Choreography: compliance training that actually sticks
Generic compliance e-learning gets watched with one eye closed and forgotten the next day. With AI generating micro-content specific to each role's risk, training becomes something people remember right when they need it, not a module they click through just to close out the certificate.
Every year, the same forty-five-minute "ethics and conduct" e-learning module reaches three thousand employees, from intern to CFO, with the same generic examples of bribery and conflict of interest. Most click "next" with the tab minimized and take the quiz with one eye closed, because none of that module's examples look like anyone's actual workday. Six months later, a buyer accepts an expensive dinner from a supplier without thinking twice, because the module never showed a scene like their own. You swap the single module for micro-content generated per role: for the buyer, a two-minute scenario about a supplier gift; for the manager, one about approving a direct report's expense. Each one sees the risk of their own job, not a generic lecture that speaks to no one.
You build a training with AI on fraud risk in reimbursements, using the company's real average reimbursement amount and the most common type of expense on the team. The context-free draft brought examples from a generic company, with amounts that didn't match the team's reality. With the right context, the scenario becomes recognizable, and the finance team remembers the rule when reviewing the next suspicious invoice.
You build a training scenario with AI for the legal team about the risk of citing case law without checking it, using a real, anonymized case that already happened at the firm. The generic draft, without that context, talked about "some lawyer" and nobody recognized themselves in it. With the firm's real case, even disguised, the team recognizes the situation and remembers the rule at the moment it matters: check the source before filing.
You build a training with AI for the marketing team about ad claims that need regulatory backing, using a real example of a piece that almost aired wrong. The generic draft talked about "some company" running misleading ads, distant from the team's routine. With the actual campaign's real case, even anonymized, the scenario engages and the team starts checking backing before approving the next piece.
You build a training with AI for managers about conflict of interest in hiring a relative, using the company's actual hiring approval process. The generic draft talked about "some policy" without citing the real approval flow. With the real process, the manager recognizes exactly where they'd need to declare the conflict, and remembers it the next time they hire someone close to them.
You build a training with AI for the product team on when a feature needs a data protection impact assessment, using a real feature the team already launched that required this assessment. The generic draft talked about "some app" collecting data, distant from the team's real product. With the real example, the team recognizes the pattern and starts flagging early when a new feature touches sensitive data.
You build a training with AI for the sales team about gifts and hospitality to clients, using the company's real commercial policy limit. The generic draft cited a limit that wasn't the company's, causing confusion. With the right number, the salesperson remembers the exact limit the next time a client suggests an expensive dinner.
You build a training with AI for the operations team on proper disposal of documents with personal data, using the company's actual disposal process already in place. The generic draft described a disposal process that didn't match the real one. With the right process, the team recognizes exactly the step they tend to skip and remembers to follow it next time.
You build a two-minute scenario with AI for the finance team about split payments, giving it real context: the company's approval threshold, the most common type of supplier in the area, how approval works in their system. The first draft, without that context, came out generic, with a threshold that didn't match the real one. You fix it with the right numbers and the scenario becomes recognizable: "this almost happened here." Six months later, an analyst remembers exactly this scenario when a supplier asks to split an invoice in two, and refuses. You save that scenario in the finance compliance training library, ready to reuse next cycle.
You build a training with AI for the technology team on privileged credential use, using a real, anonymized incident of improper access that already happened at the company. The generic draft talked about "some company" suffering an incident distant from the team's reality. With the company's real incident, the team recognizes the pattern and starts questioning out-of-scope access before granting it.
You build a training with AI for the design team on consent screens that need to clearly state what's being collected, using a real screen from the product that had to be redone for failing to meet that standard. The generic draft showed a screen example distant from the real product. With the company's real screen, the team recognizes the pattern and starts reviewing this point before any new flow goes into production.
You build a training with AI for the executive committee on compliance risk in expansion decisions, using the real history of a country the company already evaluated and dropped due to regulatory risk. The generic draft cited examples from markets distant from the company's reality. With the real history, even summarized, the committee recognizes the pattern and starts asking about regulatory risk before approving the next expansion.
How many times have you clicked "next" on a compliance training without paying attention to a single word? Whoa, almost everyone does this, and it's not a lack of character, it's that the generic module never talks about anyone's specific job. It talks about "some employee" at "some company," with an example that doesn't look like anyone's actual desk. And here's the point most people miss: the problem isn't the training, it's what nobody gave it before generating. If you give AI the role's real context, the specific risk of that function, the company's real process, it starts generating micro-content that the person recognizes as theirs. It's not magic, it's feeding. Fair?
