A prospect asked us last month: "You already build RAG chatbots over documents. Isn't AI Mentor the same thing with a different landing page?"
Fair question, and the honest answer is no - but the reason is not a feature list, it is which problem each one solves.
The chatbot's job ends at the answer
A retrieval chatbot is measured on one thing: did the person get a correct, grounded answer quickly. When that works, it is genuinely valuable. Support staff stop interrupting seniors, customers get consistent responses, and the answer cites the paragraph it came from so nobody has to trust the model's memory.
Then the conversation closes and the system forgets the person entirely. The next query starts from zero. Ask the same agent the same policy question in March and they will look it up again, because looking it up is all the system ever taught them to do.
For a customer-facing assistant, that is the correct design. Nobody wants a shopping chatbot quizzing them on last week's conversation.
The training system's job starts at the answer
For internal procedures, the same behaviour is a failure. An agent who looks up the escalation rule every time is an agent who cannot act while the customer is still on the line. The goal is not to answer the question, it is to make the question unnecessary.
That requires something the chatbot deliberately does not have: a per-person model of what has been learned and what is slipping. AI Mentor keeps one.
- Every SOP becomes assessment items, generated from the document itself.
- Every answer a person gives is recorded against them and that item.
- The next review is scheduled by FSRS - per person, per card - for the day the model expects recall to fall below target.
We built the AI Mentor demo around the scheduler rather than around a chat window, precisely because a chat window would look identical to the Gen AI demo two pages away. Rate a card "Forgot" and watch its stability collapse and its next review jump to tomorrow. That recomputation is the product.
Why not just add memory to the chatbot?
We tried the obvious shortcut: keep a history of what each user asked and use it to personalise answers. It does not work, for a reason that is easy to miss.
Asking a question tells you the person did not know the answer. It tells you nothing about whether they know it now, or will next month. A retrieval log is a record of ignorance at a moment; a review log is a record of recall over time. The second is what you need to schedule anything, and you cannot derive it from the first.
You also need the person to be tested, not just informed. Someone who reads a generated answer feels they understood it - the fluency of a good explanation is routinely mistaken for competence. Only an attempt to retrieve it later, unaided, distinguishes the two.
What this means when choosing
The honest split we give people:
- You need correct answers, fast, at scale - a customer assistant, an internal helpdesk, a document Q&A. That is Gen AI & Agents. A mentor would be overkill and annoying.
- You need people to be competent without the tool - regulated procedures, onboarding, anything where hesitating in front of a customer costs you - that is AI Mentor. A chatbot there is a crutch that quietly prevents the learning it looks like it is providing.
- You need both, often. The chatbot handles the long tail nobody should memorise. The mentor handles the twenty procedures everyone must know cold.
The measurement changes too
This is the part that decides which one you actually bought. A chatbot reports deflection rate and answer quality. A training system has to report retention: what fraction of the team can produce this policy without looking, and is that number going up.
If a vendor sells you a training system and reports usage minutes, they sold you a chatbot with a course catalogue.