Where AI Agents Can Actually Pay Off In Procurement And AP
Shaz Khan is the CEO of Vroozi, a procurement platform helping businesses modernize how they manage spend and supplier relationships.
gettyFor as long as I’ve been in the industry, procurement and accounts payable (AP) teams have heard the same promise: that their operations will finally run themselves. New waves of technology keep claiming they’ll finally deliver it, yet the same people are still chasing down approvals, tracking down why an invoice doesn’t match a PO and fielding calls from suppliers asking when they’ll get paid.
AI agents are the first technology I’ve seen that could genuinely close that gap, and the reason is simple. Older automation follows rules someone wrote down in advance. If a situation falls outside those rules, a person has to step in. An agent can actually look at a request, pull information from more than one system, figure out what needs to happen and take the next step itself. Procurement and AP are full of exactly the kind of work that fits, with high transaction volume, plenty of documents, established policies and the same handful of exceptions coming up again and again.
I want to be clear about what this means for procurement and finance teams. The goal is to free up their time for the decisions that need a person’s judgment, not to replace the people doing this work.
If a company needs 25 laptops for new hires, its employees are logging into a system, searching a catalog, picking a supplier, entering the right accounting codes and routing it for approval. An agent can take that same request in plain language, check preferred suppliers and negotiated pricing, confirm it meets policy, figure out who needs to sign off and build the requisition without the employee touching half of those steps.
Structured intake has been part of procurement software for years. What’s changed is how much of the surrounding decision-making the system can now handle on its own.
Procurement teams negotiate hard for good contract terms, and then those terms don’t always show up in what actually gets purchased. An agent can sit in the background comparing requisitions, POs and invoices against the contract itself, flagging a purchase from the wrong supplier, a price that doesn’t match or a missing volume discount.
McKinsey has cited a life sciences company with about $4 billion in annual R&D procurement spend that ran an AI review covering roughly 10% of that spend and found 4% in verified leakage, more than $10 million. That means catching a problem before the payment goes out instead of finding it three months later during a quarterly audit.
Invoice capture has been more or less automated for a while now. The hard part has always been everything that happens after.
A traditional system spots the exception and hands it to a person to sort out, but an agent is capable of doing that sorting itself. It can pull the invoice, the PO, the receipt, the supplier record and the contract; work out why the numbers don’t line up; and decide whether it falls inside the company’s own tolerance rules. Routine ones get closed out on their own, while the ones that carry real risk still go to a person. But by then, the evidence is already assembled instead of sitting in someone’s inbox for a week.
Take a three-way match failure as an example. A $52,000 invoice comes in against a $50,000 PO. A conventional system may just flag the mismatch, but an agent can tell you that expedited freight was added, that the contract allows it and that it falls inside your tolerance threshold. That’s a completely different starting point for the person reviewing it, since they’re checking a decision instead of doing the research themselves.
Every AP team fields the same supplier calls. Did you get my invoice? When am I getting paid? Why was this rejected? An agent connected to procurement and ERP data can verify who’s asking, find the invoice, explain where it stands and move it forward, whether that means pinging the employee who never logged a receipt or updating the workflow once they respond. A chatbot can explain what’s wrong, but an agent can fix it.
That distinction matters because even with years of investment in AP automation, a significant share of invoices still require human intervention. Ardent Partners puts straight-through invoice processing at just 32.6% industry-wide, versus about 49% among best-in-class AP teams. The opportunity for agents isn’t simply to automate another step in the process; it’s to close some of the gap between invoices that enter the system and invoices that can move through it without someone having to intervene.
That optimism comes with real conditions attached. Before an agent executes anything on its own, procurement and finance teams need to draw some clear lines. Dollar thresholds are the obvious one, where a small requisition or payment clears automatically, but anything above a set amount still needs a person’s sign-off. New suppliers, changed bank accounts and first-time contract exceptions should route to a human by default, since that’s exactly where fraud and costly mistakes tend to hide.
Permissions matter just as much. An agent that can flag a discrepancy is very different from one that can release a payment, and most organizations aren’t ready to hand over that second kind of authority without a lot more testing. I’d rather see agents earn broader permissions gradually, starting with recommend-and-escalate before moving toward act-and-report.
None of it works without a clear audit trail. Every decision an agent makes, and the reasoning behind it, needs to be logged and reviewable, the same way you’d want a paper trail from a person handling the same transaction.
Pricing is shifting, too, moving from a flat license fee toward paying for outcomes an agent actually delivers. Ultimately, getting these guardrails right is what lets leaders confidently decide what to delegate and what to keep human-led.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
