UVU's Kahlert Institute for Applied AI trained faculty to build custom AI tools this summer. Across six profiles, a pattern emerges: instructors engineering restraint into AI, and shifting from teaching AI use to teaching AI management.
Orem, Utah — August 10, 2026
Utah employers are about to start seeing something new on incoming resumes: graduates who list "managing an AI agent" as a job skill, not just "using AI."
That shift traces back to a two-week stretch in May, when roughly 37 Utah Valley University faculty members went through the Kahlert Institute for Applied AI (KAAII) Summer Institute — UVU's answer to a problem every Utah employer hiring new grads is already living with, whether they've named it yet or not: what, exactly, should a 2026 graduate be able to do with AI that a 2021 graduate couldn't, and what should they still be expected to do without it.
Rather than hand instructors a policy memo, KAAII gave them two weeks, a knowledge base, and a prompt: build something you will actually use this fall. What came out the other side is a set of faculty-built AI agents, tutors, and workflow tools that will show up in UVU classrooms, and eventually in Utah's workforce, starting this month.
It's the latest entry in a broader pattern of Utah institutions building explicitly toward the state's tech economy, the same instinct behind the spaceport siting process, the Weber County nuclear symposium, and the Utah Quantum Roundtable: state institutions positioning early for where the workforce is headed rather than reacting once it arrives.
One faculty member in this cohort said so directly. George Rudolph, chair of UVU's computer science department, is already using this summer's momentum to prep for what he sees as the next wave after this one: he's building out quantum computing coursework now, years ahead of the technology's expected commercial arrival, on the theory that Utah's tech sector will need that workforce ready before it's obvious the demand exists.
But the more interesting story, once you talk to enough of these instructors, is what they actually think Utah employers want from the students on the other end of this. It isn't fluency with a chatbot — for most incoming students, that bar is more or less already cleared. It's judgment over an AI's output, and increasingly, the ability to manage AI the way a manager runs a junior hire. Noah Myers, who teaches accounting, put it most bluntly: entry-level "staff" work, the manual reconciliation and data cleanup that used to be a junior accountant's first few years on the job, is increasingly done by an AI agent now, and public accounting firms are telling him directly that they want graduates who can walk in ready to supervise and review that work, not perform it by hand. Diego Alvarado-Karste, in marketing, reaches for a music metaphor to describe the same shift: AI has handed everyone a fully capable instrument, so the only differentiator left for a hiring manager is who can direct the whole performance rather than just play a note competently. Rudolph built an entire semester's coursework around that same distinction, moving his graduate AI course from single-agent code to multi-agent orchestration specifically because that's the shape of work he expects employers to actually be hiring for. A marketing professor, an accounting professor, and a computer science department chair arrived at the same conclusion from three unrelated directions: the scarce skill in an AI-saturated workplace isn't generating an answer, it's knowing what to do with one.
The tools themselves, when you look closely at how they're built, tell a second story worth an employer's attention: these faculty spent the summer engineering restraint into AI, not capability. Angie Carter's writing tutor won't draft a sentence for a student. Majid Memari's course chatbot is built to refuse rather than guess. Across English, computer science, chemistry, marketing, accounting, and sociology, instructors independently arrived at the same design instinct — the risk isn't AI getting something wrong, it's AI being too willing to just hand over the answer — which is a fair preview of the kind of AI-supervision habits these same students are being trained to bring into a Utah workplace next.
Coach Carter's 3 a.m. problem
Dr. Angie McKinnon Carter has taught writing at UVU since 2005. She has been full-time since 2010 and a senior lecturer in the English department since 2018. Her students call her Coach Carter. She holds a Ph.D. in composition and applied linguistics from Indiana University of Pennsylvania.
The problem she brought to the institute wasn't abstract. This fall she's teaching a hybrid composition course. She sees students in person only once a week, which sharply cuts the number of times she can walk the room during in-class writing and catch a student before they get stuck. And when they do get stuck, it's rarely during office hours.
"I'm hoping that students, as they are writing, they get to a point where they're like, I'm stuck," Carter said. "That's where I'm hoping they're going to use the tutor... you should be asleep, but you're not. So here's a tutor bot."
