Michele Meloni
Contact

Write to me directly.

For a collaboration, a question about my work or an initial conversation, use this form. For a detailed estimate, use the dedicated flow on the Services page.

Your message is sent directly to info@michelemeloni.com.

Back to articles
Article · AI & Interaction DesignMichele Meloni8 min read

When AI fails: errors, fallback and human handoff

How to design failures that are understandable, recoverable and safe instead of leaving people with an unexpected result or a dead end.

Published · 4 September 2026Updated · 4 September 2026

00 · Premise

Failure is not an exception to hide.

Ambiguous input, insufficient data, unavailable tools and inaccurate results are part of an AI experience. Designing for them protects the task, explains the next step and limits the impact of failure.

Fail usefully

Recognise, explain, preserve and guide.

In a conversational experience such as KIKU, an error should not erase context or interrupt the journey. The system must distinguish what it did not understand, what it cannot do and what requires human verification.

  1. 01Recognise

    Identify the type and impact of the problem.

  2. 02Explain

    Describe what happened without blame.

  3. 03Preserve

    Protect valid data and completed work.

  4. 04Guide

    Offer retry, another route or human help.

Italian and Japanese turn-taking conversation in the KIKU app
KIKU · language, turn and state should stay legible when the system cannot continue.
01

Map failure families

Failures are not interchangeable. Separate misunderstood input, model limitations, tool errors, missing permissions and temporary outages.

Practical actions

  • Connect each failure to an observable cause.
  • Assess impact and recoverability.
  • Set thresholds that stop automation.
02

Write messages for recovery

A useful message avoids vague language, protects trust and says what the person can do. It does not blame the user for a system problem.

Practical actions

  • Explain the problem in one plain sentence.
  • State what has been preserved.
  • Offer one or two actions that are genuinely available.
03

Design coherent fallback

When an AI feature is unreliable, the experience should provide another route proportionate to the task: editing, manual search, contact or deferral.

Practical actions

  • Keep previous input and results visible.
  • Allow the request to be narrowed or rewritten.
  • Never imply success without confirmation.
04

Define the human handoff

After repeated errors or before high-risk actions, human intervention should be simple and contextual. People should not have to reconstruct the entire history.

Practical actions

  • Transfer the summary, attempts and relevant failures.
  • Explain response time and channel.
  • Confirm when control has passed to a person.

Ready to use

Failure-state preflight

  • Main failure families are mapped.
  • The message separates cause and consequence.
  • Valid work is preserved.
  • Retry and alternative routes exist.
  • High-risk actions stop safely.
  • Human handoff includes the needed context.

Key principle

A good error protects the task, not the image of the AI.

Quality lies in recognising a limit, reducing harm and helping someone continue without losing information or trust.