The pitch is seductive
Preply just shipped AI-generated lesson summaries for language learners, and the retention numbers look fantastic: 75% of active learners still engaging with the feature more than a year after launch, 4.7/5 satisfaction from 300k+ ratings, and 70% of tutors using it regularly.
On paper, this is the perfect AI augmentation story. Tutors teach, OpenAI's API watches the transcript, and learners get personalized grammar corrections, vocab highlights, and pronunciation feedback within minutes of their lesson ending. No one gets replaced, everyone gets better.
But the more interesting question isn't whether this works—it clearly does—but what kind of work it's actually doing, and what assumptions about learning it's quietly encoding into production.
The automation target is telling
Preply's CTO Dmytro Voloshyn frames the opportunity clearly: tutors provide "irreplaceable energy, motivation, cultural nuance" while facing "repetitive tasks: writing personalized plans and lesson notes." The AI handles the repetitive stuff. Classic augmentation narrative.
Except lesson notes aren't purely administrative. Writing a summary forces the tutor to synthesize what happened, identify what mattered, and decide what comes next. That cognitive work—the act of reflecting and prioritizing—is part of how good tutors get better at teaching.
When you automate that loop, you're not just saving time. You're moving the locus of pedagogical judgment from the human to the model. The tutor becomes a performer; the AI becomes the analyst.
Maybe that's fine! Tutors on a marketplace platform are optimizing for volume and consistency, not deep reflection on every single lesson. The question is whether Preply's learners are getting a tutor who deeply understands their progress, or a tutor plus an OpenAI layer that understands it better.
Personalization as intermediation
The bigger architectural move here is turning transcripts into personalized practice exercises. After the lesson, the AI generates homework based on "exactly what you've been talking about, your goals, the topics you've been covering, your tutor's feedback." This sounds like personalization, and it is—but it's personalization mediated entirely by the model's interpretation of the conversation.
There's no indication that tutors review or edit these exercises before they're assigned. The loop is: lesson happens → transcript → API call → homework appears. The tutor's role is to deliver the lesson; OpenAI decides what to practice next.
This is efficient, and the retention numbers suggest learners don't mind. But it's worth naming what's happening: the human teacher is being disintermediated from curriculum design at the micro level. The model is now the instructional designer.
For a marketplace with 100,000 tutors across 90+ languages, maybe that's the only way to deliver consistency at scale. But it's a trade. You gain uniform quality and automation; you lose the idiosyncratic judgment of individual teachers who might notice things the transcript doesn't capture.
The novelty test they didn't need to run
One genuinely interesting data point: 75% of active learners are still engaging with Lesson Insights more than a year later. Emily Stott, the product lead, frames this as proof it's not just novelty: "if people are coming back and engaging months later, time and time again, that's a very strong signal of value."
She's right. A year is long enough to wash out curiosity effects. But it also raises the question: what would disengagement look like in a system where the AI feedback is automatically inserted into your post-lesson workflow?
If you're a committed language learner paying for 1-on-1 tutoring, and every lesson comes bundled with a structured summary and personalized exercises, opting out requires active effort. Engagement might reflect genuine value, or it might reflect that the feature is now load-bearing infrastructure for the learning experience.
The 4.7/5 satisfaction rating is harder to dismiss—that's direct user feedback, not just usage telemetry. But satisfaction with AI-generated content often reflects "good enough" rather than "better than the alternative I never tried."
The internal adoption story is more revealing
The case study spends a lot of time on Preply's internal ChatGPT Enterprise rollout: 95% weekly active usage among 600+ employees, company-wide enablement sessions, custom GPTs for brand voice.
This is where the story gets honest about what AI adoption actually requires. It's not just dropping a new tool into Slack. Preply made it "a company priority," reflected it in "strategy, roadmaps, and objectives," and ran structured training. Translation: leadership mandated it, and teams complied.
The result is impressive—95% WAU is genuinely high—but it's also a reminder that enterprise AI adoption is as much about organizational pressure as product-market fit. When your CTO says AI is "at the center of how we operate as a company," using it becomes an expectation, not a choice.
The Codex adoption numbers follow the same pattern: 94% of engineers using AI coding assistants. That's great for velocity, and Dmytro is clearly excited about engineers focusing "more on architecture" instead of "routine coding tasks." But it also means nearly every engineer at Preply is now dependent on OpenAI's infrastructure for daily work.
That's a bet, and it's probably the right one. But it's a bet.
What they're building toward
The roadmap section is where the quiet ambition shows up. Preply wants to build "a deeper understanding of every learner over time—capturing goals, progress, strengths, challenges, and learning preferences across months of study."
This is the logical endpoint: a longitudinal learner model that tracks everything, understands everything, and continuously adapts the curriculum. The tutor becomes one input among many.
In some ways, this is exciting. Genuinely personalized education at scale has been a fantasy for decades. If OpenAI's APIs can make it real, that's a win.
But it also locks Preply—and its learners—into a future where the AI layer is mandatory. You can't deliver hyper-personalized, continuously adaptive learning without a model that watches every lesson, processes every transcript, and adjusts every exercise. The human tutor is necessary but no longer sufficient.
Maybe that's fine. Maybe the future of education is human-led, AI-enabled, just like Dmytro says. But it's worth asking: when does "enabled" become "dependent," and when does "augmentation" become "substitution with extra steps"?
The critique isn't that it doesn't work
Lesson Insights clearly works. The numbers are real, the use case makes sense, and learners seem happy. Preply found a high-value automation target, shipped it well, and is iterating toward something bigger.
The critique is that we should be honest about the trade-offs. AI-mediated personalization is not the same as human personalization. Automating lesson summaries is not just saving tutor time—it's shifting who decides what matters. And building a system where the AI "understands exactly what you've been talking about" means building a system where the AI, not the tutor, holds the canonical model of the learner's progress.
These might all be good trades! But they are trades, and the case study doesn't really grapple with them. It's all upside, all efficiency, all augmentation.
The reality is messier. Preply is building a future where OpenAI's models are load-bearing infrastructure for learning, and where tutors are increasingly performers in a system designed and optimized by AI. That future might be better than the alternative. But let's at least name what we're building.