When an AI employee fails: how you find out and how to fix it

When an AI employee's run fails, open the run from the notice and press Explain and fix. Ask xMap reads the error and the blocks that failed, tells you in plain words what went wrong and, where it can, repairs the flow, with Undo if you disagree. If the cause was a service that did not answer for a moment, use Retry instead: it starts again from the failed block and reuses what the earlier blocks already found, so nothing is fetched or sent twice.

This article is for whoever owns an employee that stopped working. It covers how the failure reaches you, the causes that come up most, a made-up repair from start to finish, and the settings that let an employee keep going when one step has a bad day.

How you hear about it

You are told once when an employee starts failing, not every time it fails again. The same failure shows up in several places so you do not miss it:

• an urgent note in the employee's thread in the Team inbox, written at once;
• a notice in the bell and in Needs you at the top of the notification centre, with Fix (with xMap AI) and Open (the run);
• an email to its owner, at most one a day, linking straight to the fix;
• the employee's card on the Flows page, which shows it is not working, and the Needs attention tab.

When a later run on its own passes again, you get a short works again email, so you know the streak is over without opening anything.

What usually went wrong

Every failed step's message is shown in words in the run panel.

A worked repair (made up)

Aisyah owns a Rival watcher in Kuala Lumpur that posts new competitor branches to a Slack channel every morning. On Tuesday the bell shows Run failed. She presses Fix. The editor opens on the run and Ask xMap writes: the Slack step failed because its account is no longer connected; it was the personal connection of a colleague who left the team, so the flow cannot use it. The places were found and scored correctly.

Ask xMap cannot sign in to Slack for her, so it says so and offers the connection window. She connects the team's Slack account, the block picks it up, and she presses Retry. The retry reuses Tuesday's search and scoring and only posts the message, which arrives a few minutes late. Wednesday's run passes on its own and the works again email follows.

Had the cause been a setting rather than an account, Ask xMap would have changed it directly and shown the change under its reply with Undo this change.

Retry, or run it all again

• Retry from the failed block (the default when it can): blocks whose settings and inputs have not changed are reused, so the search is not paid for twice and anything already sent is not sent again.
• Run it all again: starts from the top. Use it after you changed an early block, such as the city, so the new setting is used everywhere.
• A run is retried once. If the retry fails too, retry the retry after fixing what it says.

Letting a step fail without stopping the employee

Most blocks have an If this block fails section under Advanced in their settings:

1. Try again first: one to three more tries, a few seconds apart, before the step counts as failed. Good for outside services that sometimes answer slowly.
2. Carry on without it: the items go on as they came. Good for an extra detail you can live without, such as the weather for each site.
3. A failure path: drag from the block's red dot at its bottom to another block, such as an Email to yourself. It runs only when the block fails, with the error in {{error}}, and the run does not count as failed.

Mistakes that make failures worse

• Ignoring the first notice. You are told once per streak. A failure ignored on Monday is still failing on Friday, and you will not be told again.
• Running it all again after every failure. That fetches everything again. Retry from the failed block unless you changed an early step.
• Turning on Carry on without it for the step that matters. An employee that skips its own email every day looks healthy and does nothing.
• Building on a personal connection for a team job. If the person leaves, the flow stops. Use a team connection for anything a whole team relies on.
• Not testing after a big change. A test run (nothing is sent) before deploying catches most missing settings.

When you run it yourself

If a run you start from the editor fails, Explain and fix is pressed for you, so the explanation is already in the chat when you look. If you prefer to read the error first, choose Let me fix failed runs myself in that chat.