Run an action in bulk
A batch action run takes one published action and runs it over many inputs at once — many files for an extract action, many records for a generate or fill action. Every input becomes a row in a review grid, and nothing is committed until you approve it. Extract approvals create records; generate and fill approvals add documents to the project’s files.
A run always lives inside a project, and it uses that project’s files and records. Florent does the reading and drafting; you approve. For the citation chips and confidence scores you’ll see while reviewing, see Reviewing AI work — this page shows them in the batch grid, but that page owns the full explanation.
You need permission to start playbooks to launch a run, and permission to approve to act on rows — see Roles and permissions. An action must be Published before you can run it — see Actions overview.
Launch a run
There are two ways to start a batch run. Both land you on the same launcher.
From the project
From the action
Pick the action and inputs
The launcher has two steps: choose the action, then choose what it runs over. The input picker swaps automatically based on the action’s kind — you don’t pick the axis yourself.
Choose an action
Pick from the Action dropdown. Only published custom actions appear. If the list is empty, there are no active actions yet — publish one first (see Actions overview).
Select the inputs
- For an extract action, you get a Files picker — a checkbox table of the project’s files. Tick the files to process, or use Upload more to add files first.
- For a generate or fill action, you get a Records picker — a checkbox table of the bound record type’s records. If the action uses a free-form template, choose a record type first. A generate action that combines records into one document shows the note “This action creates one document covering every record you select.”
One run processes at most 1000 inputs. If your project has more, the picker shows the first 1000 with a note. Items process 4 at a time, so a large run fills the grid gradually — you can leave the page and come back.
Review the grid
The grid is where every input becomes a row you review. The heading is the action’s name with a run-status badge. The summary line reads, for example, “12 approved · 3 to review · 1 failed · 16 files” and updates live as items finish.
Each row shows:
- A status badge for the item.
- A warning count when something needs a closer look.
- A grounding/confidence chip that points back to the source — see Reviewing AI work.
Columns are the action’s output fields. Generate and fill runs lead with a Records column (the source record) and a Document column (the generated filename) instead.
Approve in bulk
You have two fast paths and a per-row path.
Approve all clean
Approve all clean approves every row that needs review and has no warnings. Rows with warnings are skipped on purpose — open and handle those by hand. The button’s count tells you exactly how many clean rows it will approve.
Select rows for the bulk bar
Tick the checkbox on one or more rows (or the header checkbox to select all) to reveal the bulk bar:
- Approve selected — approves the chosen rows that need review.
- Retry failed — re-queues the chosen failed rows.
- Clear — clears your selection.
Edit a value in the grid
For an extract row that needs review, you can fix a wrong value inline in its cell. The value re-validates when you save. Editing a cell removes its citation chip, because the value is now yours, not the AI’s.
Use Export CSV in the header to download every row — filename, status, each output field, and the record id — to share or archive the results.
Review one row at a time
Click Review on a row (or click a field’s citation chip) to open the full per-row review dialog. It fills the screen with two columns: the source on the left, the fields on the right.
Check the source
The left column shows the source. For an extract row, it’s the source PDF, scrolled to the selected field’s citation with the cited region highlighted. For a generate or fill row, it’s an inline preview of the generated document. Word documents can’t preview here — use Download to open the file.
Verify each field
The right column lists one card per output field: the label, the value, a confidence score, and a count of cited sources. Select a field to see its citations and jump the source viewer to that evidence. Read the cited text and confirm the highlight says what the value claims. The full citation behavior is covered in Reviewing AI work.
Fix anything wrong
While the row needs review, the values are editable. Make your changes, then click Save changes. You must save before you can approve — Approve stays disabled while there are unsaved edits.
Approve or reject
- Approve commits the row. For extract, it creates the record (with each field’s citation carried as provenance). For generate and fill, it moves the generated document into the project’s files.
- Reject opens a reason box. Type why it’s wrong, then click Confirm rejection. For generate and fill, rejecting deletes the generated file.
- Retry re-queues a failed row.
Approving is what makes a result real. An extract approval creates a record; a generate or fill approval adds a document to the project’s files. Rejecting a generate or fill row deletes the produced file — the row stays as a record of what happened.
Statuses you’ll see
Run statuses (the whole run)
Row statuses (per item)
Run history
The Project actions tab keeps a history of every run in the project. Each row shows the Action, its Status, Progress (approved out of total), how many rows are still To review, and when it Started. Click any run to reopen its review grid.
When a row fails
A failed row shows a short message explaining what went wrong. Most are fixable — correct the cause, then retry the row (or use Retry failed in the bulk bar).
The file or record went missing
- This file is no longer in the project — it was deleted after the run started. Re-upload it and run again.
- A record this document needs was deleted from the project — re-add the record, then retry.
The document couldn't be read or converted
- We couldn’t read this document. It may be corrupted or password-protected — fix or replace the file.
- We couldn’t convert this document to PDF — usually temporary; retry.
- Generating this document took too long and was stopped — usually temporary; retry.
The action or its template needs attention
- This action is no longer available — it may have been deleted.
- This action is no longer active. Activate it again, then retry.
- This action’s template is missing or no longer active — check the action’s template, then retry.
- The template changed after this action was last activated — re-activate the action, then retry.
- The template has placeholder errors — fix the template, re-activate, then retry.
- The selected records don’t match the record type this action’s template expects — pick records of the right type and run again.
The result couldn't be saved
- A required field was empty, so we stopped rather than generate an incomplete document — fill in the record, then retry.
- The generated document is larger than 50 MB, the project file limit — the document is too big to save.
- This project is out of file storage — free up space, then retry.
Where to go next
Set up an action that reads files and pulls out structured data.
Set up a Word template and the generate action that fills it.
How citations, confidence scores, and field review work.
Where approved documents land and how to find them.