AP Automation Has an AI Governance Problem
Most readers will be familiar with the age-old adage: when one door closes another opens. While the intention of this phrase is to inspire optimism, in Accounts Payable (AP) it can be interpreted differently. The door that closed was manual effort but the door that opened, quietly behind everyone’s back, was governance. And this critical trade was never by design. The loosening reins on process oversight were just an unintended side effect of the automation.
At Springtime’s SSOW 2026 Split Plenary Session, “Accountable AI in Finance: Trust It, Trace It, Tie It To Results”, we explored this reality beyond hypotheticals.
Rob Bullen, Group Head of Global Business Services at British American Tobacco, explained an all-too common scenario that he had personally witnessed unfold:
An AI agent enforces your no-PO, no-pay policy to the letter. Every invoice without a purchase order gets stopped, exactly as designed. A supplier goes unpaid, a tax obligation lapses, and now you’re explaining to a court why your factory went dark.
He walked a Lisbon audience through this failure mode during our plenary session, and he again, wasn’t just speculating but issuing a warning. Because the AI in this scenario hadn’t failed but instead did precisely what it was told, and like most black-box automation, it did so without any critical reasoning applied. It couldn’t do the thing that humans had always done instinctively: recognize when following the rule creates the harm the rule was meant to prevent.
That gap is now the central problem in AP automation, and most implementation plans barely acknowledge it.
Where Accountability Breaks Down in Today’s AI Deployments
Our delegation attended many sessions on the topic of AI in AP but we kept ours purposefully candid. Rares Stoian of EY moderated; on the panel were Stefan Maschek, COO of Springtime Technologies, Bullen from BAT, and Victoria Slowikowska, a Partner in KPMG’s Global Business Services practice.
The abstract framed the issue politely on paper: agents are making financialhj decisions while accountability hasn’t caught up. What unfolded on stage had more edge.
Stefan opened with a distinction the industry often muddles:
“Accountable AI is not the same as auditable AI,” he said. “Auditable AI means you can reconstruct what happened. Accountability is much broader. It’s about outcomes and who owns them.”
You can have a complete audit trail with every step logged and still have no one accountable for the result. An auditable system shows its steps; an explainable one can show its reasoning; an accountable one has defined ownership. Yet most AP deployments, as enterprises repeatedly tell us, struggle to deliver even one of these cleanly.
An audit trail proves the system ran but it says nothing about whether it ran in the right direction or whether the logic was ever designed to be executed literally at scale. And most AI doesn’t pause to wonder if the results make sense.
Rob’s no-PO, no-pay story gave us a perfect example of this failure in action. Victoria Slowikowska took it further: the governance frameworks most AP teams are building obsess over what the AI did and ignore what it should have done instead. She highlighted three recurring issues:
- Controls bolted on top of agents
- Monitoring added after go-live
- Policy exceptions – where human judgement was critical – only surfacing after the damage is done
We've seen this before and it didn't end well: the RPA lesson
Rob named the parallel everyone in the room recognized but few said aloud. “In my lifetime, RPAs became a burden. Yes, there was some initial excitement… we built bots like they were going out of fashion and then we wondered how to fix them when they all started going wrong.”
The bot-farm hangover of 2018 to 2021 is being replayed now with sharper tools and bigger downside. Back then, automations were built fast, integrated badly, and left to marinate until the upkeep cost more than the saving. A broken bot strands an invoice. Now, a miscalibrated agent operating at scale can produce systemic payment failures, damage supplier relationships, and create compliance exposure that nobody notices until the quarter closes.
He went further, pointing to the way risk scales across complex AI ecosystems. In BAT’s P2P landscape alone, he listed five AI tools in active use: Hicks, Fair Market, RO, Xelix, and Springtime. Each adds another decision layer that must be understood in context. “I do see a bit of a risk in terms of technology proliferation,” he said. “How does one AI platform talk to another? Tech is almost becoming a risk for us.”
A GBS leader of his seniority, at an event sponsored by several of those vendors, making that statement is notable. The implication is clear: risk compounds with every additional integration.
Who Controls the Controllers? The Risk of Closed-Loop AI in AP
One of the most noteworthy moments of the session came directly from the floor. Sandy Khanna, VP of Finance at Mars Global Services and an SSOW advisory board member, asked: “If controls are critical, yet we have controls managed by AI (you’ve got AI doing the process, you could have agents doing the controls), how does that work?”
