The use of AI in finance is slowly graduating from theory to testing. Across the FP&A landscape, experimentation is abundant with companies testing machine learning pilots for forecasting, natural language generation for reporting, anomaly detection in cost centres, and much more. And yet, most of these initiatives remain firmly stuck in the exploratory phase. Despite proofs of concept, internal workshops, and technical demos, very few projects have actually reached live deployment or scaled impact.

This phenomenon, which we’ll call pilot purgatory, is a defining challenge for finance leaders in 2025. The technology is real. The intent is genuine. But progress stalls at the transition point between potential and production.

At its heart, this is not just a technological problem. It’s also an operational and organisational one. Moving from exploration to execution requires rigour, alignment, and structural readiness. Without them, AI remains a shiny object rather than a lever of value.

The invisible barriers slowing AI in finance

Many FP&A teams underestimate the operational groundwork required to successfully implement AI. While it’s tempting to start with ambitious plans like “let’s build a forecast optimiser” or “let’s automate variance explanations”, the real constraints tend to sit beneath the surface.

Data governance is the most immediate bottleneck. AI models, especially those trained on internal data, require structured, clean, and consistently tagged inputs. Unfortunately, financial data is rarely housed in a single source of truth. It’s often distributed across legacy systems, spreadsheets, and disconnected planning tools. Without a robust data foundation that covers access control, quality validation, and integration, AI projects become fragile and unreliable.

Closely tied to governance is the issue of security. AI models raise new questions about data privacy, regulatory compliance, and auditability. For FP&A functions operating in tightly regulated industries, or across borders with varying data laws, the risk tolerance for deploying AI models on sensitive financial data is low. That often pushes innovation into sandboxed environments, further isolating pilots from day-to-day workflows.

These structural issues compound over time. Pilots remain technically viable but politically sensitive, or operationally awkward to scale. Confidence erodes. Momentum fades.

The ambiguity of success

Another less obvious, but equally damaging factor is the lack of clear success criteria. Many AI pilots are launched under the banner of innovation, but without concrete metrics of what constitutes a win.

In traditional finance transformation projects, success might be defined by reduced cycle time or improved forecast accuracy. However, AI pilots often aim at softer outcomes such as greater insight, more automation, or improved decision support. These are laudable goals, but difficult to quantify, especially in the early stages.

Without agreed definitions of success that are ideally aligned with business value, pilots fail to build trust. Executive sponsors lose interest. IT teams deprioritise support. And finance teams, already time-poor, revert to tried-and-tested methods.

A useful principle here is to design AI pilots backwards: start with the end-state business decision you want to improve and work back to the technical tool or model required. Define not only what the AI should do, but also what action it should take and how that action will be evaluated.

Misaligned investment and overhyped returns

A further complication arises from the mismatch between perceived potential and actual investment. AI is often pitched as a strategic leap forward – capable of driving step-change improvements in performance and productivity. But the resources committed to these pilots rarely reflect that ambition.

Too many projects are launched with minimal cross-functional support, no change management budget, and limited IT bandwidth. The result is predictable: underpowered pilots that demonstrate technical promise but lack the operational scaffolding to scale.

This creates a dangerous feedback loop. Modest pilots yield modest outcomes, which are then interpreted as a sign that AI “isn’t ready” for finance. But in truth, the pilot wasn’t designed, or indeed funded, to test readiness at scale.

Breaking this cycle requires reframing AI not as a standalone initiative, but as a capability build. The investment should reflect not just the technology itself, but the organisational muscle required to use it effectively: process redesign, skill development, data infrastructure, and governance mechanisms.

A pragmatic roadmap for scale

So what does it look like to move AI in finance from exploration to execution? While there is no universal blueprint, certain principles can increase the likelihood of successful scale:

• First, embed pilots within live business processes, not artificial environments. A model tested only in isolation will always face new challenges in production. Aligning pilots with real workflows ensures better feedback and faster iteration—and makes success harder to ignore.

  
  

• Second, involve end users early. Too many AI projects are designed top-down, by data teams or external consultants. But true adoption depends on trust from the people expected to use the outputs. Finance professionals should be co-creators of AI tools, not just recipients of them.

  
  

• Third, treat AI deployment as a programme, not a project. This means appointing sponsors, creating training pathways, and aligning incentives with adoption. Without formal ownership, even the best pilots become orphaned.

  
  

• Finally, build a portfolio of use cases. AI will not transform FP&A in one dramatic leap but through a series of cumulative, compounding gains. A few modestly successful pilots, scaled and integrated, will deliver far more value than a single overambitious experiment that never makes it out of testing.

Learning from those further ahead

Some organisations are beginning to break out of pilot purgatory. These teams share a common characteristic: they do not treat AI as a magic solution. Instead, they approach it as a practical tool – useful when pointed at the right problems, with the right support systems.

They also recognise that the most important enabler of AI in finance is not the model, but the mindset. A willingness to test, learn, and evolve. A commitment to transparency and auditability. And above all, a focus on solving business problems, not just experimenting with technology.

As more finance teams build on this foundation, we’ll see AI move from the margins to the mainstream. But that transition will not be automatic. It will require deliberate choices, consistent investment, and a clear-eyed view of both the promise and the limitations of the technology.

Closing thoughts

Pilot purgatory is not an inevitable fate. It’s a signpost – a signal that an organisation is willing to explore but not yet ready to commit. Moving beyond it means embracing AI not just as an experiment, but as a serious part of financial transformation. 

On the 25th of June, we’ll be co-hosting the Executive Finance Summit where like-minded leaders from across the industry will be coming together to discuss how to navigate these new developments. If you’re looking to go beyond the pilot phase and explore what meaningful AI deployment could look like for your finance function, then you might be a perfect fit for the event.

You can apply to attend here.  We hope we see you there!