Over roughly five to six hours spread across the two-week institute, Carter built what she calls the Research Argument Revision Coach — an AI tutor trained specifically on her own course materials, not a generic chatbot. She started from a walkthrough video a colleague circulated by email showing how to build a custom GPT, then built a parallel version as a Claude Project.
The knowledge base is the point. Carter loaded in her assignment prompts for both the research paper and the literature review, sample student papers she had permission to use, syllabus materials, textbook PDFs, and the PowerPoint decks she uses to teach synthesis, quotation integration, and what she calls "entering the conversation," pushing students past summarizing sources toward making an actual argument.
That specificity, she said, is what separates her tool from a student simply asking a general-purpose chatbot for writing help. Generic models, in her experience, routinely coach students toward a source-by-source literature review structure — the opposite of what she teaches, which organizes sources by thematic category. "If the large language model is telling you to do it source by source, you should be able to evaluate that and know that that's not what we're doing in this class," Carter said. "And yet, for various reasons, they don't."
By design, the tutor doesn't draft for students. If a student asks it to write an example paragraph, it's built to redirect, or to help without doing the sentence-level work for them. What it does do is diagnose: a student says where they feel stuck, and the tool responds with suggestions and analysis for them to bring, in writing, to a scheduled meeting with Carter.
That last step matters to her. It's not outsourcing the teaching; it's front-loading the meeting so the two of them aren't starting from zero. With sections that can run up to 92 students and only two required one-on-one meetings per semester, the tutor is meant to close a bandwidth gap she said is real, not a convenience: "This is solving a real problem of my bandwidth."
The limits she's building in on purpose
Carter isn't positioning the tool, or herself, as an unqualified adopter. She described herself as "a cautious adopter," someone in the middle of her department's spectrum, between colleagues who embrace AI fully and colleagues, largely fellow writers, who see large language models as a threat to creative and academic voice.
She's candid about the tool's limits: hallucination risk, and a knowledge base that can still be misinterpreted. Her bigger worry is that students aren't yet equipped to know when the tutor's advice is good and when it isn't — "what if you don't know enough to know whether it's giving you garbage or something gold?" — which is exactly why the human meeting stays in the loop.
She also flagged the ethical objections some of her students raise, such as water and energy use tied to data centers, and concerns about how training data was sourced, as legitimate, not something to wave off. She won't require use of the tool; she'll encourage students to try it while leaving room for those who opt out.
Carter's broader concern, one she said she's raised previously on a podcast with Matt Siebold, is timing: that higher ed is pushing AI fluency hard enough, fast enough, that students risk losing foundational writing skills before anyone knows whether current AI tools and pricing are durable. "I'm still not entirely convinced that AI is here to stay," she said, pointing to market-valuation coverage in the Wall Street Journal as part of her skepticism. Her goal for the semester is threading that needle: teaching students to write with AI and without it, in case the tools get "prohibitively expensive" or simply go away.
Her advice to colleagues considering the same path: start with the tutorial video, borrow her materials as a template, and, more than anything, talk to students directly about where the line is between using AI as a tool and using it to plagiarize, since instructors and students, she said, usually aren't operating from the same definition of cheating. She's floated an informal, no-slides faculty meetup this fall for anyone who wants to compare notes the same low-key way she started.
The marketing professor who traded a fake boss for a real one
Dr. Diego Alvarado-Karste is an assistant professor of marketing at UVU, where he's taught social media marketing and content marketing since 2020. He grew up in Cuenca, Ecuador, earned an undergraduate degree in economics and an MBA in Mexico, then a Ph.D. in marketing with a minor in anthropology from the University of North Texas. His published research spans brand rivalry, consumer perception of brand gender, and psychological ownership.
