Part II. How Preply uses AI to give you personalized lesson insights
In our first chapter, we explored how AI-powered learning took its first steps at Preply. Launching a feature is a critical milestone, but the journey doesn’t end there—it’s equally essential to iterate, refine, and scale it to maximize its impact on learners and tutors.
Today, Emily Stott, Senior Product Manager, takes us behind the scenes of lesson insights — how it’s evolved, what surprised us, and where AI-driven personalization is headed next. Let’s dive in!
Emily, congrats on the launch! Give us the lowdown on lesson insights — what’s the thinking behind it, and how did we make sure it really delivers for our learners and tutors?
Lesson insights is your smart personal lesson recap. For learners, it answers the question: “How am I doing? What should I focus on next?” It’s all about keeping progress front and center and turning lessons into tangible growth. Think of it like a personalized report card that gives you key takeaways after every class, so you always know where you stand and what to work on next. It includes:
✅ A quick summary — the key takeaways from the lesson.
✅ Personalized feedback — what you nailed and what needs work.
✅ Clear next steps — personalized activities based on your lesson.
We started with a simple goal: to give learners and tutors the tools to track and celebrate progress throughout the learning journey. The advancements in AI and machine learning allowed us to turn this into something powerful — lesson insights.
For learners, it’s about motivation and clarity: no more wondering, “Am I getting better?” — now you’ll see exactly where you’re improving and where to focus next. For tutors, it’s about an easy way to track their students’ progress, as well as less admin and more impact: with lesson insights, they can focus on teaching rather than taking notes.
One of the biggest challenges in designing this feature was making sure it’s safe and doesn’t create any risks for our learners and tutors. AI is evolving fast, and while more people are starting to understand it, it’s still early days — so getting the user experience right is key. We wanted to make sure AI feels helpful, not intimidating. Our team worked hard to minimize risks, ensure responsible outputs, and stay transparent about data usage so learners and tutors always feel in control. And of course, we’re constantly improving and fine-tuning these safety measures with every new version.
Steven Lenoir da Silva, Design Lead
What key data or feedback did you look at?
At Preply, we built lesson insights iteratively, constantly learning and improving along the way. The first version, launched last year, was an extremely exciting launch — all about setting up the foundation: the right infrastructure, the initial value, and getting our first wave of feedback.
From there, we went deep into user insights — tracking usage data, running interviews, and collecting survey responses. One thing became clear: learners wanted deeper feedback in the areas that mattered most to them.
So, we structured insights around three key pillars:
🔹 Vocabulary — What complexity and variety of words are they using? What are the new words they’ve learned in the lesson?
🔹 Grammar — What common mistakes are they making? How can they improve?
🔹 Speaking — How fast are they speaking? How is their pronunciation?
With this, learners get personalized, actionable insights — things like synonyms to try and grammar corrections, all with real examples from their own lessons.
Since launching the second iteration, we’ve been making changes every week, leveraging continuous user research and usage data so we can understand what users love and where we can make things even better.
Were there any surprising insights from user data or feedback that significantly influenced the product’s direction?
One of the biggest surprises? Learners really love seeing their mistakes!!
We wondered if people might focus on their strengths, but we learned the real value comes from detailed, specific feedback on where they can improve. Seeing tangible examples of their mistakes and how to fix them is one of the most valued aspects of lesson insights. That insight led us to double down on detailed, personalized corrections and start developing custom follow-up activities based on individual mistakes.
More than just the technical challenges, what really excites us is the real impact lesson insights has on learners. From the very start, we made user feedback a priority — running studies throughout development and even setting up a Slack channel to track insights in real time. This helped us tweak and improve things fast, and honestly, seeing direct feedback from learners kept the whole team motivated. On the business side, engagement has been strong, with learners using Lesson Insights to track progress with tutors or guide their own study sessions. But the best part? Hearing users say it helps them ‘navigate in the dark’ — that’s exactly why we keep pushing to make it even better.
Anton Vitrenko, Senior Full Stack Engineer
Here’s what learners mentioned:
💡 They love reviewing mistakes.
“Being able to review my mistakes and successes immediately after class has been extraordinary for improving my language.”
🔥 They find it super motivating.
“Every comment helps me to try harder.”
“It keeps me going — it’s really motivating for me.”
✅ They love how actionable it is.
