The AP audit problem starts before the AI

We’ve all heard the warning that AI in payroll will not survive scrutiny. The idea is based on architecture, drawing on the belief that these systems cannot explain how they reach their conclusions, and the new wave of AI research will capture that.

Administrative Matters

PwC has already spent approx one billion dollars building a platform of this type, which is being tested now as the company moves towards its intention to identify each customer in this way by 2028, analyzing each transaction instead of the product. The pressure is mounting for AP jobs that haven’t thought AI strategies through.

As it turns out, the warning is true, but it shows up in a black box. The reason why most AP AIs fail to evaluate is not necessarily because of the architecture but because you cannot evaluate a system that is not actually defined by the team.

Used properly, the arrival of AI in finance is not what jeopardizes research but what ultimately forces the discipline that should be there in general – and there are practical steps any organization can take before introducing AI into AP.

The black box is already in your AP

Ask the AP team why a particular invoice was paid on the 14th of the month, for that amount and the person approved it. Most of the time no one can tell. The letter arrives via email, sits in the inbox, is approved by everyone at their desk, and is deleted by someone using knowledge they never wrote before. There is no record of the cause. People call AI a black box but mobile AP is also a black box, built from input boxes and behaviors instead of code. AI does not create this misunderstanding but inherits it.

This situation is more common than one might think. Ardent Partners’ 2025 Survey puts the average invoice exception rate at 18.4% – about one in five invoices falling out of the standard flow to be corrected manually. The average invoice still takes 8.2 days and costs $9.84 to process although as the most efficient organizations do it about four-fifths less, the gap is implemented through process, not through technology.

The product has been researched for many years. Test one area, test the rest, and a bad pattern exists because most of it has never been explored. Read all the people around real time and this cover is gone. The right parts of your system are no longer independent and start to become research. Some of these cost real money; 76% of US organizations experienced attempted or actual fraud last year, according to a 2026 AFP survey, with duplicate payments and mistaken entries among the biggest gaps in fraud.

Governance is a structural problem before it becomes a technical problem

The quick “show me the payment from the 14th and everything after it” only works if there is a specific, structured plan behind it. A process that lives as a habit in people’s heads cannot be transferred to a machine but neither can it be transferred to a researcher.

This is also why AI continues to take hold in financial institutions. MIT’s NANDA study last year analyzed the deployment of 300 companies and found 95% did not create a measurable impact on the bottom line. The reason, the researchers made clear, was organization, not technology. Companies never put together the tools for how to do the job. Money has no future. Ardent’s 2025 numbers show only 44% of AP teams are using AI at all, and most of that is scheduling emails via a chatbot – a long way from letting design play a central role in AP.

The pilot that didn’t reach production and the research that couldn’t be explained came from the same place – a system that was never written.

Done properly, automation makes the AP more accessible, not less

Plug a template into the chaotic writing and the search will reveal everything. Define the system first and automation can catch up. Good implementation forces the team to write down the rules, agree on who accepts what, and decide what happens with any exceptions. The result is structured information instead of PDFs in the inbox and a record of each decision instead of a shrug.

For example, take a multi-site business that operates large fleets. Their monthly fuel bill runs to 30 pages of fuel in multiple vehicles and cost centers and they use up to two days of manual reconciliation in two disconnected systems. The checking issue was not very fast but hundreds of purchase orders sat open forever in their workflow because closing them was a manual step that the team often skipped. With the right AI tools, every PO is automatically closed and every invoice posted in the ERP along with the complete accounting balances. This is what an AP interface looks like in practice.

A manual AP has never been made to explain itself, but automation done well can be a turning point. By 2028 most companies will be audited by trade. Better to install the system now than to have an inspector find out that you never had.

Four things to adjust before you add AI

The work that makes AP assessable is not skill but authority, and it must come first. Four things are worth explaining before any product approaches the payment process.

It starts with the right to decide, which must be clear. For example, the finance director can approve spending without the board’s signature up to a certain amount and the AP team leader can release funds under a different amount without two approvals. Most organizations never actually map and instead manage the culture.

Next comes math. Name the person who owns the consequences when something goes wrong, not just the team. If AP automation fails, the answer is not “money” or “ops” but man.

With accountability comes monitoring and improvement. Determine what is reported, to whom and on what basis. Importantly, it must also clearly state the cause of the increase. This is where government can fail. Organizations organize fully staffed committees and never really explain what causes something in the first place.

Finally, enter the power to change. New tools, new connections and new rules all require an internally evaluated approach. If not, the government is destroyed when the situation changes because people are silent and surrounded by it. If organizations can do this work now, they will control how the research is done. Wait, and this decision will be with the researcher instead.