His institute project started as one thing and became another. He originally wanted to build a web-based simulation where students would negotiate a project with a fake boss. However, building and clearing institutional approval for a standalone piece of software wasn't a realistic path inside a two-week institute. Talking it through with colleagues at the institute, he shifted the whole concept inside an existing AI platform instead: what he built is a single long-form prompt script, dropped directly into a student's own Claude or ChatGPT account, that runs an hour-long conversation with an AI "boss" who assigns a project and then pushes back on the student's thinking at every stage, questioning their audience assumptions, challenging their read on the competition, asking them to defend and refine their approach. All of this happens before the finished output gets pasted into Canvas for him to grade. He built it around the standard marketing funnel — awareness, interest, desire, action, conversion — with the AI walking a student through each stage rather than just producing a final deliverable on request.
The shift, he said, came out of a specific gap he'd noticed in his content marketing courses: students were becoming fluent in AI tactics, using it to generate a portfolio, a blog, a piece of content, faster than he could keep teaching around it. His worry wasn't that students were using AI, but that "doing" was crowding out marketing strategy — he described it as everyone suddenly being handed a fully capable instrument, guitar or piano, and being able to play it competently, which makes the differentiator no longer whether you can produce content but whether you can direct it toward an actual strategy. He wanted the simulation to force that discussion, a back-and-forth a student can't shortcut, rather than let a model simply hand back a finished product.
He was candid that he came into the institute without agent-building experience of his own, and said the most useful part of the two weeks wasn't a specific breakthrough but the friction of other faculty questioning his idea out loud, colleagues pushing him on why his simulation needed to exist when others were already on the market, forcing him to sharpen what made his version different. He drew a direct line from that experience to what he built: a tool explicitly designed to be harder to please than a typical AI chatbot, on the theory that AI's tendency toward agreeableness undercuts learning as much as it helps it.
His advice to a colleague just starting out was practical rather than technical: match the tool to your own weakness. He described himself as creative but disorganized, and said he's used Claude's project-organization features to get his own course materials and research sorted, not to generate content, but to manage the sprawl of it, while colleagues who are naturally organized but less generative use AI differently. His broader pitch for UVU, delivered more as a rallying line than a policy point, was that the university should aim to produce "doers" rather than just critical thinkers, since AI has made execution cheap and strategic judgment the scarcer skill.
Asked what he'd want students to carry out of that shift, he returned to his own background, arriving in the U.S. from Ecuador without an academic pedigree, and the recurring feeling of not being "enough" in rooms where he didn't have the same preparation as everyone else. His closing point wasn't really about AI tooling at all: he argued that confidence — what he called audacity — is as much a prerequisite for the AI-driven "doing" economy he's describing as any specific skill, and that building that in students matters as much as teaching them a tool.
The department chair who's been waiting thirty years for this
George Rudolph chairs UVU's Department of Computer Science and is heading into his fifth year in the role this fall. His entire academic pedigree — bachelor's, master's, and Ph.D. — is in computer science from BYU, with a doctoral research focus in machine learning and neural networks decades before either term was in wide public use. Between degrees and teaching posts, he spent seven years as a software developer and architect in Arizona and thirteen years teaching computer science at The Citadel in Charleston, South Carolina, before coming to UVU ten years ago. He describes watching generative AI arrive as, in his words, feeling like a kid in a candy store — technology he'd been researching as a graduate student decades ago suddenly becoming a public commodity.
Rudolph's institute project wasn't a new standalone tool so much as a modernization of an existing one: his graduate-level classical AI course, restructured to add an agentic, multi-agent orchestration unit on top of material that already had students building single agents. He designed the addition to be fully optional and separable. His autograder now carries a flag that turns the agentic content on or off, so another instructor could teach the same course without it if they chose. The deeper shift, he said, was in his own workflow rather than the classroom material itself: he used AI agents throughout the summer to update his syllabus, restructure his course schedule, and draft new homework questions, describing his role less as building a student-facing agent and more as using agents as a co-author on the back end of his own course design.
The single idea he said he'll actually carry into every class he teaches this fall didn't come from his own project. It came from a professor in Montana he encountered through the institute's network, who'd built a simple markdown file laying out AI "guardrails" for a course: rules a student pastes into ChatGPT or Copilot at the start of every session to define what's in and out of bounds for that particular class. Rudolph called it an elegant, low-effort protective mechanism and plans to roll a version of it out across his own courses.
