This is the moment an automation that worked perfectly starts failing without warning. The trigger has not changed, nor has the processing. What changed is the volume, and nobody had thought about it.
Definition
Batch processing means grouping several items to handle them together, rather than starting a full run for each one.
The difference shows as soon as you count. A thousand items handled one by one is a thousand runs, a thousand calls, a thousand billing lines. The same thousand items in batches of fifty is twenty runs.
Many tools run a node once per item received without saying so clearly. A Node receiving a thousand rows runs a thousand times. That is the first thing to check when a bill doubles for no apparent reason.
The three walls volume reveals
The quota
Every API limits calls per minute. A burst exceeds them and gets refused. The automation does not necessarily crash: it handles the first hundred and silently abandons the rest, which is worse.
The cost
On platforms billed per operation, volume is paid directly. On AI models, each item handled separately also resends the full instruction, so you pay for the context a thousand times: see Token.
The time
A thousand runs at three seconds each is fifty minutes. If the processing must finish before opening hours, grouping stops being theoretical.
Best practices
Group, but not too much. Batches of twenty to a hundred suit most cases. A batch that is too large becomes impossible to replay cleanly when it fails halfway.
Space out the sends. Most tools let you insert a pause between batches. One second of waiting is often enough to stay under quotas, at no cost.
Track what has been handled. A batch failing on the thirtieth item must resume at the thirtieth, not the first. Without that record you reprocess everything and create duplicates.
Test at real volume. A scenario validated on five items proves nothing. Test on a peak day's volume, not a quiet Tuesday morning.
Watch for partial failure, the most insidious of all. Three hundred items handled, seven hundred skipped, and a green "completed" status. Always check the number handled against the number received, and alert on the gap.
Frequently asked questions
How do I know if my tool processes item by item?
Run it with three items and look at the execution log: either you see one run or you see three. That two-minute check avoids four-figure surprises.
Can everything be processed in parallel to go faster?
It is possible and it is the best way to saturate the other service's quotas. Parallelism is a matter of dosage: two or three simultaneous runs are almost always enough.
Does batch processing work with an AI model?
Yes, and the saving is significant: one instruction for twenty items instead of twenty identical instructions. Some providers also offer a deferred batch mode, considerably cheaper when the result is not needed within the second.
How do you prepare an automation to scale?
By asking the maximum volume question when writing the Automation scenario, before building. Our n8n course covers that sizing from the first scenarios, rather than at the moment things break.