tanai
CV ↓

20 of 30 bug reports never reach a developer

  • Metis incident portal
  • AI
  • Live since May 2026
  • 4 min

This file covers an AI assistant that asks the right questions.

READ IN 8 SECONDS

PROBLEM
About 30 reports a day, and most were questions rather than bugs.
DECISION
An LLM checks each report against the system's data models and the code in git, then classifies it as a question, bug or idea.
RESULT
About 20 of 30 handled without a developer. Repeat escalations fell 30% in four months.

END OF SUMMARY. SCROLL FOR DETAILS ↓

9 slides · use ← → to navigate · swipe on mobile

FILE 03 · AI INCIDENT PORTAL
  1. READ IN 8 SECONDS

    PROBLEM
    About 30 reports a day, and most were questions rather than bugs.
    DECISION
    An LLM checks each report against the system's data models and the code in git, then classifies it as a question, bug or idea.
    RESULT
    About 20 of 30 handled without a developer. Repeat escalations fell 30% in four months.

    END OF SUMMARY. SCROLL FOR DETAILS ↓

    Tap right for next · tap left to go back

  2. We receive a lot of user feedback, and most of it is questions, not bugs. Someone still had to read each one.

    Incident portal overview listing incoming reports with priority and status

    Most "bugs" were questions.

  3. Ask Metis now handles the first conversation, as a good support agent or BA would: what happened, what did you expect, and can you send a screenshot?

    Ask Metis chat asking the reporter what happened, what they expected and for a screenshot

    It asks for steps, not only "it broke".

  4. It then does the research: the data models, the code in git and existing tickets.

    1. New report
    2. Ask for details
    3. Check data models + git
    4. Question? Answered
    5. Bug or idea? Jira

    About 20 of every 30 turn out to be questions, and those are answered immediately.

  5. Your turn

    Sometimes it is a genuine bug. What should happen next?

    Choose an option before you continue.

  6. A. Try to handle it anyway
    B. Hand it to a person with everything it collected my pick
    C. Ask the user to wait

    B. It creates the Jira ticket with the summary, screenshots and session recording attached.

    Principle 01

    Automate the questions, not the decisions

    The AI collects and checks. People decide what to fix.

  7. Developers reply in Jira. Replies tagged [external] appear in the portal, so the user gets a reply without anyone copying messages.

    [external]

    the tag that makes a Jira comment visible to the person who reported the issue

  8. Session recordings show what the user actually clicked, which often differs from what they remember.

    Principle 02

    Trust the recording

    People describe what they intended to do. The recording shows what they did.

  9. Guess the number

    your guess: 10

    Make your best estimate.

  10. RECEIPT · METIS INCIDENT PORTAL

    REPORTS PER DAY
    ~30
    HANDLED BY AI
    ~20
    REPEAT ESCALATIONS
    -30% IN 4 MONTHS
    LIVE SINCE
    MAY 2026
    ALSO USED BY
    PRINT.COM
    DEVELOPER READS SAVED / DAY
    ~20

    EVERY REPORT TRIAGED

    About twenty a day, handled without a developer.

  11. SWIPE UP ↑ Next file: The portal nobody used

The receipt

Every change from this file, on one receipt.

What I'd do next

Show users the answer the assistant found for a similar ticket before they finish typing.

RECEIPT · METIS INCIDENT PORTAL

REPORTS PER DAY
~30
HANDLED BY AI
~20
REPEAT ESCALATIONS
-30% IN 4 MONTHS
LIVE SINCE
MAY 2026
ALSO USED BY
PRINT.COM
DEVELOPER READS SAVED / DAY
~20

EVERY REPORT TRIAGED

Ask tan.ai