Auto-screen new Airtable submissions before they go live

Watch my Airtable submissions table with a poll trigger on the new_record event. Every time a new row appears, run this workflow once for that record. The table holds user-submitted content: community posts, reviews and public job listings. Each row has the submission text, the submitter, and the fields I want written back.

For each new record, take the submission text and send it to Mistral AI Moderations to score it against Mistral's moderation policy categories. Read the full set of category scores off the response, then work out which category scored highest and what that score was.

Also send the same submission text to Mistral AI Classifications to assign a topic or category label to the submission. Use the highest-scoring label from that response as the topic.

Use a single configurable score threshold to decide the outcome, defined as a constant at the top of the workflow and defaulting to 0.5. If the highest policy category score is at or above the threshold, the moderation status is needs review. If it is below the threshold, the status is auto-approved. If the moderation call fails, returns no scores, or returns anything I cannot read a numeric score out of, set the status to needs review rather than auto-approved. Never auto-approve a submission that the moderation step could not actually score.

Then use Airtable Update Record to write four values back onto that same record: the moderation status (auto-approved or needs review), the name of the highest-scoring policy category, the numeric score for that category, and the topic label from the classification step. Update only those four fields and leave the rest of the record untouched.

Only when a submission comes back as needs review, post an alert with Slack Bot Send a Message to my moderation channel. The message should name the submitter, include a short excerpt of the submission text (roughly the first 200 characters, trimmed on a word boundary), state the policy category and score that tripped the threshold, and include a direct link to the Airtable record built from the base id, table id and record id. Clean submissions are never posted to Slack, so the channel only ever contains things a human actually has to look at.

Keep the score threshold, the Slack channel, the Airtable base and table ids, and the excerpt length easy to change at the top of the workflow. Keep the logic strict and simple: score the text, compare against one threshold, write back four known fields, and notify only on the flagged path.