The gap between AI hype and reality in finance
Artificial intelligence and automation have become buzzwords in the world of finance, but the gap between the talk and the walk is still massive. More than three years after ChatGPT reignited global interest in AI, many companies are still figuring out, in practice, where this technology actually fits into daily financial operations.
A recent Gartner survey revealed something pretty eye-opening: nearly 60% of finance teams are already testing or implementing AI projects, but only 7% of CFOs report a real, meaningful impact from those initiatives. In other words, the difference between a pilot project and a concrete result that actually moves the needle is still a huge gap for most organizations.
As Mohit Sharma, co-founder of the startup Pinaka AI and a professional with decades of experience in finance, put it bluntly: AI is going through an identity crisis right now. But that does not mean nothing is working. Quite the opposite — there are real cases, with real people, showing where the technology delivers genuine value. 🎯
Three finance professionals from different parts of the world shared how they are using AI and automation in their own contexts, what worked, what did not, and what lessons they picked up along the way.
Predicting who will pay late before it actually happens
Mohit Sharma co-founded Pinaka AI in 2023 to solve a pain point he heard over and over throughout his global career in finance. The problem was easy to understand and extremely hard to solve: roughly 60% of invoices issued in B2B environments are not paid on time.
When a payment is late, the consequences snowball fast. The sales team may need to renegotiate contracts. The company might need to take on debt to cover the cash flow gap. It is a vicious cycle that eats up time, energy, and money across multiple departments simultaneously.
Pinaka AI’s solution, now with about a dozen employees, is software that predicts which customers will pay late and identifies the specific reason why it is likely to happen. The platform is currently being piloted by two major manufacturers in India.
According to Sharma, the company’s algorithm can make predictions with 96% accuracy. Hosted on Oracle Cloud, the platform also uses generative AI to recommend actions to the user and can draft and send personalized emails to help collect payments before the due date.
The system works by integrating data scattered across different sources — like CRM systems, ERP platforms, credit agencies, and even news outlets — to create what Sharma describes as a single version of the truth about each customer’s payment behavior. It uses four types of AI working together: a recommendation engine, decision intelligence, a classification algorithm, and generative AI.
The startup is still in its early stages, having recently completed an accelerator program, but Sharma points out that the new generation of AI tools has made it much faster and more affordable to build this kind of product. In the past, traditional automation could have addressed part of the problem, but not at the same speed or with the ability to handle the complexity of the data involved.
When we are solving complexity, we need intelligent systems, Sharma summed up.
Unifying messy charts of accounts with ChatGPT
Janice Stucke, a CPA and CGMA candidate, took on the CFO role at CREW Network last year — an association connecting more than 14,000 women in commercial real estate around the world.
When she arrived, Stucke found a finance department that was still issuing paper checks and in need of a deep overhaul. It was right in her wheelhouse — she had already implemented robotic process automation (RPA) and generative AI in previous jobs.
But before she could automate anything, there was a massive obstacle in the way. The organization’s accounting data was spread across about 50 different entities, each with its own chart of accounts. There were subsidiaries, affiliated entities, operations in multiple countries, and completely different formats. This fragmented landscape generated more than 10,000 general ledger transaction lines per month and made payment processing painfully slow.
Keeping everything running required constantly updated custom coding — a huge technical debt for an organization with just 35 employees.
The time dividend AI delivered
Stucke decided to create a consolidated chart of accounts for all the entities, but converting the entire historical structure from the old format to the new one was a massive undertaking. The traditional way, her team would have spent weeks, even with help from consultants, using spreadsheet macros or RPA — and both approaches would have required customization for the countless variations in data format and definitions.
Stucke’s approach was different: she used her corporate ChatGPT account to handle the transformation. She fed in the charts of accounts and asked the AI to map the data to the new unified format. The tool could even handle the variability in how each entity labeled its transactions — using generative AI to interpret the many different ways the charts might categorize something like event revenue, for example.
It worked — up to a point. After running the transformation flawlessly on ten consecutive charts, the AI might introduce a new, unwanted method on the next one. Stucke also ran into common generative AI frustrations, like hallucinations and errors. She tried getting ChatGPT to include formulas that would show and verify its own work, but she simply could not get the tool to do it correctly.
The fix? She implemented her own verification formulas, the same way she would with human work or RPA, to check the unified ledger against the legacy data.
My internal controls process did not change, Stucke explained.
Even with the frustrations, the AI-powered approach delivered faster and more reliable results in days — allowing her to move forward with automating the finance function. And all of this happened with a relatively generic, accessible, and affordable product.
