might look like:

‍

Hypothesis: “We believe AI-assisted course design will reduce the time faculty spend on initial course development by 30%, while maintaining alignment between objectives, activities, and assessments.”

‍

Category: Faculty experience

‍

Displacement risk: The slow, generative work of designing a course is often where faculty deepen their own understanding of what they want students to learn. It surfaces questions: What do I actually care about here? What does mastery look like? What will reveal whether students get it? Automating this process may produce coherent syllabi and aligned rubrics but shortcut the thinking that makes teaching intentional and responsive. We will watch for whether faculty report feeling less connected to their own course design, and whether courses become more generic over time.

‍

Timebox: Two semesters (to capture full design-through-teaching cycle)

‍

Leading indicators: Faculty time spent on course development, faculty sense of ownership over course design, alignment quality as assessed by instructional designers, student experience of course coherence, faculty observations about whether the AI draft was a starting point or a crutch.

‍

Decision criteria: If faculty report faster development without loss of intentionality, using AI drafts as scaffolding they substantially rework, continue and refine prompts. If faculty report “just accepting” AI suggestions or feeling distanced from the design, revisit the workflow and consider whether the tool is undermining the pedagogical value of the design process itself.

‍

Defining and implementing learning loops at the unit level preserves autonomy. The risk for institutions is either paralysis while waiting for a comprehensive AI strategy, or fragmented experimentation that never aggregates into institutional insight.

‍

There is a way to create “good enough” conditions now that can inform strategic decisions later. Institutions can agree on broad categories of value without resolving every debate about institutional purpose, such as student engagement, faculty and staff experience, operational capacity, and institutional sustainability. Individual units define their own hypotheses and indicators within these categories. A shared frame allows distributed learning to accumulate into institutional knowledge, avoiding the thousand-flowers problem I named in Part 1. A thin, agreed-upon frame allows local experimentation to produce insight that leadership can synthesize and act upon.

‍

As campus-wide AI strategy and governance conversations progress, institutions with structured learning underway will have something many won't: actual evidence. Not anecdotes, not vendor narratives, but accumulated institutional knowledge about what is working, what is being lost, and where the tradeoffs lie. That evidence will matter when leaders are deciding where to expand investment, where to pause, and where to stop.

‍

Higher education is built on inquiry. The challenge for institutions now is to apply that discipline to themselves — to remain present to what they value, accountable to evidence, and attentive to education as a fundamentally human endeavor.

‍

One way to do this is through learning loops, which is a practice borrowed from product development. It begins with an explicit hypothesis: We believe X will happen because Y. This will feel familiar to universities. It mirrors the logic of hypothesis testing that underlies academic research itself. The difference is, the object of inquiry is not a disciplinary question, but an institutional choice. 

‍

A learning loop has four elements: an explicit hypothesis, indicators to monitor, a defined observation period, and decision criteria about what happens next.

Articulating a hypothesis forces clarity about what you are actually testing and separates institutional learning from vendor claims. Vendors arrive with their own narratives about impact. Without an institutional hypothesis, adoption easily becomes an implicit endorsement of someone else's vision.

‍

When defining indicators or signals to monitor, students should be included as participants in the process, not merely subjects of study. They are experiencing AI’s effects across multiple contexts simultaneously. Their observations about how AI is shaping their learning, their workflows, and their relationship to intellectual work are data that institutions need and should collect systematically.

‍

These observations are not substitutes for controlled research. They do not establish causation. They are indicators of trajectory; evidence about whether institutional choices appear to be moving in the intended direction. This kind of monitoring is not a substitute for rigorous research on AI and learning. It is a governance tool for institutions that must make decisions while that research is still developing.

‍

In a university context, there are two broad signals worth attending to, in addition to what institutions will define locally.

‍

The first is displacement. Every efficiency gain from AI automation displaces something. Sometimes what is displaced is acceptable, even desirable. Sometimes it is not, and not all displacement is equivalent. Shedding administrative burden is categorically different from eroding the relational and intellectual core of teaching and learning: the informal moments of connection during advising, the experience of receiving feedback from professors, the slow reading that builds expertise in research. In some cases, AI shifts not just workflows but what counts as knowledge or intellectual labor. Every hypothesis in a learning loop should identify the human interaction at risk and define which displacements are tolerable and which are stopping conditions.

‍

The second signal worth attending to is adoption itself. If faculty, staff and students are bypassing institutional tools for personal AI accounts, that is not merely a policy issue. It is feedback about whether your infrastructure choices are serving the people who use them. That information should shape decisions about vendors, investment, and whether to build rather than buy. Institutions often have more capacity to build than they recognize, including technical talent and deep contextual knowledge.

‍

To break endless data-gathering, learning loops are timeboxed. A semester, a quarter, a defined observation period. If possible, before observation begins, decision criteria are established. Under what conditions do we continue? Under what conditions do we revise or stop? This prevents initiatives from drifting indefinitely and reduces the temptation to rationalize ambiguous results.

‍

A teaching and learning center is piloting an AI tool to help faculty develop course materials, including learning objectives, assessments, and scaffolded activities aligned to outcomes. A learning loop for this pilot might look like:

‍

Hypothesis: “We believe AI-assisted course design will reduce the time faculty spend on initial course development by 30%, while maintaining alignment between objectives, activities, and assessments.”

‍

Category: Faculty experience

‍

Displacement risk: The slow, generative work of designing a course is often where faculty deepen their own understanding of what they want students to learn. It surfaces questions: What do I actually care about here? What does mastery look like? What will reveal whether students get it? Automating this process may produce coherent syllabi and aligned rubrics but shortcut the thinking that makes teaching intentional and responsive. We will watch for whether faculty report feeling less connected to their own course design, and whether courses become more generic over time.

‍

Timebox: Two semesters (to capture full design-through-teaching cycle)

‍

Leading indicators: Faculty time spent on course development, faculty sense of ownership over course design, alignment quality as assessed by instructional designers, student experience of course coherence, faculty observations about whether the AI draft was a starting point or a crutch.

‍

Decision criteria: If faculty report faster development without loss of intentionality, using AI drafts as scaffolding they substantially rework, continue and refine prompts. If faculty report “just accepting” AI suggestions or feeling distanced from the design, revisit the workflow and consider whether the tool is undermining the pedagogical value of the design process itself.

‍

Defining and implementing learning loops at the unit level preserves autonomy. The risk for institutions is either paralysis while waiting for a comprehensive AI strategy, or fragmented experimentation that never aggregates into institutional insight.

‍

There is a way to create “good enough” conditions now that can inform strategic decisions later. Institutions can agree on broad categories of value without resolving every debate about institutional purpose, such as student engagement, faculty and staff experience, operational capacity, and institutional sustainability. Individual units define their own hypotheses and indicators within these categories. A shared frame allows distributed learning to accumulate into institutional knowledge, avoiding the thousand-flowers problem I named in Part 1. A thin, agreed-upon frame allows local experimentation to produce insight that leadership can synthesize and act upon.

‍

As campus-wide AI strategy and governance conversations progress, institutions with structured learning underway will have something many won't: actual evidence. Not anecdotes, not vendor narratives, but accumulated institutional knowledge about what is working, what is being lost, and where the tradeoffs lie. That evidence will matter when leaders are deciding where to expand investment, where to pause, and where to stop.

‍

Higher education is built on inquiry. The challenge for institutions now is to apply that discipline to themselves — to remain present to what they value, accountable to evidence, and attentive to education as a fundamentally human endeavor.

‍