​The Pencil And The Agent: How AI Can Be Designed Into The Classroom, Not Banned Out

Dr. Terry Oroszi, Vice Chair and Associate Professor, Boonshoft School of Medicine, Wright State University.

getty

In the 1980s, math departments banned calculators, fearing arithmetic fluency would erode. In the 1990s, faculty banned laptops from seminars, fearing typing would replace listening. Today, phones are often banned outright because generative AI is one tap away; three decades, three tools and a largely unchanged argument.

Each ban was a policy response to a design problem, and this time the people best positioned to solve it are not faculty senates but the technology leaders who decide what gets built, procured and deployed.

The Permission Trap

Generative AI does not live inside an institution’s walls; it lives in a pocket. A ban regulates what a student does while a professor is watching and has no jurisdiction over the other 23 hours when the assignment actually gets written. An unenforceable rule becomes a fiction everyone quietly agrees not to examine.

That is an opening for anyone building or selling technology in education. Institutions that cannot legislate a tool out of existence need a better tool instead. Faculty gradually shifted learning outside the classroom by favoring take-home essays finished off campus because they were easier to grade, which is a format built for a pre-GenAI world. No policy memo fixes that; redesigning the room does. That is a technology problem, not just a pedagogical one.

Two Rooms Worth Building

The first room requires very little technology. Any assessment meant to verify unassisted skill belongs in an environment where GenAI cannot be used: in person, on paper or on a locked-down device, during a block of time the institution controls. The concept is straightforward, but the logistics are not. Proctoring hundreds of students requires space, staffing and scheduling, making it a resource-intensive solution rather than a simple fallback.​​

The second room is the one I find more interesting. It is where AI stops being a threat to manage and becomes infrastructure worth building, and it's a model I'm pitching this semester: a classroom lined with laptops, each running a subject-matter AI agent, guardrailed against hallucination and scoped to a specific expertise. Students get learning objectives, not answers, and move station to station, questioning each agent to extract what they need.

For example, this might include a period-accuracy toggle that lets a student hear a primary source voiced in period language, then compare it to a modern translation. A biology class might rotate through a genetics station where the Rosalind Franklin agent will not reveal the double-helix structure until the student reasons through the diffraction patterns themselves. Nobody gets a finished answer; satisfying the objective means knowing what to ask, recognizing a weak answer and pushing further, the skill students need the moment no professor is checking their work.

What This Asks Of The People Building The Tools

​Faculty cannot build this room alone. Turning it into reality is a product design challenge, placing technology leaders at the center of the solution. Building AI that teaches without simply completing the work is exactly the kind of educational infrastructure problem the industry should be solving.​

Three design choices determine whether it works:

• Scope: An agent that answers anything is a liability; one scoped to a single expertise is an asset.

• Restraint: Withholding an answer until reasoning is demonstrated is a harder build than maximizing helpfulness.

• Auditability: Every session should leave a transcript by default. This is the feature that turns a chatbot into a tool an institution can trust, and vendors should treat it as the specification that they once treated FERPA compliance.

Grading The Second Room

Grading here does not require new detection tools. Every conversation already creates a transcript that may be more valuable than a polished final answer because it reveals how a student approached the problem, what questions they asked and whether they challenged weak responses rather than simply accepting them. The objective sheet becomes the rubric. Add a two-minute debrief defending one decision, and this room becomes harder to fake than the take-home essay ever was because the focus shifts from the final artifact to the reasoning process behind it.​​

Where This Gets Hard

Three obstacles stand between this model and a real deployment, and any technology leader here should weigh all three before pitching it as a solution.

1. Scale

​A faculty member can review transcripts for a seminar of 15, but not a lecture of 300. Scaling requires AI tools that can pre-screen transcripts, which creates a new challenge: AI is now evaluating a student's interaction with AI. If that grading layer mistakes genuine inquiry for evasion (or fails to catch the same prompt-injection techniques the classroom agent missed), the assessment breaks down for the very reason it was designed to prevent.​

2. Procurement

This model assumes technology leaders can deliver a custom, guardrailed agent into a classroom within a semester. In practice, institutional procurement rarely moves at that speed. Security reviews, accessibility requirements and budget cycles can stretch implementation timelines significantly. A pilot within one department is realistic; a campus-wide rollout on the first attempt is not.​

3. Guardrails

An agent instructed to withhold an answer is not immune to a student who tells it to ignore its instructions and answer anyway. Prompt injection is a real, unresolved vulnerability, and any vendor claiming a “restrained” agent is bulletproof is overselling. These guardrails need continuous testing, not a one-time setting trusted forever.

None of this makes the second room a bad idea. It makes it a harder engineering problem than the classroom pitch alone suggests.

Owning The Design Choice

Faculty often describe the AI cheating problem as something happening to them. In large part, academia built it by defaulting to formats GenAI is now best suited to defeat. The institutions that get this right will be the ones whose ed-tech partners treated grading infrastructure and prompt-injection resistance as real engineering problems and built tools that reward verification over performance.

The question was never whether to allow the tool into students’ pockets; that was answered the moment smartphones existed. The real question is which room a piece of learning belongs in, the pencil or the guardrailed agents, and whether the people able to build both are willing to do it on purpose.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?