Rudolph drew a sharp distinction between how most disciplines should use AI and how his own students need to. For most fields, he said, effective AI use doesn't require understanding the underlying technology. He argued people can operate at whatever level of depth gets them the result they want. Computer science students, in his view, are the exception: they're the ones who will eventually build, maintain, and certify that AI systems and their safeguards actually work as claimed, which means they can't be allowed to simply generate their way through a computer science degree the way a business or English student reasonably can with a paper.
He was candid about the tension that creates against industry expectations, where employers increasingly want new graduates fluent in what he called "vibe coding:" generating and directing AI-written code rather than hand-writing all of it, even as his program insists students understand every line. He described a genuine reversal in his own thinking a few years back, after conversations with employers made clear they wanted graduates skilled specifically at directing AI tools, not avoiding them: a moment that shifted him from viewing AI-assisted work as something close to cheating to treating fluency with it as a legitimate, expected job skill his program has an obligation to teach.
He offered the same practical, low-barrier advice as several other faculty in this series: someone with an afternoon and no background should pick any reasonably well-supported framework, clean up a dataset they care about, try it, and only dig into customization if the out-of-the-box results fall short. He said this advice applies whether the person has thirty years of machine learning background or none. On the institute itself, Rudolph was candid that he didn't come away having learned much he didn't already know technically, given his own research history, but said the value for him was social rather than instructional: watching less technical colleagues validate their own ideas, and seeing how many faculty outside computer science were independently building practical, course-specific agents once given the room to try.
Asked what he wished the interview had covered, Rudolph pivoted to a subject entirely outside this summer's work: he expects quantum computing to become commodity-accessible within five to seven years, on a trajectory he compared to generative AI's own jump to mainstream use around 2021, and predicted that a fusion of quantum computing and AI within that window could dwarf the impact AI is having on its own right now. It is a prediction he pegged at something like 80 to 90 percent confidence rather than certainty. UVU's College Engineering and Technology is already building out quantum computing coursework aimed at that horizon, with an explicit goal of building a quantum-capable workforce in Utah ahead of the shift.
The NVIDIA ambassador who wrote 5,000 lines of rules for his own AI
Majid Memari — known to students and colleagues as MJ — joined UVU in 2023 as the computer science department's first postdoctoral researcher before moving into an assistant professor role, and has worked with the Kahlert Institute for Applied AI for more than two years designing and teaching new AI courses. Since founding Nexus AI Solutions LLC in March 2026, he's also built a formal relationship with Nvidia: after training, an exam, and an interview, he holds status as an Nvidia Ambassador specializing in agentic AI, which gives him access to Nvidia's teaching content and computing resources and lets him run Nvidia-supported workshops at UVU.
His institute project, an "AI fluency module pack" built for CS 4390R, UVU's introduction to large language models course, is closer to a curriculum framework than a single tool: a student AI-use integrity guide spelling out what's allowed and how to disclose AI use, four scaffolding labs covering prompting, verification, source-checking, and code review, a running AI-interaction log where students document how they used AI on an assignment, grading rubrics built specifically for prompt quality and AI-assisted work, a risk checklist, and an instructor guide so other faculty can adopt the whole package rather than reinvent it. He's teaching the course online this fall and is already using the module pack in the current course offering, though he hadn't originally connected that the material had grown out of his institute project until it was pointed out to him during the interview.
Alongside that framework, Memari built a second, more targeted tool: a retrieval-augmented course chatbot trained specifically on his own syllabus, assignments, and due dates. He designed it deliberately to avoid hallucination — if a student's question isn't answered by the course material in its knowledge base, it doesn't guess or reach for an outside source; it tells the student to contact him directly, after a light pre-screening step. The motivation was mundane but real: a large share of student questions are things already answered in the syllabus, and offloading those frees him to spend office hours and email on questions that actually require his judgment. He said the tool also seems to shift what students ask about — toward more substantive questions rather than logistics — though he's planning a formal end-of-course survey to check that impression against real feedback.