Stucke believes this new wave of AI products is empowering small and midsize businesses to experiment. Many of these organizations feel they do not have the talent needed to implement complex systems, but now it is possible to grab an off-the-shelf product and automate processes with far less effort than before — as long as there is a clear vision behind the initiative. 💡
Fixing manual processes at a continental scale
The most valuable advances in AI and automation do not happen without a unified and accessible data source. That has been a central priority in the work of Lawrence Amadi, partner and technology risk business leader at KPMG Africa.
Among his clients are some of the continent’s largest telecom operators — one of them with more than 85 million subscribers. Amadi has been working with these companies to transform their SIM card management systems, which authenticate and manage user identities and their devices on the network.
The company relied on manual processes to manage this colossal volume of data. Roughly once a week, employees would download the data and check for incomplete or anomalous records. This manual process increased the risk of incomplete data exports and audit fatigue among staff.
It gets really stressful and demoralizing, Amadi described.
KPMG Africa’s goal was to automate the export, analysis, and exception generation from the same data — a project that took seven months and involved professionals from multiple areas of the organization. Amadi emphasized that the project required people with deep knowledge of the products, the business rules, and how the data is generated, along with data analysts capable of unpacking, analyzing, and reorganizing large volumes of information.
The solution was built with Automation Anywhere, which describes itself as an agentic process automation system combining the power of AI, automation, and RPA.
The result? Instead of periodic manual downloads, there is now an automated export running in the system. There is automated analysis running continuously. And there are exceptions being flagged automatically. All of this led to fewer errors, greater efficiency, and more consistent reporting for the board and the audit and risk committee.
With the data organized and centralized, the company is now ready to apply automation and AI to even more areas of the business.
Three practical lessons from people already implementing AI in finance
The experiences of these three professionals reveal patterns that apply to any finance team thinking about jumping into this journey. Here are the most relevant lessons that emerged from their projects.
Understand the real costs before you start
Sharma learned firsthand that running AI costs a lot more than most people think. A poorly configured product can burn through tokens — a practical measure of computational power and cost — at an alarming rate. Beyond the direct costs, there are indirect risks: what happens if a generative AI makes a serious mistake or causes a controversy, like using inappropriate language with a customer?
There are layers of cost, direct and indirect. The moment you lose visibility, you lose money, Sharma warned.
These costs and risks need to be weighed against a realistic forecast of revenues and efficiencies to be gained. The fundamental question, according to Sharma, is not which technology to use, whether OpenAI or Google, but rather: what is the return for my business? When does the solution hit breakeven? 📊
Build buy-in before you build the system
In large organizations, an AI or automation project can require combining skills and data from multiple areas of the company. According to Amadi, none of that happens if people do not understand and accept the mission.
You need buy-in from all of them. The why is crucial, Amadi stated.
That why can vary — better control coverage, improved business visibility, reduced rework. It does not matter what it is. If the purpose and methods of the project are not clear to everyone, there simply will not be any progress.
Do not do it all alone — let others learn
Stucke managed to use generative AI to complete a massive data transformation in just four to five days, by herself. What would have taken an entire team of consultants and employees two to three weeks was handled by one person with the right tool.
But that efficiency came with a price: my entire team was not taken on that journey, Stucke acknowledged. In other words, her team did not get the chance to practice using AI or to understand how and why the tool was so effective.
AI proficiency will continue to be an essential skill for finance teams, especially as automation absorbs more operational work. Getting everyone involved from the start can help prepare people’s careers for the future and accelerate the transformation of the finance function as a whole.
What these experiences tell us about the future
When we put all these stories together, some patterns become pretty clear. The first is that success with artificial intelligence in finance rarely happens in a straight line. There is almost always a frustration phase before results show up, and organizations that quit during that phase give up right when they were closest to reaping the rewards. Having realistic expectations from the start is what separates the projects that move forward from the ones stuck in pilot mode forever.
The second pattern is about scope. Smaller projects, focused on specific and well-defined problems, have a significantly higher success rate than broad initiatives that try to transform everything at once. Real transformation happens incrementally — each automated process, each forecasting model implemented, each report that no longer needs to be done manually builds a solid foundation for more complex initiatives down the road.
And the third pattern, perhaps the most important one, is about data. Without quality data, no AI strategy is going to work. Before choosing any tool or platform, it is well worth doing an honest assessment of the organization’s data maturity. That means understanding where data lives, how it is structured, how up to date it is, and whether there are significant gaps. This upfront work might seem a lot less glamorous than deploying a machine learning model, but it is what will determine whether the project actually gets off the ground with real results. 🔍
AI is not a solution you flip on and wait for a miracle. It is a journey that requires preparation, patience, and a clear vision of where you want to go.
The financial sector is at a real turning point. The tools are more accessible, the cost of experimentation has dropped, and success stories are multiplying. But the distance between the hype and actual delivery still demands a lot of work — and a lot of honesty about where each organization truly stands on this automation and innovation journey.
