AI production
Check a small AI batch before making thousands
A queue can create a thousand items faster than a person can inspect them. That is useful only when the brief is clear and the outputs are worth keeping. Start with a small, varied sample and prove the checking process before increasing the batch.
Key takeaways
Generated does not mean approved.
A unique prompt can still produce a duplicate.
Set a stop rule before spending the full budget.
Give each item its own specification, approve a representative sample, separate generated from approved outputs, and stop when quality or spending crosses an agreed limit.
- Item brief
- Small sample
- Human approval
- Controlled queue
- Output checks
1. Describe what makes each item useful
Begin with the output, not the number of prompts. Define the required format, purpose, audience, and details that must be different. For a product description, that could mean the correct material, size, care instructions and tone. For an image, it could mean the subject, composition, aspect ratio and elements that must not appear.
Use one row per requested item and give it a stable ID. Keep the source facts separate from creative instructions. Missing facts should be flagged, not invented to fill the space. A repeated template can help, but it must carry the actual item details rather than a different number attached to the same vague instruction.
| Field | Example output |
|---|---|
| Item ID | CARD-017 |
| Purpose | A background for a planning worksheet |
| Must include | One abstract calendar shape |
| Must avoid | Words, brand marks and faces |
| Output format | A landscape image at the agreed size |
| Variation | Mint accent, left-to-right composition |
| Review status | Awaiting sample approval |
2. Review a varied sample
Choose examples from different parts of the planned batch. Include an easy item, an unusual one, a long brief, and a brief with something missing. Ten near-identical examples may tell you less than a smaller set that tests the difficult edges of the job.
Write down what passes before you review. For text, check facts, usefulness, tone and unsupported claims. For images, inspect requested details, unwanted text, visual defects and the actual export dimensions. If reviewers disagree, clarify the standard before producing more.
3. Track the item through distinct stages
Use separate states for Ready, Generating, Generated, Needs review, Approved, and Failed. An output that exists on disk has only reached Generated. It should not automatically become ready to publish or send to a customer.
Record the item ID, brief version, generation attempt and output location. When a brief changes, keep that version visible so you can identify which outputs need another check. If a call fails, preserve the row and its status rather than quietly dropping it from the batch.
4. Check both format and substance
A basic check can confirm that a file exists, opens, and has the required format. Text checks can look for missing fields or an unchanged placeholder. Duplicate checks can compare exact text or file fingerprints. Those checks are useful, but they do not establish that the content is accurate, appealing, or meaningfully different.
A different prompt can still produce a very similar item. Decide which similarities are acceptable and which require review. For factual content, compare important claims against the supplied source. Structured AI output can make fields predictable, but Google's documentation warns that correct structure does not guarantee semantically correct values.
5. Set a budget and a reason to stop
Estimate the cost of the sample and the main batch, then allow for an agreed number of retries. Log actual usage as the queue runs. Keep a spending limit separate from the target item count, so a run cannot keep retrying simply because it has not reached the target.
Example stop rule: a batch of 100 items pauses after three consecutive invalid results. Set the batch size and failure threshold around the work you are producing. Check the failed items and fix the cause before restarting.
- Cap attempts for an individual item.
- Pause when the same failure keeps returning.
- Pause when a required source or permission is missing.
- Stop when the agreed spending limit is reached.
- Require approval before scaling beyond the sample.
6. Deliver a batch someone can audit
Give the reviewer an index of requested items, outputs and statuses. Keep failed and rejected items visible so the final count is honest. A simple summary can separate Requested, Generated, Approved, Rejected and Still waiting; those numbers should reconcile.
For a small pilot, measure review time and the share of outputs accepted without changes. Include discarded outputs and retries when discussing cost. If checking the batch takes longer than making it manually, narrow the task or improve the brief before adding more volume.
Your ready-to-use checklist
0 of 6 checked
Make the next step clear.
Copy this batch brief
Free templateSources and notes
Sources checked on October 19, 2026.
- Google Gemini: structured outputs
Explains schema and format checks, plus validation of the values inside a structured response.