Memari's broader research interest, and what he considers the single most valuable application of AI in education, is personalized or adaptive learning — using data on how an individual student engages with material (visual learner versus reader, for instance) to shape how content gets delivered to them specifically, rather than teaching an entire class through one fixed method. He was direct about the obstacle: doing this well requires collecting substantial data about how individual students learn, which raises real privacy and security questions he doesn't consider himself positioned to answer definitively, since his own expertise is machine learning rather than cybersecurity or data governance.
He was equally direct about the limits of giving AI systems autonomy. Describing an incident where he asked an AI coding assistant to "clean up" a code repository and it instead deleted important files, he said the experience pushed him to write out an extensive rule set — he estimated around 5,000 lines — that any AI agent he builds now has to follow, covering what it's permitted to do unsupervised and what it isn't. His broader advice to non-technical colleagues was to be cautious about handing AI systems real autonomy at all until industry-wide AI governance and policy standards mature further, arguing that current models remain weak specifically at ethical judgment and genuine creative originality — a point he tied to academic publishing norms, where he said an idea sourced from an AI model doesn't count as a researcher's own contribution because it's drawing on patterns in existing published work rather than generating anything new.
On academic integrity, Memari's position in his own AI-focused courses is close to the opposite of a ban: he actively tells students to use AI, arguing that in a computer science or AI course specifically, a student without deep programming or networking background can still tackle an ambitious, open-ended real-world project with an AI assistant filling the gaps — and that the value of the assignment is in the process of proposing an idea, implementing it, and evaluating the result, not in whether the student produced every line unaided.
His advice for a colleague with about an hour to try building something depended heavily on their background: for a non-technical faculty member, he'd point them to a high-level consumer tool like ChatGPT or Claude and skip programming entirely; for someone in a technical field, he said the field is still actively debating whether traditional programming fundamentals remain necessary at all, and that he genuinely doesn't know the answer — a rare admission from someone this deep into the subject, and one he tied to a broader point about how difficult it's become for anyone, including university leadership, to plan more than a year or two out with any confidence.
The accounting professor teaching students to manage an AI employee, not do the grunt work themselves
Noah Myers grew up in northern Utah, worked at CPA firms in Salt Lake City and Orem, then went back to grad school before landing his first teaching post at UVU, where he's now an assistant professor of accounting and a CPA with a professional background in IT auditing, systems implementation, and data analytics. He also consults and gives corporate presentations on AI in accounting, fraud detection, and ethics outside his teaching load.
Myers didn't take the path the institute's track officially offered. The default assignment was to build a chatbot; he'd already built that kind of tool before and, after talking it through with institute director Tyler Small, got explicit sign-off to go a different direction. What he built instead was a full curriculum unit — not a piece of software, but a case, a set of instructions, and a teaching structure — designed to put students in front of agentic coding tools like Claude Code and Codex and force them to act as a manager rather than a data worker.
The case itself is a forensic-accounting scenario borrowed from a major public accounting firm: a tire manufacturer fielding warranty claims from wholesalers, some of whom may be submitting claims that don't hold up, buried in hundreds of messy invoices that need to be reconciled against other records by hand. Myers picked it deliberately because it's big and tedious enough that brute-forcing it manually is a real option. One student in his spring pilot did exactly that, working through every line item herself rather than delegating it, which he said proved the point as much as an efficient run would have. The actual assignment isn't to solve the case; it's to direct an AI agent to solve it, meaning breaking the problem into steps, instructing the agent on how to process the source documents, reconcile the datasets against each other, and flag discrepancies, then reviewing and correcting its output the way a manager would review a new hire's work.
The framing he uses with students is pointed: entry-level "staff" work, including the manual, repetitive reconciliation that used to be a junior accountant's first few years on the job, is increasingly done by an AI agent now, and firms are telling him they want graduates who can walk in ready to operate at the next level up, reviewing and directing that work rather than doing it by hand. He draws a hard line for accounting specifically: unlike a general chatbot, which gives a different answer every time you ask, agentic coding tools can be directed to build a fixed, repeatable process, something he calls a genuinely deterministic, auditable workflow, which matters in a field built on reconciliation and audit trails in a way it might not in other disciplines.
