AI in Finance
The AI Hype Cycle in Finance: An FCA's Reality Check
4 April 2026
I have watched AI in finance closely for over three years. Not from the analyst benches, not from a vendor’s product roadmap, but from inside finance functions and alongside finance leaders trying to make real decisions about real implementations with real money. What I have seen is not a revolution. It is not a failure. It is something more complicated and more instructive than either of those narratives.
This is not a balanced piece in the sense of saying something nice and something critical about every claim. Balance for its own sake is a way of avoiding judgment. I have a view, and it is based on evidence.
What Was Overpromised
The claim that AI would close the books in a day. This was circulating seriously in 2022 and 2023. The vision was appealing: AI that reads documents, reconciles accounts, and produces a draft close with minimal human intervention. In a highly standardised, single-entity, data-clean environment with a limited number of integrated systems, something close to this is achievable. That describes a small fraction of the finance functions that were hearing the pitch. For the rest, the conditions required to approach that capability simply do not exist and cannot be created quickly. The promise was real in the same way that a concept car is real: it exists, it works in controlled conditions, and it bears limited resemblance to what you will find in the showroom.
Autonomous finance. The idea that AI agents would operate finance processes end-to-end without human intervention appeared in a remarkable amount of vendor marketing between 2023 and 2025. It is real in narrow, well-defined use cases with constrained inputs and clear rules. Automated bank reconciliation in a single-currency, low-exception environment can approach autonomous operation. But finance functions are not collections of narrow, well-defined use cases with clear rules. They are complex systems that interact with other complex systems. Exception handling is not a margin note: it is a significant fraction of the total work. The claim that AI would make finance autonomous was a simplification that became a distortion in the hands of people who needed to justify a sales target.
The idea that you could deploy AI on top of existing processes without addressing the foundations. This was always wrong. It has been consistently proven wrong across hundreds of deployments over the past three years. It is still being sold. The logic: your processes are too complex to change before deploying AI, so deploy the AI first and rationalise the process later. In practice, AI deployed on top of a broken process accelerates the broken process. Duplicate records get processed faster. Inconsistent coding gets embedded in the model’s training data. Exceptions multiply rather than resolve. The foundation work is not optional preparation for AI. It is the prerequisite. The organisations that tried to shortcut it are carrying the evidence.
See AI will not fix a broken finance function for the full argument on this point.
What Delivered
Invoice processing automation on structured document types from established suppliers. This is where the technology has performed to its claims in well-deployed implementations. A finance function with a stable supplier base, predominantly structured invoice formats, and clean supplier master data can achieve 70 to 85% straight-through processing on AP. That is a meaningful reduction in manual workload. The time savings are real. The error reduction is real. The economics stack up at scale. Organisations that did the data preparation work before deployment and managed the change properly are seeing the results the vendors promised. This is genuine progress.
Anomaly detection in transaction data. The coverage argument is real. Full-population monitoring is genuinely different from sampling, and the difference matters for controls quality. Organisations running AI-based anomaly detection on their AP and journal entry data are catching duplicate payments, identifying unusual transaction patterns, and maintaining a stronger control environment than sampling-based manual review can achieve. The technology is not infallible. The false-positive management requires discipline. But the fundamental capability, applied with good governance, is delivering value that was not available to finance functions three years ago.
Forecasting support tools in narrow, data-rich contexts. The AI is not replacing the forecast. A finance professional who believes the AI forecast is the forecast misunderstands what the tool does. What it does is improve the accuracy of the statistical baseline. For a finance function with 18 months of clean, granular transaction data and a reasonably stable operating environment, AI-assisted forecasting tools are producing baselines that are measurably more accurate than the prior manual approach. The human judgment layer still determines the final forecast. The AI layer makes the starting point better. That is a useful contribution, even if it is not the transformation that was advertised.
The Honest Position on Where We Are Now
The technology is real and useful in the right conditions. That is the honest statement.
Those conditions are more demanding than the vendor literature suggests. Clean, consistent, well-structured data. Sufficient transaction volume to train the models meaningfully. Process standardisation that gives the AI something coherent to work with. Governance frameworks that preserve human accountability at the right points. A finance team with the AI literacy to use the tools well rather than over-relying on them or avoiding them.
These conditions are also less frequently met than the analyst reports imply. Most finance functions have data quality issues. Most have process inconsistencies that were manageable with manual workarounds and become visible problems when you try to automate them. Most have not done the organisational development work needed to build AI literacy into the team.
Finance functions that did the foundation work are getting real value. Finance functions that bought the hype without the preparation are carrying AI shelfware: tools that are technically live and functionally underperforming because the conditions for their success do not exist.
The gap between these two groups is not the technology. The technology is the same. The gap is everything that came before the technology.
What Is Genuinely Coming
Agentic AI will change finance operations significantly within three to five years. This is not more hype. The capability trajectory is observable and accelerating. Agentic AI systems, which can take sequences of actions, make decisions across multiple steps, and operate with minimal human input in bounded contexts, are already functioning in early finance deployments. The capability will mature. The deployment friction will reduce. The economics will shift.
Within five years, the plausible scenario is this: a finance function with the right data foundations and governance structures will be able to run significant portions of AP, AR, reconciliation, and variance analysis through agentic systems that operate largely autonomously within defined parameters and escalate to humans at genuine decision points rather than at every step.
This is a meaningful change in what finance functions look like. It changes the ratio of humans to processed transactions. It changes the skills required of the humans in the function. It changes the oversight model and the governance requirements.
The question is not whether this is coming. It is whether finance functions will be ready when it arrives. Most are not, but the runway is long enough to change that. See agentic AI in finance for more on what these systems are actually doing now and what the near-term trajectory looks like.
The Right Frame
AI is a chapter in a longer finance transformation story. It is not a revolution that makes the previous chapters irrelevant. The organisations getting the most from AI in 2026 are not the ones that adopted it earliest or spent the most on it. They are the ones that took transformation seriously before AI was the agenda item. They standardised their processes. They invested in data quality. They built finance teams that could work analytically rather than just transactionally. They did the foundation work because it was the right thing to do for a professional finance function. AI became a powerful addition to those foundations.
The organisations that skipped the foundation work and bet on AI as a shortcut are now discovering what that costs. The tool is live. The process is still broken. The data is still inconsistent. The team is still resistant. The expected returns are not materialising. The business case is being revised. The implementation partner is working on scope change number four.
This is not a technology story. It is a management story. The technology is a tool. The management decision to build proper foundations, or not to, determines the outcome.
The hype cycle in finance AI has followed the standard Gartner pattern. The peak of inflated expectations was somewhere in 2023 and 2024. We are now moving through the trough of disillusionment: the deployments that did not deliver are becoming visible, the shelfware is being written down, and the organisations that rushed to AI without preparation are reassessing. What comes next, for the organisations that do the work, is the plateau of productivity. This is where the genuine value is. It is not as exciting as the peak. It produces real returns on real investment in real finance functions.
The time to get ready is now. Not to deploy. To get ready. Assess your data foundations honestly. Identify the process standardisation work that needs to happen before automation can work. Build AI literacy into your finance team as a development priority. Design the governance framework before you need it.
The organisations that do this work in 2026 will be positioned to capture the real value of agentic AI as that capability matures. The organisations that wait for the technology to mature before starting the preparation will find the same problem they have always found: the technology arrived before they were ready for it.
Start with the AI readiness assessment to understand where you actually are. Then use the CFO’s guide to AI strategy to map the path from where you are to where you want to be.
The full context for how AI fits within a longer-term finance transformation programme is at AI finance strategy.
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.