Preparing a batch of payments used to mean a person checking each invoice, confirming it matched what was approved, and grouping everything by hand. Coupa just showed how much of that work software can now do on its own. The company’s Payment Batch Creation Agent ran 14 payment batches on its own, covering 2,395 individual payments worth $20.1 million, in its first five weeks in production, Coupa said in a Thursday (Aug. 20) release. More than 450 customers now use Coupa’s AI agents in production. That agent does everything up to the last step, but it still doesn’t have the final say. A human, or a separate system, has to approve the release before the money moves.
That distinction is the real story. An AI system reading an invoice or pulling a purchase-order number is handling information. An agent deciding which approved invoices belong in a payment run, and preparing that batch for release, is participating directly in the transaction itself. The economics back up that change. One Coupa customer said the agent cost about $27 in AI credits to run and saved an estimated $2,000 in staff time in a single week. Coupa also reported requisition cycle times falling by as much as 50%, though those figures are self-reported and may not hold for every customer.
Preparing a Payment Is Different From Sending One
The larger opportunity is not just processing an invoice faster. An agent can validate an invoice, confirm approvals are complete, check payment terms, find an early-payment discount and place the transaction into the right batch, all before a person looks at it. Humans may come to handle only the transactions that don’t fit that pattern: a supplier that suddenly changes its bank account, an invoice much larger than usual, or a new beneficiary nobody has paid before.
That boundary between preparing a payment and having the authority to send it is already shaping how card networks build infrastructure for agents. Mastercard has built what it calls Verifiable Intent, a system that creates a tamper-resistant record of exactly what a person authorized an AI agent to do before that agent acts. “As autonomy increases, trust cannot be implied,” Pablo Fourez, Mastercard’s chief digital officer, said. “It must be proven. And if something goes wrong, everyone needs facts, not guesswork.”
Visa is building a similar Trusted Agent Protocol, PYMNTS has reported, letting a merchant confirm it is dealing with an approved agent and cryptographically linking that agent to the payment details being used.
The Hard Problem Is Proving Who Did What
Coupa is also tackling that same accountability question from another angle: opening the door for third-party AI tools to reach directly into a company’s live financial data. Its latest release includes more than 30 connections spanning procurement, invoicing, contracts and expenses that let external AI systems read and act on that data without a company building a custom integration first. Access still inherits whatever permissions a human employee already has, and remains auditable, according to Coupa.
That points toward a future where the agent preparing a payment may not even live inside the software that ultimately controls the transaction. The $20.1 million Coupa has processed this way is still small against global corporate payment volumes, but it shows the shift is no longer theoretical. Seventy percent of firms already use at least one AI tool to manage some part of their cash flow, according to PYMNTS Intelligence’s “Time to Cash” report, and that number is only going to grow as more of the preparation work moves to software.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