The core idea of this lesson. Generic e-learning delivers a generic example. With role context, AI generates a recognizable scenario. The choreography has four beats: you give the context and risk of the role, AI generates the specific micro-content, you adapt with judgment (because AI doesn't know the exact process or the real incident at the company), and what engaged becomes a new piece in your training library by role. The danger to name right away: publishing a scenario without reviewing the real risk it describes is a recipe for training that teaches the wrong thing. AI accelerates the draft; the responsibility for the content stays yours.
Back in lesson N.jur.6 of this course you saw that a workflow that works shouldn't be rediscovered every time, it should be packaged. Here's the same economics applied to training: the scenario that engaged the purchasing team shouldn't be rewritten from scratch every cycle. It should be saved, feeding the next one.
01Why generic e-learning doesn't stick
Think of AI as a competent instructor who's never worked at your company. You ask for "a conflict-of-interest training" and it delivers the textbook training: technically correct, but with none of your process, your threshold, the specific risk of the role of whoever's going to watch it. It comes out soulless not from incompetence, but from lack of context.
The cure isn't switching AI tools or polishing the isolated request. It's giving the role's context before asking: what's the real risk of that function (the buyer receives a gift from a supplier, the manager approves a direct report's expense, the salesperson negotiates with a client who asks for a favor), what's the company's real process, and, when possible, a real anonymized incident that already happened. When AI has that in front of it, it stops inventing the example from scratch and starts generating the scenario the person recognizes.
02The training library by role feeds generation
The fuel for generating the right risk is your training library by role: the set of scenarios you've already tested and that truly engaged. Supplier gift for the buyer, expense approval for the manager, ad claim for marketing, privileged credential for technology. Each one is a ready piece, organized by role, not by generic theme.
Notice the dashed arrow closing the loop: the scenario that engaged goes back to the library. Every well-made training doesn't end at the "completed" click; it deposits a new piece in the stock, and the next cycle already starts out better. You're building a training asset while you work, not just fulfilling an annual obligation.
03The choreography in four beats
- Beat 1, you give the role's context. Present that function's real risk, the company's process, and, if you have it, a real anonymized incident.
- Beat 2, AI generates the micro-content on the right risk. With the context in front of it, it builds a short scenario (two to five minutes) that the person in that role recognizes.
- Beat 3, you adapt with judgment. AI built the skeleton, but it doesn't know the exact incident at the company or the right tone to avoid exposing anyone. This beat is yours, and non-negotiable.
- Beat 4, what engaged becomes a new piece. The scenario the team commented on, that generated a real question in the session, enters the library by role. The cycle closes and the stock grows.
AI accelerates beats 1 and 2, which used to consume the most writing time. Beat 3 stays human work, with responsibility for what will teach three thousand people.
04The danger: publishing without reviewing the real risk teaches the wrong thing
I'll be direct, because this is the part that separates those who use it well from those who get burned. AI writes fast and writes well, and that fluency is a trap: the scenario looks so ready you feel like publishing it directly. Don't. AI doesn't know that the threshold it used is wrong, doesn't know the approval process it described changed last month, doesn't know that gift example could read as a veiled accusation against a specific area of the company.
The rule is simple and has no exception: AI accelerates the draft, the responsibility for what it will teach is yours. A training that describes a wrong process isn't just boring to watch, it teaches the employee to follow a rule that doesn't exist, and that's worse than not training at all. Think of it as a talented writer who's never sat in one of your expense-approval meetings: you'd never publish their text without checking whether the process matches reality.
Do it now
Take a real role at the company that needs specific compliance training for your real task and run the choreography once, start to finish:
- Build the role's context. Write the real risk of that function (what the person could do wrong, even unintentionally) and the company's real process that applies.
- Ask for the micro-content. Ask AI for a two-to-five-minute scenario, citing the role, the risk, and the real process, and ask it not to use a generic textbook example.
- Read and adapt with judgment. Mark what AI got right in format but wrong in real process or tone. Adjust with your own head, not its.
- Deposit what engaged. After running it with the team, save the scenario that generated a real question in a "training library by role" folder. Next cycle, it already goes in as context.
Compare: how long would it take to write this specific scenario from scratch, and how long this choreography took.
Practice
1. Why does generic compliance e-learning tend to get watched without attention and forgotten right after?
2. In the four-beat choreography of training by role, which step is human work and can't be delegated to AI?
3. What is the risk of publishing an AI-generated training scenario without reviewing the process it describes?
For the board
On why it does not stickthe generic module talks about some employee at some company. Nobody recognises themselves in it.
On what cannot be delegatedthe AI gets the format right and can get the real process or the tone wrong. Reviewing the actual risk is yours.
On the trapa generated scenario that looks finished, built on an outdated process, teaches the employee to follow the wrong rule.
Thanks for the feedback. It helps sharpen the next lesson.