Local vs. Global Maximum: Why Internal AI Projects Often Underperform

The plateau that feels like progress

Many of us know this: someone on the team uses ChatGPT to extract invoice data. A colleague builds a small script that generates posting suggestions. Maybe there’s even an RPA bot copying data between systems.

It works. The automation rate climbs from zero to 30%, maybe 40%. Great progress for little effort.

But then something typical happens: progress slows down. Every additional percentage point costs disproportionately more effort. Edge cases, error handling, audit logs, approval workflows, model updates – suddenly you need one or two full-time developers maintaining the whole thing. And every month those developers spend in back-office automation is a month they’re not working on the actual product.

The question is: are we climbing the right peak – or are we standing on a small hill, unable to see the big one?

What is a local maximum?

The term comes from optimization theory. Imagine a blindfolded hiker crossing a mountain range. They can only feel the ground directly beneath their feet and always walk uphill. Eventually they reach a summit – it slopes downward in every direction.

Have they reached the highest point? Possibly not. They might be standing on a 500-meter hill, while the actual peak is at 2,000 meters – separated by a valley they can’t cross without first going back down.

This is exactly what happens with internal AI projects: teams optimize within their current approach and reach the best result that approach can deliver – the local maximum. But the global maximum – the best result that’s possible overall – sits on a completely different hill.

The problem: you can’t see the global maximum from the local one.

The global maximum in finance

What would the global maximum look like for a finance team? The answer is fairly concrete, because others have already reached it:

>80%

Touchless Processing

Invoices fully processed automatically – from capture through coding to approval. No human intervention required.

<3 Days

Monthly Close

The top 18% close their books in three days or less. The median is 6.4 days (APQC).

80%

Cost Reduction

From $10–15 per invoice (manual) to $2–4 (automated).

Why internal AI initiatives get stuck at the local maximum

According to McKinsey, 88% of companies use AI in at least one function. The problem isn’t a lack of interest – it’s the leap from individual solutions to a system. 62% remain stuck in the experiment or pilot stage; only 31% scale enterprise-wide (McKinsey).

Four reasons why this happens:

Limited visibility

When you only know your own documents, edge cases, and errors, you optimize for your own horizon. But document processing follows industry-wide patterns – and if you only see your own slice, you can’t recognize those patterns.

No learning across company boundaries

What shows up as an edge case at one company today has long been solved at another. Internal teams can’t benefit from this knowledge transfer – they only learn from their own mistakes. Platforms serving many customers benefit from every single resolved case.

Missing reference points

If you’ve never seen 80% touchless processing, you don’t aim for it. 50% feels like the ceiling – until you learn that others have long surpassed it. Without that comparison, the ambition level stays low.

Underestimated engineering effort

An AI prototype runs in a week. But the path to a production system – audit logging, regression tests for model updates, monitoring, compliance documentation – takes 12–24 months and ties up one or two developers permanently. That capacity is then missing from the actual product.

The honest counter-arguments

Let’s be fair. There are two objections that are legitimate:

Data sovereignty

Financial data is among the most sensitive data categories. GDPR, industry-specific regulations – these are real requirements, not a smokescreen. But: anyone automating internally also needs LLMs – and sends the same data to OpenAI, Anthropic, or Google. The data protection question arises in both cases. A specialized provider can often offer more here: dedicated instances, EU hosting, DPA contracts, and encryption as a core product rather than an IT side project.

Vendor dependency

A specialized vendor can be acquired, raise prices, or discontinue their product. That’s a structural risk. The counter-question: what’s the risk of not automating? And: open data formats and export capabilities significantly reduce lock-in.

A third argument we hear often: “We’re too complex for a standard solution.” In practice, this is almost always process debt – historically grown complexity that was never cleaned up. The truly unique parts of a finance department (strategic FP&A, valuation decisions, business judgments) are exactly the parts that shouldn’t be automated anyway.

Leaving the hill – without crossing a valley

The mountain metaphor suggests that switching to the global maximum means a setback first. But that’s exactly the advantage of an external partner: they’ve already crossed the valley. The infrastructure is in place, the typical edge cases are solved, the system is running in production at other customers.

In practice, this means: no months-long valley of in-house development, but a faster path to the higher peak. An internal team would have to make every mistake themselves, build every integration themselves, document every regulatory requirement themselves. A specialized partner has already been through that – with other customers, in other industries, with other systems.

Shortcut, not detour

Most companies that have reached the global maximum didn’t get there alone. They bought the experience – and could then focus on what only their internal finance team can do best: strategic analysis, business judgments, planning.

Conclusion

The question isn’t: “Can we use AI internally?” – probably yes. ChatGPT, Claude, Copilot, simple scripts – that works for individual tasks.

The question is: are we aiming at the right peak?

The global maximum in finance isn’t “AI helps with individual steps.” It’s: fully automated document processing including posting and approval, a monthly close in under three days, and a finance team that works strategically instead of moving data around.

The first step isn’t technical – it’s conceptual: accepting that there might be a higher peak beyond the current hill. The benchmarks show that others have already arrived there. That shifts the question from “How do we optimize our approach?” to “Is it the right approach?”