Based on an interview with Global Process Owner Adam Mroczkowski

Ask about AI in purchase-to-pay and you usually get two very different answers.

In one camp, vendors describe a finance function that basically runs on autopilot: invoices processed without human effort and exceptions resolved before anyone even notices. In the other, skeptics write the technology off as not ready, a conclusion usually drawn after a single pilot stalled and got abandoned.

Most P2P teams sit somewhere in the middle. Recent survey data shows that only 17% of finance teams currently use AI in a core workflow, while 45% remain stuck in limited pilots. What holds the rest back usually isn’t a lack of interest. It’s simply not knowing where to start, or which use cases actually hold up once you take them out of the demo environment.

So we spoke with Adam Mroczkowski. He is a Global Process Owner at Rockwell Automation with over a decade in P2P across large, complex organizations, and a regular voice on finance-transformation stages. He is someone who does the day-to-day work while tracking where the function is heading. We skipped the futurism and asked him where AI is doing real work in P2P now, what is still just a demo, and what he would try first.

“Finance was one of the first functions to start trying AI. Last year, we were talking about it, this year we’re doing it.”

Some of what follows is what Adam sees across the industry. Some he is testing in his own team, and some is what other finance leaders are putting into practice right now.

Start by pointing AI at your policies

If you point AI at a process and ask it to run the steps or guide your team through them, it will get things wrong more often than not. Before automation can work, you need to know if the process is documented consistently across entities, or if the rules drift from site to site.

The place to start is upstream, in the rules governing the whole system: your approval thresholds, payment terms, and local site procedures. These documents were written at different times, by different teams, across different countries. Over time, they drift. One site pays a supplier on 30 day terms, while another pays the exact same vendor on 60. One entity accepts invoices with no PO at all, while the next blocks them.

None of this is obvious from inside a single document, and comparing every policy side by side by hand isn’t realistic. This is the exact work AI is built for.

“One policy telling you to go left, the other telling you to go right. And the third one saying it depends.”

In Adam’s view, this is the single most common lesson learned on AI projects today: teams discover their procedures never aligned in the first place.

• Payment terms and discounts. The same supplier sits on net-30 in one entity and net-60 in another, or an early-payment discount is captured in one region and left on the table everywhere else. A consistent relationship with that supplier is worth more than a collection of separate ones. Handled at group level, the same spend could earn volume discounts and stronger terms across every entity.
• Three-way-match tolerances. One entity clears anything within 2%, the next within 10%. Read across sites, the useful question is not which is stricter but how many exceptions each threshold actually raises: one that almost never triggers is a rubber stamp rather than a control.
• PO thresholds and approval limits. Spend over €5k needs a PO and two approvers in one business unit, while in another the bar sits at €50k with one. Most of the spend in the second unit never hits a control at all.
• Spend categorization. The exact same purchase gets logged under completely different categories or GL accounts depending on the site. What one entity treats as standard IT operating spend, another tags as external consulting, throwing off spend tracking and approval routing before the invoice is even processed.
• Vendor master data and onboarding rules. Each entity requires something different to create a vendor, which is the exact root of the duplicate supplier records and misrouted payments that show up downstream.

Test it this quarter: Take one process across two or three sites; payment terms is a clean place to start, and gather the policy documents that govern it. Hand them to a model and ask it to lay out how each site’s rules compare and where they conflict. What comes back is a comparative picture you could never have assembled by hand, and a short list of the gaps worth closing first.

Clear repetitive work

The next move is to take the dull weight off your team. More than a fifth of AP staff time spent answering supplier inquiries, according to research done by Ardent Partners. Most of it is the same loop: read the question, look up the invoice or payment status, write the reply. AI does that loop now. It reads the incoming email, pulls the status from your system, and drafts the response. Your team checks and sends, a quick review in place of a from-scratch lookup.

As Adam points out, caution matters here. AI does the visible part, but it only delivers because it rests on years of process standardization. That underlying clean process is what makes the output worth trusting.

“Do your homework before you use the tool. Do it properly and you can trust the output enough to take the human out of the loop. And that’s where the real productivity is.”

