A familiar pattern is emerging in AI training initiatives. The budget is approved. The workshops are held. Staff have learned to prompt. Six months later, no one can say what was actually achieved.
The problem is not the AI and not the staff. The problem is in the structure of the initiative, and the structure is something we can fix.
The same structural flaw, at two levels
Subject-matter experts' time is valuable and they are extremely busy. Questions within their own field are routine for them, built on years of work. A lawyer spots a compliance risk in a document immediately. A safety manager grasps a process hazard through experience that cannot be copied. A headhunter recognises quickly whether a candidate fits the role.
When a specialist is asked to build training materials about their own field, two additional jobs are placed on their shoulders: pedagogical design and technical production. Pedagogical design is its own discipline, involving learning-objective formulation, cognitive-load management, assessment validity and feedback design. Technical production covers system integrations, data architecture, information security and accessibility. Neither should be part of the specialist's job description.
The same triple burden is distributed across the whole organisation when the proposed solution is "let's train everyone to use AI". Now, instead of a single specialist, every employee is expected to combine subject knowledge, pedagogical structure and technical production. Marketing runs one tool with its own prompts, HR runs another, finance runs a third. Content quality varies from team to team, and no one is responsible for what staff have actually learned.
In both cases the outcome is the same. The subject matter is correct, but the content is pedagogically weak and technically difficult to use. At worst, months of valuable time are spent on work that should have been done elsewhere. And because project success is measured by whether the AI tool is in use, the real question goes unexamined: did staff learn what they were supposed to learn?
A model that actually supports learning
A working model distributes the work across four roles. Each does what they do best.
The subject-matter expert provides the knowledge. In practice, one hour of interview, delivery of existing materials, and approval of the finished content. They do not prompt, fill in templates or learn AI tools.
The learning designer structures. Learning objectives, learning structure, assessment design and feedback design belong to this role.
The AI agent produces volume. Interactive exercises, questions, practice dialogues and feedback. This is the strength of AI, but only when the agents have been trained carefully with prompts and memory files, and when their work is properly guided. Without this, you get apparent volume whose quality does not stand up to pedagogical scrutiny.
The technical architect integrates. System connections, data architecture, information security and accessibility are their own discipline.
In this model, the subject-matter expert gets to focus on their own work, and the rest of the production happens in professional hands. The content is also produced significantly faster.
Where this matters most
We see the same mistake repeated in many situations, such as:
Compliance training. Employment law, data protection, occupational safety, sector-specific regulations and ethical guidelines. This kind of training is mandatory. In most cases it is produced in a way no one would recommend: a specialist writes the text, someone else hurriedly turns it into a course, staff click through it. The box is ticked. Learning stays superficial. When the next audit arrives, the same work is done all over again.
Strategy rollout. Project-based, scheduled, organisation-wide. The subject matter sits with leadership, but rollout requires pedagogical structure and production capacity. Those are not part of leadership's day job, and a couple of workshops will not bring them in.
Onboarding. Traditionally under-resourced, considered dull, yet critical for how a new employee integrates. Onboarding is precisely the point where a well-produced AI-supported learning companion changes the experience. The employee gets the answer when they need it.
Three things to consider
A tool alone is not enough to produce the content. A share of the investment should go to the content itself, or to a partner who produces it for you.
Measure learning, not usage. "Is the AI tool in use" is easy to measure. "Did staff learn what they were supposed to" is harder, but that is the metric that tells you the project's value.
Keep your subject-matter experts on subject-matter work. Their time is valuable, and the most productive use of it is in their own work. Pedagogical design and technical production belong to a specialised team, whether internal or an external partner.
In closing
At 3DBear we are building a production model where each role has its own place. We are happy to have a conversation and give you an honest assessment of your situation.
If you would like to hear more, book 20 minutes. We will tell you whether this kind of solution would suit your needs.
Author
| Name | Jussi Kajala |
| Role | CEO and founder, 3DBear |
| Background | PhD in computational physics (Aalto University), MPhil from the University of Cambridge (with the title of Master of Philosophy). Previously at Spinverse, leading key enterprise accounts. Tekes Employee of the Year 2015. |
| jussi@3dbear.fi | |
| Phone | +358 50 561 5411 |