Kajabi Will Rent You an AI Teaching Assistant for $79 a Month. The First Big Trial of One Just Came Back With Lower Grades.

The pitch is good. You've built the course, you're answering the same four questions in the comments every week, and Kajabi will now sell you something that answers them at 2am instead.

A Teaching Assistant is the first type of Expert Agent, an AI chat that lives inside a Product and answers member questions using the Lessons, transcripts and Community posts they already have access to. It speaks in the third person about your expertise, never claims to be you, and refers members back to you for deeper engagement.

The price is the part everyone's been staring at. Expert Agents are priced per agent, per month, with 1,000 messages and 250 AI credits included. The standard rate from October 1, 2026 is $79 per agent per month. Agents bought between August 1 and September 30 locked in $49 for twelve months. Want the Sales Agent too? That's $158 a month for the pair.

Then, on September 30, a paper landed that nobody selling AI tutors wanted to read.

2,379 students, 30 instructors, one randomized trial

"The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial," by Jing Liu, Tracy Sweet, Marty Haoyuan Chen and colleagues including Jennifer King Rice, went up through Brown University's Annenberg Institute in October as EdWorkingPaper 26-1598.

It evaluates access to a course-integrated generative AI tutor in a randomized controlled trial involving 2,379 undergraduates and 30 instructors across multiple disciplines at a large U.S. public university. The trial ran in fall 2025; instructors were randomly assigned a GPT-4o tutor built into the learning platform and drawing on course materials they selected.

Not a side experiment. The tool was built into the course environment, available to students as a standard part of their work. Which is, structurally, what Kajabi is selling.

The result: among sections of the same course, tutor access reduced final grades by 0.37 standard deviations and learning management system participation by 0.90 standard deviations.

Break the engagement number apart and it gets worse. In the exact-match sample, AI tutor access reduced course page views by 0.37 SDs, active days by 0.51 SDs, and recorded participation in instructor-designed online activities by 0.90 SDs. Students with access visited course pages less, logged fewer active days, and completed fewer discussion responses, assignments and quizzes.

The sentence course creators should read twice

Here is the mechanism, straight from the paper: compared with nonusers in treatment classrooms, AI tutor users were more likely to report asking instructors fewer questions about course content and sending the instructor fewer emails or messages.

The authors' own framing is that an AI tutor "may displace engagement with existing learning activities without improving academic performance."

Displace. Not supplement.

If you sell a $997 program, you already know that the students who finish are the students who talk to you. They post in the community, they show up to the call, they email the question they're slightly embarrassed to ask. That contact is not overhead. It is the completion mechanism, and completion is what produces testimonials, renewals and the referral engine underneath an evergreen funnel.

A bot that absorbs the embarrassing question before it reaches you is removing the exact friction that was doing the work.

Where the comparison breaks down

Be honest about the gaps, because they're real.

This is a working paper, not a peer-reviewed journal article. Treatment was assigned at the instructor level, so inference relies on a relatively small number of clusters — the authors flag it themselves. One read of the results notes the 0.37 SD drop holds among sections of the same course, with no significant grade effect in the full sample. It is one trial, of one tool, at one institution, and a differently designed tutor introduced differently might produce different results.

Uptake was also thin. A breakdown of the study reports only about 15% of offered students used the tutor, nearly three quarters of requests asked for direct answers, and just 0.7% asked for feedback on their own work. Most instructors kept the default direct-instruction mode.

A bored sophomore in a required lecture is not a buyer who paid four figures of their own money at 11pm because they want to change careers. Motivation changes everything, and your students chose to be there.

And AI tutoring has won before. At Harvard, Greg Kestin, Kelly Miller and colleagues ran a randomized trial in fall 2023 with 194 students in an introductory physics course; the AI lessons won, and students needed less time, a median of 49 minutes against 60 in class. What's new here is a large, well-powered randomized test of the product category that has been sold hardest.

If you're buying one anyway

Treat it as an experiment with a control group, not a feature you switch on across the whole catalog.

Kajabi gives you the instrumentation. Each agent's page counts total messages, active conversations, unique people and the average messages per conversation. Those are usage metrics. They will go up. They are not the metric.

The metric is module completion, community posts, live call attendance and refund rate, measured against a cohort that never had the agent. If messages to the bot rise while posts in your community fall, you have bought the Maryland result.

Know the fence too. Expert Agents run only on Kajabi-hosted pages, one site per agent. They're placed through the Theme Editor as a floating button or embedded block, and they're text-only at launch. The 1,000-message cap is generous for a small course and tight the week after a launch.

Kajabi's own documentation already contains the correct instruction, almost by accident. The agent is supposed to refer members back to you for deeper engagement. The trial suggests that when a bot is sitting in the course, a meaningful number of students never make that trip.