Batch jobs

Batch jobs let you run any AI operation across a large set of products or categories. The work runs in the background, so you can keep using the admin while the job processes.

Job lifecycle

Every job goes through this state machine:

pending — created, waiting for the worker to pick it up
running — worker is processing items
completed — finished, all items processed
failed — fatal error before completion
cancelled — manually cancelled from the UI
suspended — paused by a recoverable provider error (e.g. quota exceeded). Top up your AI account and use Resume to continue from where it stopped
awaiting_manual — job is in manual review mode and waiting for the next user step
blocked — an automation step job waiting for the previous step to finish before it runs (shown as Waiting for previous step)
await_review — the job finished its AI work but produced suggestions that must be reviewed before the automation continues — e.g. Image Finder running inside an automation (shown as Awaiting review)

Transient provider errors (a brief network blip, a short rate-limit) are retried automatically a few times before a job is suspended, so you usually never see them.

The progress bar on the jobs list shows processed / total. The detail page breaks down accepted, rejected and applied counts per item.

Each suggestion also has its own status:

pending — produced by the AI, needs a decision
accepted — user accepted but not yet applied
rejected — dismissed (audit-only, no change)
applied — accepted and written to the product/category
confirmed — AI examined the field and the proposed value matched the current one. No change to apply, but the field was processed
reverted — an applied suggestion has been rolled back from its snapshot
failed — the suggestion itself errored (e.g. apply failed against the entity)
target_deleted — the product, category or image the suggestion targets no longer exists, so it can’t be applied

Creating a job

Go to AI Assistant → Jobs → New job.
Pick the type (Generate SEO, Translate, Extract measures, etc.).
Filter the records — by category, manufacturer, sales channel, active state, “missing field” predicates. The wizard shows a live preview count so you know the scope before continuing.
Pick the AI connection, target languages and any type-specific options (preset for Improve, source mode for Generate Content, etc.).
Look at the cost estimate. Adjust the scope if it is too high.
Pick a run mode:
- Auto — the worker processes every matching item in the background.
- Manual review — the wizard creates the job in the awaiting_manual state and opens the review modal; see below.
Start.

Manual review mode

Manual review mode lets you step through items one at a time, edit each AI draft inline before saving, and hand the rest of the job off to the background worker at any moment.

The review modal walks the job one item at a time:

Next — picks the next eligible item, runs the AI, and shows you every draft side-by-side with the current value. Drafts where the AI’s proposed value already matches the current one are recorded as confirmed, so they appear in the audit trail without needing a decision.
Save — applies the (optionally edited) drafts for the current item. Anything left unticked is rejected, and the item won’t come up again.
Skip — rejects all drafts for the current item and moves on. Useful when you don’t want either the AI proposal or the existing value to win.
Resume in background — hands the rest of the job to the background worker, which continues from where you stopped and skips every item you already handled.

If you close the modal mid-review without Saving or Skipping, the pending drafts stay on the job. The next time you open the modal on the same item, those drafts are reused — no second AI call, no duplicates.

Reviewing results

When the job is running, the jobs detail page polls every few seconds. As suggestions land, they show up in the Suggestions tab with confidence, cost, reasoning and per-row actions.

You can review item by item or use bulk actions:

Apply selected — applies the picked suggestions to the products
Reject selected — marks as rejected (audit-only, no change)
Revert selected applied — undoes already-applied suggestions, restoring the snapshot stored on the suggestion row

Recovery

If the background worker crashes or stalls, a job can get stuck in running. The plugin re-queues such jobs automatically every few minutes, so they recover on their own. A job that keeps needing recovery points to a deeper problem — check the API logs for that job.

Recoverable provider errors (rate limit, payment required, quota exceeded) move the job to suspended instead of failed. Fix the underlying issue (top up, raise the quota) and click Resume — processing continues from the next unprocessed item.

Cancelling

Click Cancel on a running job. The worker stops at the next item boundary. Already-produced suggestions remain on the job and can still be reviewed and applied.