“Gave me clear and actionable steps to improve myself.”
“It’s amazing! No human being would be able to make a real-time summary like this. I read it every time a lesson ends.”
That’s exactly why we’re focusing on making Lesson Insights even more interactive — so learners don’t just see their mistakes, but have personalized exercises to keep improving based on them.
How did advancements in AI contribute to the refinements made in lesson insights?
Advancements in AI, especially in large language models (LLMs), have been a game-changer for lesson insights. These models, like those from OpenAI, are the backbone of how we generate insights, and the leaps in quality over the past year have been incredible. They’ve enabled us to deliver richer, more accurate, and more personalized insights to our learners.
But it’s not just about using the best models, it’s about how we work with them. We’ve gained valuable learnings on evaluating the quality of LLM outputs and refining how we integrate them into our systems. By collaborating with top providers and applying best practices, we’ve optimized our processes to deliver even more impactful results for our learners and tutors.
Instead of general takeaways, lesson insights is fully based on your personal lesson — your progress, your strengths, and your areas for improvement. Here’s how AI makes it smarter:
- Personalized feedback — lesson insights generate grammar corrections, key vocab, and synonyms tailored for you
- Personalized practice — Based on your lesson, AI generates custom exercises for self-learning, so you get the most relevant practice to keep improving.
What were the biggest challenges during the iteration of Lesson Insights?
Ensuring the quality of the insights. It’s not just about generating feedback — it has to be accurate, relevant, and genuinely valuable to learners.
Take grammar feedback, for example. In the early days, we relied on manual reviews to evaluate prompt outputs: linguists would go through hundreds of insights, giving them a thumbs-up or thumbs-down. Helpful? Yes. Scalable? Not at all.
Now, we’ve come to more robust and objective processes:
- Benchmarking against labeled datasets — Ensuring AI-generated insights match real expert-reviewed data.
- LLMs as a judge evaluation – We make AI check its own outputs with simple evaluator prompts before it’s shown to the user.
How did you determine that lesson insights was providing value for learners and tutors?
It’s still early days, and we’re continuously refining Lesson Insights based on user feedback. But the signals so far have been very promising.
- Retention is strong. We’re seeing high activation and sustained engagement, with a smiling retention curve to back it up.
- Users really don’t want to lose it. Aside from positive feedback, we ran the classic Product-Market Fit (PMF) survey, asking learners: “How disappointed would you be if you could no longer use lesson insights?” Even with lesson insights live for just a few months, 66% said they’d be very disappointed 🎉 We see this as a clear sign that we’re solving a real problem for our users.
And we’re just getting started!!
What cross-team collaborations were essential to the success of Lesson Insights?
The success of Lesson Insights was powered by collaboration across many teams at Preply. We worked based on the strong foundation of the first launch, and collaborated closely with the “Self-Learning team” to ensure learners received relevant and personalized exercises, keeping engagement and progress at the core. We also engaged with legal and security teams to make sure we’re not just compliant, but ahead of the curve on best practices. And because tutors play such a key role in delivering impactful learning experiences, we collaborated with the Tutor Supply team to keep them informed and aligned with these updates (as well, of course, with tutors directly to help us shape the best version of the product!).
One of the most powerful examples of collaboration happened at the start of the project. After gathering initial user feedback and testing concepts, we had clear signals about what needed to be done, and the team came together like an orchestra — each member playing their unique part in harmony. We shared a vivid vision of how the feature should work and look, and everyone worked purposefully toward that common goal. It was amazing to see how each craft within the team contributed their expertise to solve this challenge.
Steven Lenoir da Silva, Design Lead
The future of personalized learning: where is Preply headed?
The future of learning is deeply personal, highly efficient, and powered by AI — without losing the human touch.
Tutors will always be at the heart of it. The connection, motivation, and expertise they provide is irreplaceable. But AI will enhance their capabilities, handling repetitive tasks so they can focus on what they do best: teaching. For tutors: AI will become their teaching assistant, freeing up time and providing data-driven insights to make every lesson more impactful.
For learners: the perfect balance — expert guidance from tutors + AI-powered self-study. With smarter feedback and personalized exercises, learners will be able to progress faster than ever.
We are actively seeking talented candidates to help us build AI-powered learning that enables people’s progress 🚀🚀 If you’re excited by this challenge apply to one of our open positions!