Test it this quarter: Scope a pilot to one narrow, safe category, payment-status questions, and route anything that involves a commitment, an early payment, a terms change, or a dispute straight to a person. Keep a human on every send, and watch two things:

1. Data correctness. A wrong status in the system becomes a wrong answer to a supplier.
2. Remove the room for hallucination. Let the AI do its job, interpret the query, and come to a conclusion, but keep the final answer to a couple of standard responses per business scenario. Don’t let it be creative, and don’t let it make commitments. Get that guardrail right and this is the cheapest hour-saving you’ll find.

Build a chatbot that actually works

Most internal chatbots run on a fixed script. They manage the generic questions and fall apart on anything specific. The version worth building is a knowledge hub: point a model at your own procedures, policies, and contracts, and let anyone query it directly. They get an answer in seconds, and finance stops fielding the same questions all day.

Every global team knows the problem it solves. Hours lost in SharePoint, the search for who owns what, the wait for someone three time zones away to answer something that was written down all along.

It is a safe place to start, too. Gartner’s late-2025 survey found knowledge management was the single most common AI use case in finance, ahead of AP automation and anomaly detection. It works, it keeps working, and that makes it worth early effort.

Test it this quarter: pick the one process your team gets asked about constantly, the thing the wider business and every new joiner keep pinging you about, and point a model at that documentation alone. One domain, tight scope, every answer grounded in your own documents. Prove it there before you let it widen.

Make sense of unstructured data

For years, P2P teams have spent enormous effort trying to force structure onto their suppliers. Submit invoices in this format, use this portal, fill in this template. With several thousand suppliers, it never fully works, and Adam calls full standardization mission impossible.

Handling unstructured data is what changes that. AI reads whatever a supplier sends, an invoice, an email, a statement,  in whatever shape it arrives, and turns it into something clean on your side. The pressure to make every supplier conform to your systems comes off.

It also reopens a question most teams settled years ago. Adam’s example is supplier onboarding, which usually comes bundled inside a third party’s service. With AI carrying the manual load, there is a real case for bringing it back in-house, where your own people understand the internal rules and the real friction far better than a BPO ever will.

Test it this quarter: take one slice of supplier onboarding you currently outsource and run it in-house, with AI handling the document and data work. Keep it small enough to fail safely and real enough to show you whether your own team, with the tool, can do it better than the BPO.

Find where you can save

This is where AI starts paying for itself. Most companies cannot see their own spend clearly. It sits in different systems across regions and entities, the same supplier spelled four different ways, the same category labelled differently in every country.

Pulling that together and cleaning it up is the work that used to eat analysts’ weeks, and it is the work AI now does well. It connects to the systems, recognizes that “Bosch GmbH” and “Robert Bosch Ltd” are the same supplier, and lines the spend up so you can finally compare how and where you buy, and at what price. Adam sees the opportunities across supplier panels, payment terms, DPO, and terms-and-conditions compliance.

Take payment terms, for example. You pay the same supplier 30 days net in Poland and 60 in Germany. Tools have flagged gaps like that for years. What AI adds is the context around the flag: it reads the terms across every entity at once and points you to where it is worth looking. A person still decides which gaps are real, since the reason usually sits in the background where only a person can see it. What changes is that you start from a map of where to look.

Test it this quarter: point a model at the payment terms for your top suppliers across your entities and ask where the same vendor sits on different terms in different places. Each gap is a question worth asking: a chance to free up cash, or a process that drifted. Both are worth finding, and neither surfaces on its own.

How to start when you’re already maxed out

The teams who would gain most from AI are usually the ones with no time to set it up, buried in the close and the reporting, the very work AI could take off their plates. So the first step is the dullest one: protect a small slice of time.

Then be deliberate with it. Run one small pilot on real work, and skip anything done for the sake of having AI in the building. Decide up front where a person stays in the loop, especially where the output is a judgment call, because a confident answer and a correct one are different things.

And find your champions, the one or two people who are hungry for this and will drive it for you. Give them the time and the cover.

The teams innovating are the ones who carved out the hours, started small, and let the curious people run.