Check invoices before an AI agent pays them
Have people confirm totals, line items and vendors match before automated payments go out.
How it works for this decision
Your agent sends
{
"question": "Does the invoice total match the purchase order?",
"context": "Purchase order: 40 units at $105.45, total $4,218.00. Invoice: 40 units at $105.45, shipping $180.00, total $4,398.00.",
"options": [
"Matches",
"Does not match",
"Unclear"
],
"verifiers": 3
}A verified person sees
Task$0.05
Does the invoice total match the purchase order?
Purchase order: 40 units at $105.45, total $4,218.00. Invoice: 40 units at $105.45, shipping $180.00, total $4,398.00.
You get back
{
"answer": "Does not match",
"agreement": 0.667,
"answers": 3,
"votes": {
"Matches": 1,
"Does not match": 2,
"Unclear": 0
}
}The result shown is an example. Try answering the task yourself.
Why a human check
Document agents misread totals, miss a duplicated line or accept an invoice that doesn't match the purchase order. Paying the wrong amount is expensive to undo.
A human glance at the invoice next to the order catches the mismatch an extraction model misses, before the money leaves.
- Send an image_url of the document along with the extracted fields.
- Ask a narrow question with clear options rather than an open review.
- Use verifiers: 5 when the amount is large.
Questions
Can verifiers see the document itself?
Yes. Add image_url with an https link to the document image; verifiers see it with the question.
How do I start?
Get a free test key at plures.si, send tasks from your agent in test mode, then add credit in your dashboard and live turns on.
More use cases
- Catch upset customers your support agent misreads“Is this customer upset?” Yes / NoHave people check tone and intent before an AI support agent closes, refunds or escalates a ticket.
- Verify refund and dispute decisions before money moves“What is this customer asking for?” Full refund / Replacement part / Information onlyAsk verified people whether a refund request is legitimate and what the customer is actually asking for.
- Human review for content moderation edge cases“Does this review break the rule against personal attacks?” Breaks the rule / Does not break the ruleSend the posts your moderation model can't decide on to verified people, and get a consensus answer back.