Example scenario. A customer calls for the third time. This time they want to know why they were given conflicting information. The adviser can read the first date on screen or verify what was promised. This short decision is what the role play should bring into view.
What does clinical communication research tell us?
A 2026 JMIR study of simulated patients explores opportunities and limits of AI conversation practice in clinical training. A workplace pilot is needed to establish whether the approach improves customer service performance.
A workplace example
Have the customer say they received two different delivery dates. The employee should acknowledge the inconvenience, confirm the facts and agree on a callback time. Use fictional customer details.
Find the decision point
The customer is calling for a third time. “You told me yesterday that someone would call today.” The adviser can end the conversation quickly by offering a new delivery date, but that may create another unsupported promise. The scenario tests what they choose under pressure.
A 2026 JMIR study assigned 50 medical students to AI-supported practice or standard training. After one academic year, the difference in communication improvement between groups was not statistically significant. The study also reported a usable practice environment with limitations in reproducing the emotional range of real consultations. Consider both findings when designing a workplace pilot.
Design the first conversation
- Hear the customer’s experience
- Check the conflicting information
- Offer an option within policy
- Agree the callback time
| Moment | Weak example | Behaviour to observe |
|---|---|---|
| Opening | “That’s what the system says.” | Summarises what happened to the customer |
| Uncertainty | “It will definitely arrive tomorrow.” | Distinguishes verified information from estimates |
| Closing | “Someone will get back to you.” | Names who will respond, when and how |
Make a second attempt count
Select one priority from the first report. If the adviser omitted a callback time, test that skill in a different problem. Use a billing error instead of a late delivery, for example. Keep difficulty comparable and look for the same behaviour.
Compare simulation results with a manager’s observations of sampled real conversations. If high scores do not correspond to the behaviour at work, review the scenario design, job conditions and scoring before deciding how to expand the programme.
How should the customer react?
The fictional customer can cooperate when asked for a specific detail and refer to the previous promise if the adviser brushes them off. Tie reactions to observable behaviour. If the character gives the same reply every time, the exercise offers fewer chances to observe listening and adaptation.
Keep authority to resolve the case separate from communication quality. An adviser without refund authority can still explain the process and escalate correctly. Ask managers to try the scenario before the pilot and flag unrealistic reactions or requests that conflict with policy.
Set the scenario boundaries before scripting reactions
Specify the product, customer history and adviser’s authority in a few sentences. An adviser may be able to check delivery but unable to approve a refund. Pressing them to make a promise the company cannot honour adds difficulty without clarifying the intended skill. Decide where to test fact checking and where to test how the adviser explains the limits of their authority.
For the COBIDU pilot, give the team a practice example and a second conversation of similar difficulty. Choose one development goal per person from the first report and record the wording that shows the behaviour in the second conversation. Check that the cases are comparable before interpreting a change in score.
How to apply it with COBIDU
Create the scenario in COBIDU AI Simulation and assess listening, verification and clarity of next steps. After the first attempt, choose one point of feedback. Give the employee a similar but unfamiliar scenario and compare the result with a manager’s review of a sample conversation.
Two attempts raise a further question. If two reviewers hear the same conversation, can they agree on what counts as success?
