AI in Finance
AI Won't Fix a Broken Finance Function
8 December 2025
I have sat in enough boardrooms to know what happens when a new technology gets positioned as the answer to an operational problem that has not been properly diagnosed. It happened with ERP. It happened with shared services. It is happening right now with AI.
The pitch is familiar: your finance function is slow, manual, error-prone. AI will automate the repetitive work, flag the anomalies, close your books faster, and free your team for higher-value activity. All of that is possible. None of it is guaranteed. In a function with structural problems, AI will make things worse, not better.
This is not a pessimistic view of the technology. AI represents a real step change in what finance functions can do. But the step change accrues to functions that are ready for it. The rest will spend significant budget finding out that automation is not transformation.
AI is an amplifier, not a fixer
AI amplifies what already works. It also amplifies what does not.
Give an AI tool a well-designed reconciliation process with clean, consistent data and clear exception handling logic. It will automate a large proportion of the work, surface the right exceptions, and close faster with fewer errors. That is what vendors demo.
Give the same tool a reconciliation process that has accumulated workarounds over ten years: data coming from three systems in inconsistent formats, a team that manually adjusts figures before running the report. The AI will automate those workarounds. It will do it consistently and at scale. The result is a faster, more expensive version of the same problem.
This is not a hypothetical. It is the pattern I see in the gap between AI pilot results and AI production results. Pilots are run on curated data and clean processes. Production is real life.
The three failure modes AI gets used to patch
Every finance function I have worked with has at least one of three structural problems. Frequently, all three.
Bad process design. The process was designed for a different volume, a different system, or a different team. It has been patched and extended but never redesigned. Manual steps exist because someone worked around a system limitation in 2016 and it became standard practice. Approval steps add no value because nobody remembers why they were added. Reconciliations reconcile to a number that does not mean what people think it means.
Bad data. The same entity has three names across three systems. Dates are stored in four formats. The chart of accounts has been extended by multiple people without governance, so the same type of cost sits in five different cost codes depending on who raised the purchase order. Reports are manually adjusted before circulation because the source data cannot be trusted without human interpretation.
Bad adoption. The system was implemented, the team was trained, and then the team went back to doing what they knew. A sophisticated ERP sits underneath a set of spreadsheets that replicate the reports the ERP should be producing. The data in the ERP is a month behind because that is when someone re-enters it from the spreadsheets.
AI does not fix any of these. It assumes them away. Most vendors do not do this maliciously. They build tools for the idealised version of your process, not the actual version. The gap between those two things is yours to resolve.
See my earlier post on why finance functions fail at scale for a longer treatment of how these problems compound. The three-tier automation framework is also relevant here: the tier structure only works if the foundations are sound.
What fixing the foundation means in practice
I am not saying you need to spend three years on a transformation programme before you touch AI. That is the other failure mode: using “we need to fix the foundations first” as a reason to do nothing.
Fixing the foundation means being clear about which processes you are asking AI to work on, and whether those processes are coherent enough to automate.
Start with a process map. Not the aspirational one. The actual one. Walk the process as it is executed today: what triggers it, what data it uses, what decisions are made, what exceptions arise, what happens when something goes wrong. This takes time. It is the work most people want to skip. It is also the work that determines whether your AI investment returns anything.
Identify the data dependencies. What data does this process consume? Where does it come from? How consistent is it? If the answer is “it depends on who processed it”, you have a data quality problem that will defeat an AI tool before it gets started. I cover the data dimension in more depth in the upcoming post on data quality and AI in finance.
Document the exception logic. AI tools for finance automate the standard case. The non-standard case still needs human judgment. If you cannot clearly define what “standard” looks like, neither can the AI. Documenting exception logic is not just a prerequisite for automation. It forces a conversation about whether your current exception handling is consistent and defensible.
None of this is glamorous work. It is exactly the kind of work that gets deprioritised when a finance team is busy closing the books and a vendor is demonstrating a slick AI interface. Do it anyway.
Why this matters if you have done the transformation work
Finance leaders who have done the work of process design, data quality, and system adoption are in a strong position. Their functions are AI-ready. The tooling now available to them is impressive and improving quickly. The gap between what AI can do today and what was possible three years ago is significant.
Clean reconciliation process, consistent data, a team that uses your systems properly: AI-assisted reconciliation can reach 80 to 90% automation rates on high-volume matching. Well-designed intercompany process with consistent data across entities: AI-assisted consolidation can compress your close timeline materially. Consistent expense coding and a governed chart of accounts: AI-assisted expense processing can remove a large proportion of manual intervention.
These are real outcomes. They are achievable. They are not achievable by skipping the transformation work and buying an AI tool instead.
The vendor conversation you need to have
When an AI vendor pitches you, the most important question you can ask is: what does my data need to look like for this to work?
A vague answer is a signal. Good vendors know exactly what their tool requires because they have seen it fail when those requirements are not met. A vendor who says “it works with any data” is either selling you something that does very little, or setting you up for an expensive disappointment.
Ask to see a customer case study where the implementation was difficult. Not the reference customers on the website. The customers who took longer than expected, who had data problems, who had to go back and fix process issues before the tool worked. That is the conversation that tells you what you are actually buying.
AI is the next chapter, not a shortcut
The finance functions that will get the most from AI over the next five years are not the ones that invest the most in AI tooling. They are the ones that invest in AI readiness: clean data, coherent processes, capable and engaged teams.
AI is the next chapter of finance transformation. It is not a shortcut past the earlier chapters. Organisations that skipped those chapters will find that out in their implementation projects. Organisations that did the work will find that the tooling now available to them is more powerful than anything they have had access to before.
I have watched finance transformation from the inside for twenty years. The technology changes. The principles do not. Get the foundation right. Then build on it.
Explore the full AI in Finance Strategy framework, or read the next post on what a coherent CFO AI strategy actually looks like.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader. She writes about AI in finance transformation from the inside out.