The training in this branch is repetitive, examinable and exactly the shape an AI tutor handles well: quiz a learner on FAIS, explain why an answer is wrong, track progress across a curriculum. We default to edge models for cost and data-locality, reach for frontier APIs when partners need them — and a human expert moderates the content. The model is a delivery mechanism, not the authority.
A patient, repeatable drill agent over examinable regulatory content — quiz, explain, track.
An AI tutor here is an agent that works a learner through a defined curriculum — FAIS fit-and-proper, RE1/RE5 framework knowledge, PCI awareness, AML/FIC duties — by quizzing, explaining and tracking. It asks scenario questions, marks them, explains why the right answer is right and the wrong one wrong, and keeps a record of where each learner is strong or weak.
This works precisely because the subject matter is bounded and examinable. Unlike open-ended advice, regulatory training has known correct answers grounded in the Act, the codes and the standards. The AI is a patient, infinitely repeatable drill partner over content that already has a defined truth — not an oracle inventing the syllabus.
The honest framing: the model is a delivery mechanism. It makes good content reachable, repeatable and adaptive. It is not the source of authority — that is the curriculum and the human experts behind it.
Edge for the common case, frontier or open when a partner needs it — never welded to one vendor.
The default runtime is edge models on Cloudflare Workers AI. For tutoring and assessment over a fixed curriculum, a capable open model at the edge is fast, cheap per interaction, and keeps learner data close — a real advantage for SA institutions weighing data locality and cost-per-seat. Most quiz-and-explain interactions never need a frontier model.
But the architecture is deliberately model-portable. Where a partner requires it — harder explanatory reasoning, a contractual model preference, an existing API relationship — the same tutor can route to frontier APIs (Claude, GPT, Gemini) or to a self-hosted open model (Ollama). We never lock the surface to one model or one vendor; the model is configurable, the curriculum is constant.
Cloudflare Workers AI for everyday quiz/explain — low cost per seat, low latency, data stays close.
Claude / GPT / Gemini via API when a partner needs deeper reasoning or has a model preference. Routed, not hard-wired.
Self-hosted (Ollama) for institutions that want full control of weights and data. Same tutor, different backend.
No lock-in to a single model or SDK. The backend is a config choice; the curriculum and the moderation do not change.
The content an AI tutor teaches against is authored and moderated by human experts, on the same review cadence as the rest of pay.2nth.ai. The model does not decide what is correct under FAIS or the FIC Act — the curriculum does, and a domain expert signs it off. The AI’s job is to deliver and assess against that signed-off content, not to improvise regulatory positions.
This matters most exactly where AI is weakest: regulation is specific, dated and consequential. A confidently wrong explanation of a fit-and-proper requirement is worse than no tutor at all. So the design keeps the human in the loop on content correctness, uses the model for delivery and adaptivity, and grounds explanations in the moderated material rather than the model’s open-web memory.
Every leaf in this branch carries a review stamp for the same reason: the authority is the expert and the source, surfaced transparently — the AI sits on top of that, it does not replace it.
AI tutors are strong where the content is bounded, repetitive and self-assessable — RE framework drills, PCI awareness refreshers, AML scenario practice, CPD-style knowledge checks. They scale patient repetition to any number of learners at a per-seat cost a human tutor cannot match. That is the case to deploy.
They are weak, and should not be trusted alone, on novel or high-stakes interpretation — a genuinely ambiguous regulatory question, anything where a wrong answer carries real consequence. There, the tutor should defer to the moderated content and to a human, not generate a confident guess. Design the escape hatch deliberately.
For a partner deciding the backend: start at the edge for cost and data locality, and only reach for a frontier API where the explanatory quality genuinely demands it. Paying frontier prices to drill multiple-choice RE5 questions is the easy money-waste to avoid.
A fluent but incorrect regulatory explanation is dangerous. Ground answers in moderated content; do not let the model freelance the law.
Practising with an AI tutor builds readiness; it is not the FSCA exam, a CPD hour, or a PCI sign-off. Keep the distinction explicit to learners.
Most drill/track interactions run fine at the edge. Reaching for frontier compute by default burns budget for no learning gain.