Myers traced the idea back further than the institute itself, to a week in January when he was out sick and, with unplanned free time, sat down to try one of these coding agents on a personal project. It is a website concept he'd wanted to build for years but had stalled on for lack of time to gather the underlying data. Within two hours, he said, he had a working front end and the data itself, scraped and categorized automatically. That experience, more than anything at the institute, was what convinced him the technology had crossed a real threshold rather than being another overhyped tool.
What the institute specifically provided, in his account, wasn't the idea but the space to execute it: dedicated time away from a full course load, plus a room of colleagues comparing notes on what they were each discovering the tools could do, a cohort effect he credited as much as any formal mentorship, including learning about internal UVU resources he hadn't known were available. He's kept his lower-level accounting information systems course largely unchanged, on the theory that the foundational material, such as understanding databases, understanding where a number actually comes from, still has to be taught the traditional way before a student can meaningfully supervise an AI agent doing the same work; the new curriculum lives specifically in his upper-level elective, which he treats as a sandbox.
His advice to a colleague with an afternoon to try this: treat it exactly like delegating to a new hire. Gather the relevant files into one folder, write out instructions as if you were handing the project to a person. The goal, the raw material, the steps all point one of the free agentic coding tools (he named Claude Code, Codex, and Cowork specifically) at that folder, and let it work, then check what comes back the way you'd check any new employee's first assignment.
The sociologist making room for the students who won't touch AI at all
Yi Yin is an assistant professor of sociology in UVU's Department of Behavioral Science, with a Ph.D. in sociology from the University of Wisconsin-Milwaukee completed in June 2022. Her research sits specifically at the intersection of sociology and AI: one track studies AI's role in health informatics: whether tools can help someone without medical training reliably judge the accuracy of health information they find online. This role is built in collaboration with scholars in health informatics and machine learning, and a second track has her leading a UNESCO-affiliated project examining how faculty, staff, and students at partner universities in the Philippines experience and adapt to AI differently depending on the resources and constraints of their institutions. As a woman from Bouyei, an ethnic minority community mostly living in province of Guizhou in Southern China, she's also spoken about wanting her work to widen access to education and opportunity for underrepresented students specifically.
Her starting premise for the institute was that AI in the classroom isn't a computer science problem to hand off to computer science departments. She wanted her sociology students to see their own discipline as directly relevant to understanding AI's social effects, not adjacent to it. What she built is an AI-assisted literature-review tool for her classical social theory course, a class many students arrive at already dreading as a dry degree requirement. She uses AI on the instructor side to translate dense theoretical material into more accessible language for her course materials and to design activities that connect abstract theory to present-day scenarios, and on the student side, she built an assignment around research tool Research Rabbit chosen specifically because, unlike a general chatbot, it surfaces real, checkable sources rather than citations a model might fabricate. Students run the same research question through both a traditional method (library databases, Google Scholar) and the AI-assisted tools, then compare what each approach turned up and where each one's limitations showed.
Yin was candid that most of her students walk in with close to no prior AI experience, which shapes how she sequences the assignment: before any research tool, students start with basic prompt-writing practice, learning that an imprecise prompt gets an imprecise result, and only build toward more complex AI-assisted research once that foundation is in place. She argued against assuming any classroom is full of confident AI users by default, and said she runs an end-of-semester survey specifically to find out what exposure, or lack of it, students are getting from other instructors, so she can offer extra support to students who arrive behind.
The part of her interview she returned to more than once was what she does with the small minority of students. By her estimate, one or two out of thirty in a given class object to AI use outright, often on ethical grounds. Rather than requiring participation, she's built an alternative assignment path: students who opt out research AI's ethical, privacy, or environmental costs directly and reflect critically on the technology instead of using it. She was pointed about one pattern she's noticed: the students who push back hardest aren't the ones with the least AI exposure. In her experience, it's often the opposite, students who've used it enough to have specific, informed reservations. Her broader point was that an educator's job isn't to produce uniform enthusiasm for AI, but to make room for competing, well-reasoned positions on it, including the position that a given student wants no part of it.
