An AI tutor that meets students where they are
Answers questions about course material, explains concepts at the right level, adapts to learning context. For edtech platforms, corporate training, and any operation where learners need on-demand support.
Discuss a pilot →Quick facts
- Business size
- Edtech platforms, training-operations, corporate L&D teams
- Timeline
- 4-6 weeks test, 2-4 months pilot, 4-6 months production
- Budget range
- Pilot from €30k.
- Hardware
- Cloud-based.
- Data needed
- Course material (text, video transcripts), learning objectives, sample student questions.
- Evolution
-
- Genesis
- Custom-built
- Product
- Commodity
New ground. The models exist but are still shaky. We work it out as we go.
What this scale means
Further right means more proven and cheaper. Further left means newer and riskier. Here is the test for each step.
- Commodity
- You could get the result yourself from a ready service, with almost no work. We rarely take these on.
- Product
- A vendor already sells this result turnkey, like shelf recognition from Trax or document reading from ABBYY. If one of them fits you, use it. You come to us when it does not: when it has to run on your own servers, cost less, or fit systems the product cannot reach.
- Custom-built
- Vendors sell only the parts. A tool like Tableau hands you charts, but the dashboards and metrics for your business still have to be built. That build is the work, and that is us.
- Genesis
- The approach exists but does not work reliably yet. You are betting on it maturing, so it costs more and carries more risk.
Expected outcomes
The Problem
Students get stuck at 11pm. Instructors are asleep. Conventional course platforms cannot help. They were built for content delivery. They do nothing for the moment when a student does not understand the third paragraph of chapter 4.
The “human tutor” alternative does not scale. A typical course has hundreds of students, one instructor, and a few TAs. Most questions get answered slowly or not at all. Students drop out at the friction points. Completion rates suffer. Reviews trend down.
Modern LLMs are good enough at explanation, at meeting learners at their level, and at adapting tone to learning context. An AI tutor does not replace human instruction. It is the always-available first line that absorbs the question volume instructors cannot reach.
What the Solution Does
A tutor specialized for your course material.
- Question handling, student asks, tutor answers grounded in course content with citations.
- Concept explanation, adapts depth based on prior context (“I asked about derivatives earlier; explain integrals assuming I got that”).
- Practice problems, generates appropriate practice based on the student’s struggle areas (where supported by your content).
- Escalation, out-of-scope or ambiguous questions route to instructor / TA.
- Learning analytics, surface what students struggle with most (course-improvement signal).
Where It Fits
This makes sense if you…
- Operate an edtech / training platform with structured content
- Have measurable cost from instructor / TA load OR from student drop-off at friction points
- Have content material that can be ingested (text, slides, video transcripts)
- Are willing to design the tutor’s pedagogical tone
This is probably not the right time if you…
- Need fully accurate technical / medical / legal tutoring (LLMs hallucinate; not for high-stakes contexts without review)
- Have unstructured course material that resists ingestion
- Cannot route escalations to human instructors
Business Value
Coverage. The AI tutor typically answers 70-85% of routine student questions. Out-of-scope cases route to humans.
Instructor time recovery. Routine questions get absorbed, so instructors focus on complex, high-judgment questions and structural course improvements.
Completion-rate lift. Students who get stuck at 11pm and get help complete more often than students who get stuck and wait. The lift varies, and it is typically meaningful for friction-point content.
Learning analytics. What students ask and where they struggle surfaces course-improvement signals you cannot see without aggregated tutor logs.
How It Works
1. Content ingestion
Course material, text content, slides, video transcripts, and sample questions get ingested and embedded. It uses the RAG architecture from knowledge-base-assistant, specialized for educational content. RAG means the tutor looks up the relevant course material before it answers.
2. Pedagogical prompting
The tutor’s system prompt is tuned for teaching: ask clarifying questions, build on the student’s prior context, explain at the right level, and encourage the student through the problem.
3. Student context tracking
The tutor remembers what the student has asked, what topics they have engaged with, and where they have struggled. Future answers draw on this context.
4. Escalation
When the tutor can’t answer (out-of-scope, beyond the LLM’s confidence, student requests human), the conversation transfers to an instructor / TA with full context.
5. Analytics
Aggregated questions, common confusions, escalation patterns, all surface for course designers and instructors.
Stack
OpenAI, Anthropic, or self-hosted LLMs. RAG over course content. Conversation history per student. Datapipe for ingestion and retraining. Integration with your LMS or course platform.
What You Need to Make This Work
Data. Course material (text, slides, transcripts). Sample student questions help us tune.
Integrations. LMS / course platform integration. Escalation path to human instructors.
Hardware. Cloud-based.
Team. Course designer / instructor for content review. Pedagogical lead for tone calibration. LMS / IT contact for integration.
Implementation Roadmap
1. Test (4-6 weeks)
Pick one course, one subject. Ingest content. Tune pedagogical prompts. Test with sample student questions. Output: working tutor with a measured coverage rate.
2. Pilot (2-4 months)
Deploy to one course cohort. Measure question coverage, instructor time reduction, completion rates. Output: working production deployment with documented outcomes.
3. Production (4-6 months)
Expand to multiple courses. Continuous improvement.
Keep in Mind
- Tone matters. Pedagogy is more than Q&A. The tone has to encourage learning, and answering the question is only half of that. We tune carefully.
- LLMs hallucinate. For high-stakes content (technical certifications, medical, legal), hallucinations are unacceptable. We design with strict source-grounding and human review.
- Cheating concern. AI tutors can become “do my homework” tools. Pedagogical prompts can encourage learning over shortcut-giving, but the line is fuzzy. We discuss policy per platform.
- Content quality matters. Bad course content produces bad tutor answers. Sometimes the most useful finding is “the course material is wrong”.
- Multi-language adds work. Per-language tuning of pedagogical tone.
FAQ
Will this enable cheating?
Risk exists. Pedagogical prompts can encourage learning: asking clarifying questions, walking through reasoning, drawing the answer out of the student. Platform policy on what AI is allowed is your call.
Can the tutor grade student work?
Limited. Automated grading of well-structured problems works. Subjective grading of essays and open-ended projects is harder. We typically position the AI as a feedback tool, with the grade staying with the instructor.
Multi-language support?
Yes, with per-language tuning.
How does this compare to commercial edtech AI (Khanmigo, etc.)?
Commercial AI tutors are excellent for their platforms. Our approach fits when you need one of these: custom integration with a non-commercial LMS, fully on-prem deployment, or substantial customization of the pedagogical approach.
Ready to Discuss?
If you operate an edtech or training platform where student support is a real cost and quality line, this is a worthwhile pilot. We will walk through your content and your platform, and tell you what to expect.
Part of: AI Assistants ↗