Training that is generic gets ignored. AI is good at making training specific to your own procedures, which is the only kind anyone remembers.
Key takeaways
Yes, AI can deliver compliance training, and it is most useful where generic training is least useful: turning your own written procedures into scenarios, questions and short refreshers that reference the way your practice actually works. It can also keep the records, spot who has been missed and generate a fresh question bank when a procedure changes.
What it cannot do is decide what your obligations are, or make training count towards a professional requirement. Accreditation is a regulator’s decision, and the content still has to be right, which means somebody in the practice reads it before anybody is trained on it.
If your practice is a reporting entity, FINTRAC’s compliance programme guidance requires you to develop and maintain a written, ongoing compliance training programme for employees, agents or mandataries and other authorised persons, and separately to institute and document a plan for that ongoing training and to deliver it. Note the two limbs: the programme and the plan. Practices routinely have material and no plan, which is the finding that surfaces in a review.
Privacy law adds a second, quieter training duty. Clause 4.1.4 of Schedule 1 to PIPEDA requires organisations to implement policies and practices giving effect to the principles, including training staff and communicating to staff information about the organisation’s policies and practices. The federal Voluntary Code of Conduct on Advanced Generative AI Systems points the same way, listing among its accountability measures the establishment of policies, procedures and training so that staff are familiar with their duties.
Turning procedure into scenario. Nobody remembers a policy paragraph. People remember “a client asks you to invoice a different company than the one you onboarded — what do you do?” A model is good at generating a dozen such scenarios from your own procedure text, which a senior person then culls to the five that are realistic.
Question banks that change. The same twelve questions every year train people to remember answers, not rules. Regenerating a bank from the current procedure each cycle is cheap and materially better.
Targeted refreshers. If your quiz results show that six people misread the same rule, generate a two-minute refresher on that rule for those six rather than repeating the whole module for everyone.
Records. Attendance, version trained on, score, date, and who has not completed. This is the part a reviewer asks for and the part practices most often reconstruct after the fact.
Translation and accessibility. Producing the same module in French for a bilingual team is a translation job with a review step, covered in AI translation for Québec client work.
A practice has an AML procedure nobody has read since it was written. The rebuild takes a week.
Day one, the procedure is reviewed by the compliance officer and the sections that carry real decisions are marked — roughly eight of thirty pages. Day two, a model drafts fifteen scenarios from those eight pages; the officer keeps six and rewrites two of them because the model made the client too obviously suspicious. Day three, a twelve-question bank is generated and pruned to ten. Day four, the module is assembled and the record structure is set up. Day five, it is delivered and the plan — who, what, how often, and what triggers an off-cycle session — is written down.
The result is not that training got cheaper, though it did. It is that the training now describes this practice. The scenario about invoicing a different company is one that actually happened in the firm two years ago, which is why people remember it.
Whether internally generated learning counts towards a professional requirement is a matter for your regulator. CPABC publishes its Act, Bylaws and Code of Professional Conduct alongside its CPD requirements, including a four-hour verifiable ethics requirement. In insurance, the Insurance Council of BC publishes continuing education requirements for licensees, FSRA regulates Ontario’s life and health sector, and CISRO is the forum through which Canadian insurance regulators pursue consistent qualification and conduct standards and publishes its member list. For lawyers, professional development requirements sit with the law society, alongside the conduct rules harmonised through the Federation’s Model Code.
Ask the question before you build, not after. A module that turns out not to count is still useful internal training; a module you told staff would count and does not is a different kind of problem.
Three places. Someone qualified confirms the content is correct before delivery, because a confidently wrong training module propagates the error across the whole practice at once. Someone decides what happens when a person fails — that is a supervision decision, not a score. And someone owns the plan, because the plan is the artefact a reviewer asks for. The employment framework that makes training mandatory rather than optional is the handbook; our sister firm covers how workplace policies are written to be enforceable.
Can AI mark the quiz?
For multiple choice, yes. For free-text answers on a judgment question, use it to draft feedback that a person reviews — a wrong answer on a compliance question is information about a person’s understanding and deserves a person’s attention.
How often should the training change?
Whenever the procedure changes, plus the cycle your programme commits to. Regenerating a question bank from an updated procedure is now cheap enough that there is no excuse for training against a superseded version.
Do we still need external training?
Often yes, particularly where a regulator expects verifiable or accredited hours. Internal AI-built training is best at the practice-specific layer — your procedures, your systems, your clients — which external courses cannot cover.
A 30-minute call is enough to tell you whether AI pays for itself here.
