When Dr. April Crenshaw realized her online precalculus students needed more support, she didn't turn to an off-the-shelf AI tool,she built one from scratch with help from a student in her class. The result offers a rare glimpse into what happens when educators and learners collaborate on AI education technology, and the lessons they discovered challenge common assumptions about how AI tutors should work.

Why Did a Community College Professor Start Building AI Tools?

In summer 2025, Crenshaw faced a familiar problem at Chattanooga State Community College. Her mathematics department had received a grant to reimagine support systems for struggling students, but they had also eliminated a bridge course that previously helped students strengthen their algebra skills before tackling precalculus. Crenshaw worried about the impact on her online learners especially.

Most of her students juggle work and family responsibilities alongside their studies. This reality made it nearly impossible for them to attend traditional office hours or visit the math center during regular operating hours. Crenshaw understood something crucial: the students who needed the most help often struggled most with asking for it. "It's easier to reach out to somebody you're seeing in person four times a week," she explained. She hoped an AI tutor could offer a lower-stakes way for students to get help exactly when they needed it.

What Went Wrong When She First Tried to Build It?

Crenshaw began experimenting with ChatGPT, attempting to customize it as a virtual learning assistant. But she hit a critical roadblock: the tool behaved inconsistently. It would work perfectly on one math problem, then fail on the next one, even when the problems were nearly identical. "One of the worst things I could do was to present this new tool to students and then have it give the wrong answer," Crenshaw noted.

She tried rewriting prompts repeatedly, but the real issue was architectural. That's when David Escalante Gonzalez, an electrical engineering student in her class, stayed after class to ask about her project. After hearing about her struggles, he offered to help. Together, they realized the AI was getting lost in multiple overlapping sets of instructions. Instead of dumping all instructions at once, they needed to build a modular system that routed student requests into different instructional pathways depending on what the student needed.

How Did They Solve the Consistency Problem?

Escalante suggested moving the work off ChatGPT's interface and instead using OpenAI's API (Application Programming Interface), a more flexible tool for building custom applications. Then the pair did something surprisingly low-tech: they grabbed pencil and paper and mapped out exactly what they were trying to accomplish. They designed a foundational layer of instructions that could route student requests into different modes based on the conversation flow.

The result was a system with three distinct pathways, each offering different levels of support:

• Step-by-Step Mode: The default setting where the learning assistant breaks problems into steps and directs students to work through them, providing formative feedback along the way.
• Quick Hints Mode: Offers short prompts for the student's next action, useful for learners who need minimal guidance.
• Detailed Explanations Mode: Provides the reasoning behind each step, including conceptual rationale and regular comprehension checks throughout the solution process.

Crenshaw also paid careful attention to tone. She didn't want students to feel judged or talked down to, so she intentionally wrote warmth and respect into the AI's responses. Many of her students arrived already feeling discouraged about their abilities. "We want to give them encouragement and big wins early," she explained.

How to Build Trust in an AI Tutoring System

• Extensive Testing: Crenshaw and Escalante tested and evaluated their prompts extensively until they were confident the learning assistant could consistently provide either correct tutoring support or accurate answers.
• Accuracy Improvements Over Time: They waited for OpenAI to release updated models of ChatGPT, which showed improved accuracy, especially for mathematical responses, before deploying to students.
• Transparent Communication: Escalante, who had been using AI tools since ChatGPT's release, was familiar with hallucinations and the tendency of AI to generate incorrect information. This awareness shaped how they built safeguards into the system.
• Emotional Design: Building warmth and respect into the AI's tone helped students feel supported rather than judged, addressing the emotional barriers that often prevent struggling learners from seeking help.

"I had been using AI tools since ChatGPT came out, so I was very familiar with hallucinations and aware that they were prone to giving you the wrong answer," said Escalante.

David Escalante Gonzalez, Electrical Engineering Student at Chattanooga State Community College

What Were the Real-World Results?

Crenshaw piloted the AI learning assistant first in Escalante's class, then expanded it to her online students in subsequent semesters. The adoption rates were striking. National survey data typically shows that about one in eight students use tutoring and academic support services. In Crenshaw's spring 2026 pilot, 91 percent of students used the learning assistant, and 85 percent of those students returned for at least one additional session.

These numbers suggest something important: when AI tutoring is designed thoughtfully with student needs in mind, adoption looks dramatically different from traditional tutoring services. The on-demand nature, the lack of social friction, and the ability to access help at any hour appear to have removed significant barriers to seeking support.

What Does This Mean for Other Educators?

Crenshaw doesn't believe other instructors need to build their own AI tutoring systems from scratch. Her experience building one taught her what actually happens behind the scenes and which features genuinely help students learn. That knowledge made her a more informed consumer of AI tools. She was later able to ask publishers more pointed, technical questions when piloting AI tools for Chattanooga State's largest statistics course.

Crenshaw and Escalante, along with colleague LeAnders Burns, eventually published their approach in a contributed chapter in the 2026 open-license book "The Pedagogical Promptbook: Enacting Evidence-Based Teaching and Instructional Design Practices with Generative AI." The chapter includes a template that other instructors can adapt for their own courses, making their hard-won insights available to the broader education community.

This story arrives at a moment when AI in education remains deeply polarized. While some school districts have implemented broad bans on student use of generative AI tools, citing concerns about academic integrity and skill development, higher education institutions are generally taking a different approach. Universities tend to view AI as an inevitable and transformative force, focusing instead on teaching students how to use it ethically and effectively.

The Chattanooga State collaboration suggests a middle path: AI tools designed with pedagogical rigor, tested for reliability, and built with student psychology in mind can achieve remarkable adoption and engagement. The key isn't whether to use AI in education, but how to design it in ways that support learning rather than shortcut it.