Season · AI, Trust & Human Oversight · 9 of 16
When automation stops, people still need a way to carry on.
Concept of the Week · Give people more than “Try again”
A fallback lets people complete a task when the system they rely on becomes unavailable. It keeps the information and actions needed for the task accessible without that system.
Design rule: Keep essential information, actions, and progress clear and accessible, so people can continue if AI fails.
During a ward round, a physician uses a summary generated by an artificial intelligence (AI) system to review overnight changes in a patient’s condition. The AI system stops responding, while the rest of the clinical system—including the patient record—remains available. A screen displaying only “Try again” offers no way to open the notes and results from the same view.
A proposed design keeps the patient’s notes, observations, and results accessible throughout the review, including when the AI system is unavailable. It shows which source entries support the summary and keeps unfinished checks visible, with a named physician responsible for each. The physician can review the patient record and document their assessment without waiting for the AI system.
The AI system can draft a summary, but the physician still reviews it against the patient record and makes their own assessment. If essential information is unavailable, the team follows its agreed escalation or downtime procedure.
Ship this by making patient records, assessment documentation, and review progress accessible with or without AI, then testing, revising, and retesting the design with representative physicians in simulated rounds against defined usability and safety criteria before release.
Left, a ward physician can only retry the unavailable AI system; right, the proposed design keeps the patient record and unfinished checks within reach.
Asset & Resource
Design and human factors teams can use this Work With and Without AI guide when designing an interface for a task that people may need to take over.
Build the human path: Identify the information and controls people need to complete the task themselves, then integrate that path into the everyday interface.
Show the work: As the AI system works, show the information it uses, the actions it takes, and what remains; let people inspect completed steps without following every detail live.
Keep controls within reach: Keep the same information and controls accessible during AI use, with a clear way to take over and prevent conflicting actions.
Preserve the next step: If the AI system fails, retain completed work, flag anything unfinished or uncertain, and keep the next permitted action available.
Test learning and takeover: Let representative users observe repeated tasks, then try a similar task and a simulated mid-task outage without AI; use missed steps and assistance needed to revise and retest the design.
Show how the task is done while the AI system performs it, so users can learn to use the same information and controls themselves. Encourage them to follow the steps occasionally and to practice when needed. Test whether they can perform the task without AI, including after a period of relying on it.
Microsoft’s UX design for agents (2025) offers guidance on making AI actions visible and controllable, even when the AI system works in the background.
Light Wisdom & Reflection
“Everybody has plans until they get hit for the first time.”
If the AI system were unavailable today, which task would your team be unable to complete—and what would make it possible again?
Mindful Practice
On a familiar walk, leave your phone in your pocket for one minute and notice what’s around you.
Which landmarks tell you where you are?
What do you notice when no prompt tells you the next turn?
Next time, notice one detail that helps you feel oriented.
Best,
Andreas Walden
Share this with someone designing how work continues during an outage.
For another way a quiet system can mislead, explore what a clinical response changes about later alerts.