About six months ago, I sat down with the CFO of a massive industrial conglomerate to discuss their invoice-to-pay challenges. This veteran finance leader, who had climbed from accountant to C-suite over three decades, delivered a sobering assessment:
“We run accounts payable today the same exact way we did when I started 30 years ago.”
I assumed this was hyperbole until we started examining their operations. Yes, they had cloud-accessible ERPs. Yes, they’d spent millions on consultants promising “automation” and “transformation.” But the operational reality remained unchanged: hundreds of staff were still manually processing documents, just digital ones instead of paper. They were performing the same core tasks with marginally fancier tools.
This wasn’t an isolated case. Another finance transformation leader told me:
“We spent years moving the majority of our suppliers to an EDI system — but we still have dozens of accounting specialists on staff to handle the 30% of transactions that don’t clear automatically.” The reason? Often, it’s trivial discrepancies in parts codes. A missing dash. A trailing space.
Let that sink in. It’s 2025, and the supposedly “automated” EDI solution collapses when confronted with a missing hyphen. Finance systems with less intelligence than the average smartphone are running trillion-dollar global supply chains.
The Process Crystallization Problem
What we’re witnessing isn’t a failure of technology capability but a case of emergent process lock-in. As organizations grow, they develop operational patterns that gradually crystallize into rigid structures. These structures aren’t designed — they emerge through a kind of organizational natural selection where “what works” becomes “what must be done.”
The dysfunction follows a predictable template:
- Data becomes trapped in fragmented systems — digital fiefdoms jealously guarding their portion of enterprise reality
- Workflows calcify around institutional knowledge — critical expertise locked in employees’ minds or buried in desktop procedure manuals
- “Automation” devolves into manual data gymnastics — exporting from one system, normalizing data in Excel with hand-built macros, and uploading to another
The result? Extended cycle times. Bloated costs. Frequent errors. And processes that unravel when key personnel call out sick.
Why Integration Bridges Keep Collapsing
Enterprise software rigidity wouldn’t be so problematic if integrations weren’t so brittle. But they are, and for structural reasons.
When two systems need to communicate, they require perfect alignment across data formats, field mapping, transaction sequencing, and state management. The subtle impedance mismatches between different systems manifest in pernicious bugs and ever more complexity under management in the glue code layer.
This is why EDI implementations — technology from the 1970s — still fail on missing hyphens. The integration assumes perfect data conformity because the alternative (true resilience) would require actual intelligence.
And as a result, every new SaaS point solution and RPA script adds another potential failure point.
- Point solutions create automation islands outside the ERP
- RPA tools layer brittle scripts over existing interfaces
Each connection becomes another bridge that must be maintained, patched, and eventually rebuilt when the underlying systems change.
Neither works because both typically worsen the complexity: adding another integration point, another data silo, another Excel cleanup ritual. The debt compounds.
The Human Complexity Sink
What happens when software is rigid and integrations are brittle? Complexity gets dumped onto humans.
This manifests as:
- BPO contracts where costs scale linearly with transaction volume
- Elaborate SOPs that expand with each edge case
- Desktop procedures that grow to hundreds of pages
- Managers who spend their days triaging exceptions
- Regional variations of processes for essentially identical work
Human labor becomes the flexible component that absorbs all the variance that software cannot handle. It’s enterprise’s preferred complexity sink — because humans adapt faster than software can be rewritten.
The problem is that this approach doesn’t scale. It’s why finance organizations grow linearly (or worse) with transaction volume, when they should be growing logarithmically at most.
Why Current AI Approaches Fall Short
Enter generative AI, which promised to eliminate this human complexity sink. Train agents on your SOPs, desktop procedures, and institutional knowledge, then let them handle the variance that humans currently manage.
But first-generation AI approaches are falling short for several structural reasons:
- Context fragmentation — Agents lack access to the complete information universe needed for decisions
- Guardrail complexity — Creating effective policies that prevent harmful actions quickly becomes a project as complex as the original automation effort
- Decay of coherence — Agents drift from desired behavior as edge cases accumulate
- Security control limitations — Enterprise-grade controls over AI actions remain primitive
Most critically, these agents operate as glorified macros — taking action based on predefined patterns but without a true understanding of the process context or business objectives.
The result? Agents that can handle the simplest 70% of cases but fail on the complex 30% — precisely where human judgment was previously required. We’ve recreated the same EDI problem, just with more sophisticated technology.
The Principal-Agent Problem in AI
There’s a deeper issue here that relates to the classic principal-agent problem in economics. Today’s AI systems operate as agents in the delegation sense: they’re given instructions to follow, but they lack true alignment with organizational objectives.
What we need are AI systems that operate more like principals — entities with contextualized understanding of organizational goals that can make appropriate trade-offs when facing novel situations.
This requires a fundamentally different approach:
- Data-centric foundations that unify information across system boundaries
- Contextual awareness that extends beyond immediate tasks to business objectives
- Coherence guarantees that ensure behavior remains aligned over time
- Governance frameworks that treat AI components as first-class organizational citizens
The New Substrate for Finance Operations
The path forward isn’t about building better agents to navigate the existing landscape of brittle systems. It’s about creating a new substrate upon which more capable agents can operate effectively.
This substrate must:
- Liberate and unify data from fragmented systems
- Enable workflows defined by outcomes rather than procedures
- Support autonomous decision-making while maintaining security and compliance controls
- Increase the “coherence half-life” of AI systems
This approach recognizes that finance operations aren’t just a collection of tasks to be automated. They’re a complex adaptive system that requires intelligence at every layer.
By building this new substrate, we can shift from merely automating procedures to orchestrating outcomes — letting AI handle the complexity that humans currently absorb while maintaining appropriate governance and controls.
That’s the future we’re building at CoPlane — not just smarter agents, but a fundamentally better foundation for them to operate upon. If this diagnosis resonates with your finance operation’s reality, we should talk.