Victoria provided a very honest answer: the agent might find a loophole in the controls within three months and route straight around them.
The problem is baked into the design. Imagine your process agent, control agent, monitoring agent, all operating in one closed loop – something already being wired into some of Europe’s most advanced AP functions. If approval and control rely on the same underlying logic, the safeguard only holds if that logic does. And AI is highly effective at identifying its weak points.
At that point, transparency isn’t a feature but the only thing standing between you and outcomes you can’t explain or defend.
What True AI Governance Requires (and Most Teams Don't Have)
Genuinely auditable AI in AP needs three things most deployments only half-deliver:
- a time-stamped record of every decision
- a clear explanation of why that decision was made
- risk-tiered review aligned to financial and compliance impact
There’s also a scaling issue that goes ignored. Industry-wide, only about one in three invoices is processed touchlessly; even best-in-class teams sit around 49% (Ardent Partners’ 2025 ePayables benchmarks). The human layer is still doing the hardest work and it’s where accountability either survives or dies.
Bullen’s stated ambition at BAT – moving from 25% to 75% touchless – means deliberately shrinking that human layer. Do that without a risk-tiered review model underneath and you’re not removing oversight from low-stakes work; you’re removing it everywhere at once.
This is where the EU AI Act stops being abstract. As enforcement approaches for high-risk automated decision-making in financial services, regulators won’t ask whether your AI kept a log. They’ll instead ask who owns the outcomes, how exceptions are handled, and what happens when the agent is wrong.
For many AP teams today, those answers remain unclear.
The Hidden Cost of Building AI Without Governance Expertise
For teams planning or mid-build, Bullen’s RPA retrospective still applies: organizations that built proprietary bot estates spent years maintaining them. Those that relied on vendors who had already learned those lessons fared better.
Same logic, higher stakes, for AI. Building governance frameworks in-house is possible but it’s slow, costly and resource intensive. It also requires a level of expertise most AP functions don’t have.
The real question is whether the current vendor market offers anything better than faster processing: explainable decision logic, traceable trails, and accountability that holds up in front of an auditor.
In practice, very few platforms were designed with accountability at the core. Most generate audit trails after decisions are made – logs layered onto workflows that were never designed to explain themselves in the first place.
The alternative is systems where decision logic, execution, and auditability exist in the same layer. Where every decision is recorded together with its reasoning in real-time – not reconstructed after the fact.
This architectural difference matters because it’s what determines whether AI decisions can actually be defended when something goes wrong.
At Springtime, this is the principle behind Invoicetrack. Accountability isn’t added through monitoring or logging layers; it’s built into how every decision is made, recorded, and retrieved.
Bullen’s view on commercial models reinforces the point. BAT now links vendor fees to agreed touchless targets. He’s seen the opposite extreme on offer too: zero implementation cost, zero ongoing fee, 100% outcome-based pricing. He’s wary of it, because pure outcome models reward the wrong thing.
Pure throughput incentives prioritize volume over decision quality and therefore, his preference sits “somewhere in the middle.”. The implication is bigger than pricing: the buying conversation has shifted from features to accountability architecture, and most AP teams haven’t updated their evaluation criteria accordingly.
The One Questions Every AI Deployment Must Answer: Who Owns the Outcome?
The industry has a confident answer to “can we automate this?” It is still in search of one for “when it goes wrong, who owns it?”
In AP, where the decisions are financial, high-volume and compliance critical – that second question isn’t optional.
We’ve been in AP automation long enough to know that the teams that navigate these transitions successfully won’t be the ones deploying the most agents. They’ll be the ones that defined in advance:
- what a wrong decision looks like
- how it gets detected
- how it gets traced
- and who is accountable when it happens
That requires more than AI capability. It requires governance by design.
Most organizations are only beginning to confront this reality. The AI, meanwhile, cannot distinguish between a decision that is technically executable and one that is operationally responsible.
And yet…it’s already approving the invoice.
At Springtime, this is exactly the problem Invoicetrack was built to address. Not by adding more monitoring around AI decisions, but by ensuring every decision is made with accountability built in – fully traceable, fully explainable, and tied to a clear outcome.
If you’re evaluating how to bring AI into AP without losing control of it, you can explore our approach to Accountable AI in more detail here.