How should organisations choose work problems for AI learning?

Asking learners to bring real work into AI training sounds practical. It can also create a session filled with sensitive information, problems too broad to resolve and tasks whose answers cannot be checked.

Generic exercises avoid some of these risks but may feel disconnected from professional decisions. The design challenge is to preserve the part of real work that makes learning useful while controlling scope and consequence.

Choose the work problem before designing the AI activity. The right problem makes capability visible.

Real work creates relevance and risk

Recognisable tasks help learners connect a method to their role. They bring familiar terminology, constraints, source material, quality standards and consequences into the learning experience. This creates a shorter bridge to later workplace application.

Live work can also be a poor learning environment. A current customer case may contain information that cannot enter the tool. A strategic problem may depend on many teams and unresolved facts. A high-consequence decision can make experimentation inappropriate. Learners may focus on delivering the work rather than examining how they reasoned.

Realism should come from the decisions and conditions that matter, not from exposing live data or reproducing every complication.

Tool-led and open-ended selection often misses the learning need

A tool demonstration can efficiently show a feature such as summarisation or classification. It becomes weak learning when the activity asks only whether the feature works. The learner may never decide whether the task is suitable, how to check the result or what professional standard applies.

Open learner choice can increase relevance and ownership. Without a selection screen, however, people may choose whatever work is urgent, easy to describe or personally interesting. Five unrelated problems can make feedback shallow and comparison difficult.

Start with a performance need. What should someone be able to do differently? Then find a problem that requires that behaviour. The AI technique follows the learning objective.

Screen candidate problems for learning value

A suitable problem should meet several criteria:

• Role relevance: learners recognise the work and its purpose.
• Observable decisions: task selection, evidence checking and professional judgement can be seen.
• Accessible evidence: there is a source, standard or rubric against which work can be reviewed.
• Bounded scope: the task can be attempted and reflected on within the available time.
• Manageable consequence: errors remain inside the learning environment.
• Appropriate information: material is synthetic, public, sanitised or specifically approved.
• Repeatability: variations can be created so learners apply a method rather than memorise an answer.
• Transfer opportunity: similar reasoning is likely to appear in the learner's work.

A business problem can be important yet fail this screen. That does not mean it lacks value. It may belong in a separate discovery, governance or implementation process rather than an early learning activity.

Role experts and learners can help identify recurring problems. Learning designers then assess which candidates reveal the target capability safely and clearly.

Convert the work without losing its professional core

Simplify the problem by removing dependencies that do not support the learning objective. Replace confidential material with synthetic or carefully sanitised equivalents. Preserve the terminology, ambiguity, evidence and decision points that require expertise.

For example, a live claims decision may be unsuitable. A fictional file can still contain conflicting dates, an omitted exclusion and an output that needs referral. The professional reasoning remains while the live consequence is removed.

Pilot the scenario with a role expert and someone similar to the intended learner. Check whether instructions are clear, evidence is sufficient, controls are realistic and the activity produces the decisions expected. Include accessible formats and avoid unnecessary technical complexity.

Reject or redesign a problem when it has no clear success criteria, depends on unavailable information, introduces too many new concepts or rewards a polished output without evaluation. Also reject it when AI is obviously inappropriate and the only learning would be a predetermined refusal; a balanced suitability exercise may serve that goal better.

End with reflection on transfer. Learners should identify where the method could apply, what would differ in live work and which controls remain necessary. A well-chosen problem teaches a reusable way of thinking rather than one impressive use of a tool.

Example

A claims team proposes three training problems: summarising a fictional file, predicting a disputed settlement and organising public policy guidance.

The learning designer and claims expert screen each candidate for evidence, consequence, scope and transfer. They reject the settlement prediction as too consequential for early practice. The policy task is safe but less relevant to the target behaviour. They select the fictional summary, add a material omission and create a source-based review rubric.

The chosen problem preserves real claims reasoning while avoiding a live decision and giving learners clear evidence against which to evaluate the output.

FAQs

• Must practical AI learning use live work?

  No. Synthetic or sanitised material can preserve professional terminology, ambiguity, constraints and decisions without exposing sensitive information or creating live consequences. The learning value comes from authentic reasoning, not the status of the data.

• Should learners choose their own work problem?

  Choice can improve relevance and ownership when candidates are screened for permission, scope, evidence, consequence and fit with the learning outcome. Early learners may benefit from a curated set before adapting the method to their own work.

• What makes a problem too complex for an AI learning activity?

  Warning signs include an unclear objective, excessive dependencies, unavailable evidence, high live consequences and too many unfamiliar decisions at once. Simplify the scenario, separate it into stages or choose a different problem.