On the institute itself, Yin said the most useful part wasn't technical polish but the exposure to how differently colleagues in other departments approached the same assignment. She said one-on-one AI coaching sessions and access to the AI Gateway helped her sharpen her own agent. However, sitting with faculty from the College of Engineering and Technology whose tools she described as noticeably more technically sophisticated than her own — and from the humanities — who she said were considerably more skeptical, raising concerns about accuracy, ethics, privacy, and AI's environmental footprint — gave her a wider view of the range of legitimate faculty positions on AI than she'd had walking in.
Her advice to a colleague with an afternoon to spare was to start with UVU's own campus resources and workshops rather than going it alone, then begin with prompt-writing fundamentals and one small, low-stakes classroom activity before attempting anything as ambitious as building a full agent. Asked what she wished the interview had covered, she turned the question back on the series itself: she suggested the Institute ask every other faculty member in the project the same question — how they handle a student who refuses to use AI at all — arguing it's a more revealing question about a classroom's culture than what any individual tool can do.
Closer: what a summer of restraint adds up to
Six instructors, six disciplines, and not one of them set out to make AI smarter this summer. They set out to make it more disciplined.
That's the throughline running under every profile in this installment. Carter's tutor won't draft a sentence. Memari's course chatbot is built to refuse rather than guess. Rudolph is rolling out a colleague's "guardrails" file across his own courses specifically to put a leash on what a student's AI assistant is allowed to do unsupervised. None of these instructors needed a more capable model — the models were already more capable than the assignment required. What they spent May building was restraint: tools engineered to hold something back, on the theory that a fast, complete answer is often the thing standing between a student and actually learning.
The second thing that came into focus across all six conversations is what these faculty think they're actually preparing students for. It isn't fluency with a chatbot — that bar is already cleared for most incoming students, whether an instructor teaches to it or not. It's judgment over an AI's output, and increasingly, the ability to manage AI the way you'd manage a junior hire. Myers says it outright: entry-level "staff" work in accounting is now an agent's job, and firms want graduates who can walk in ready to supervise it. Alvarado-Karste reaches for a music metaphor to say the same thing — AI has handed everyone a fully capable instrument, so the only differentiator left is who can direct the whole performance. Rudolph built a semester's worth of coursework around exactly that shift, from single-agent code to multi-agent orchestration. A marketing professor, an accounting professor, and a computer science department chair arrived at the same conclusion from three different directions: the scarce skill now is knowing what to do with an answer, not producing one.
What made that possible, almost every subject agreed, wasn't the institute's curriculum so much as the institute's shape — two protected weeks and a room full of colleagues comparing notes. Rudolph said flatly he didn't learn much technically he didn't already know; the value was sitting next to people working the same problem. Myers, Memari, and Yin each described some version of the same thing: time away from a normal course load, plus a cohort, did more than any formal session. Several of the sharpest ideas in this batch of profiles — Myers's whole curriculum, Rudolph's borrowed guardrails file, Alvarado-Karste's pivot away from a stalled software build — trace back to a conversation with another faculty member as much as anything KAAII taught directly. The institute, on this evidence, functioned less as a training program and more as a deadline with company attached: the thing that turned half-formed personal experiments, some of them dating back months before May, into something ready for a syllabus by fall.
And running quietly under several of these tools is a problem none of these instructors named directly but all of them are solving for: they can't be everywhere a student needs them. Carter's tutor exists for the student stuck at 3 a.m. Memari's chatbot exists for the gap between office hours.
The question this installment didn't get to ask directly, but that several subjects answered anyway, came from Yi Yin: how do you handle a student who wants nothing to do with AI at all? Carter builds an opt-out into her syllabus. AlSobeh treats a student's ethical objection as a legitimate position, not a problem to correct. Yin has a full alternate assignment ready for exactly that student. None of them treated resistance as something to argue a student out of — which may be the most consistent value in this entire cohort: an insistence that using AI well and refusing it outright are both defensible positions a student is allowed to hold, as long as they can explain why.
That question — and the rest of UVU's 37-person cohort — carries into the next installment of this series.
Additional profiles from this series will be published on the Kahlert Institute for Applied AI's website as they're completed.