Why AI Implementations Fail in Finance, and the Fix

What the Research Says About Why AI Implementations Fail

The most cited evidence comes from MIT's NANDA initiative, whose 2025 report The GenAI Divide reviewed more than 300 public AI initiatives, interviewed representatives from 52 organisations and surveyed 153 senior leaders. Its headline finding was stark: around 95% of organisations were getting no measurable return from their generative AI pilots.

The more useful findings sit underneath the headline. The researchers concluded that the gap was not explained by model quality or regulation. It was explained by approach. Organisations that bought in specialist help and partnered with external providers succeeded at roughly twice the rate of those attempting purely internal builds. And while budgets were skewed towards visible, front-office uses such as sales and marketing, some of the strongest returns came from back-office automation, the category finance sits squarely in.

For a finance leader, that combination is encouraging rather than discouraging. The failure rate is high, but it is driven by choices you control: what kind of AI you buy, which work you point it at, and whether it is fitted to how your processes actually run. The rest of this guide is about making those choices well.

It is also worth being honest about what the statistic does not say. It does not mean AI does not work in finance. Teams using AI well are saving meaningful time on commentary, reconciliations and reporting. It means that buying AI and getting value from AI are two different projects, and most organisations only budget for the first.

Dead End One: The Horizontal Assistant Nobody Owns

The first common shape is the horizontal assistant. Everyone gets Microsoft Copilot, ChatGPT Enterprise or Claude, and the expectation is that people will work out how to use it. Some will. Most will use it for drafting emails and summarising documents, which is useful but rarely changes how the finance function performs.

Three months in, the typical result is a growing usage bill and a scattering of half-built automations and agents that nobody owns. One analyst built something clever for the bank reconciliation, then moved team. Another built a variance tool that only works on last year's chart of accounts. None of it is documented, governed or maintained.

There is a quieter cost too. Every correction your team makes in a general-purpose tool (no, we treat that differently; no, the approval limit is lower for that entity) is institutional knowledge being typed into a system. Depending on your licence and settings, that knowledge may improve a product you do not own, while doing nothing to build a reusable asset for your own organisation. It is worth knowing exactly what your plan opts you into, which we cover in our guide to whether ChatGPT is safe for financial data.

Horizontal assistants are not a mistake. They are a sensible baseline. The mistake is treating the licence as the implementation.

Dead End Two: A Stack of Point Solutions

The second shape is the opposite: a specialist AI tool for each process. One for accounts payable, another for the close, another for expenses, another for cash forecasting. Each comes with a good demo and a clear business case.

The trouble is that none of them knows your process as it actually runs. Your AP process may have seven steps rather than the four the vendor modelled. Your exceptions may arrive by email, in three different formats, from a shared inbox that also receives supplier statements. The tool handles the clean path well and hands everything else back to your team, who now log into a dozen systems to get part of the gain.

Point solutions also multiply the integration and governance burden. Each one needs access to your data, its own controls, its own owner and its own renewal decision. For a mid-sized finance team, the overhead of managing six AI tools can consume much of the time they were meant to save.

There are good point solutions, and for some high-volume processes a specialist tool is the right answer. But a stack of them is not a strategy. Before buying another tool, it is worth asking a harder question: what is the work that actually consumes your team's time, and does any of these products touch it?

In our experience the answer is usually uncomfortable. The time is not in the core transaction the tool automates. It is in everything around it.

The Real Problem Is Glue Work

When we map finance processes with clients, the hours rarely sit in the steps written in the procedure manual. They sit in the glue: the person copying a number from one screen into another, chasing an email when two figures do not match, reformatting a supplier statement so it can be compared, escalating when nobody replies, and keeping a private spreadsheet that tracks what the systems cannot.

That glue work is exactly what most AI purchases miss. The horizontal assistant could help with it, but nobody has designed it to. The point solution cannot see it, because it lives between systems rather than inside one.

The shape that works is neither. It is a single layer that sits across the systems you already run, does the routine work the way your team would, and flags only the decisions that genuinely need a person. Think of it as automating the handoffs, not replacing the systems. Automate the glue and the close shrinks. Buy another tool and it usually does not.

This is also why external help tends to outperform internal builds. Designing that layer well requires someone who understands both the technology and the finance process, and who has seen enough implementations to know where the hidden steps usually are. Most finance teams have plenty of the second and too little time for the first.

How to Audit Your Worst Process in 20 Minutes

You can find the glue in your own function without a consultant or a project. Pick the one finance process everyone dreads, and write down how it actually runs, not what the procedure says. Include the workarounds, the monthly volume, what a single error costs and how often one slips through, and how exceptions are handled and in how many formats. The gap between the procedure and reality is where the money hides.

Then let AI do the analysis. Paste your notes into this prompt:

The final instruction matters. Without it, models tend to fill gaps with a generic version of the process, which defeats the point of the exercise. Remove any confidential detail before pasting, or use an enterprise account with data protections you have checked.

From Audit to Implementation

Run the audit on two or three processes and you will usually find that one stands out: high volume, mostly pattern-matching, expensive when it goes wrong. That is your first candidate. Resist the urge to automate everything at once. One process done properly, with clear ownership and controls, builds the confidence and the evidence for the next.

The audit tells you what is worth automating. Turning that into a working agent that runs inside your existing tools, keeps your data within your own environment and passes an auditor's review is the harder part, and the part most teams get wrong. The common failure points are predictable: no named owner, no exception route, no test against historical data before going live, and no plan for when a supplier or system changes format.

Two questions are worth answering before any build starts. Who owns this when it breaks? And how will we know it is still right in six months? If you cannot answer both, the implementation is not ready, regardless of how good the demo looked.

If you would like a structured version of this exercise across your whole function, our AI audit and assessment maps your workflows, finds where the hours actually go and produces a prioritised plan. You can read more about the approach in what an AI audit involves.

Frequently Asked Questions