The best initial use cases for GenAI are those where three factors coincide: a high volume of repetitive work involving text and documents, a low regulatory cost of any potential errors, and data that is already structured and accessible. In practice, this translates to four areas: internal knowledge search, document processing and summarisation, front-line customer service, and IT team support. Autonomous scoring, credit decisions and risk assessment are best left for later – or subject to strict human oversight from the outset. Below, we explain why this sequence actually pays off.
The question is no longer ‘whether’, but ‘where first’
In the financial sector, the stage of ‘whether to adopt GenAI at all’ is now behind us. According to a 2025 ABBYY survey, 91 per cent of companies in the financial sector have already implemented generative AI, and 98 per cent report positive results. A similar picture emerges from the NTT DATA report (2025, 810 banking leaders): 58 per cent of banks are making full use of GenAI, compared with 45 per cent two years earlier. And according to an EY-Parthenon analysis, the proportion of banks with fully implemented GenAI applications rose from 10% in 2023 to 47% in 2025.
What is more interesting, however, is what lies behind these figures. The same ABBYY survey shows that one in four companies (26 per cent) admit to a lack of proper governance surrounding these implementations. In other words: almost anyone can implement GenAI today, but the competitive edge will go to those who do so in the right order and in a controlled manner. That is why the real question is not ‘whether’, but ‘which process to start with in order to see results quickly and avoid running into regulatory pitfalls’.
The three criteria that determine the order
Before we identify specific processes, it is worth having a simple filter in place. A good candidate for the first implementation meets all three criteria at once.
Volume and repeatability. GenAI comes into its own where people carry out a lot of similar text-based work: reading the same types of documents, writing similar responses, and searching for the same information across disparate sources. One-off, unique tasks are a poor starting point.
The cost of an error and the level of regulation. The more serious the consequences of a mistake, the more cautious one must be. A process in which an error means ‘the adviser received a poorer recommendation’ is a safer starting point than a process in which an error means ‘the client was refused a loan’. The latter category usually falls within the scope of high-risk systems as defined by the AI Act and requires a completely different regulatory regime.
The quality of data and documentation. GenAI is only as good as the data it has access to. If knowledge resides in people’s heads or in undocumented systems, the model has nothing to work with. This criterion is most often overlooked – yet it can be decisive.
The items that pass through all three filters make up your shortlist.
Where to start – four quick wins
Searching for internal knowledge. Procedures, interpretations, product documentation, project findings – all of this is usually scattered across the intranet, hard drives and email inboxes. An assistant that can answer a question and point to the source reduces the time taken to access knowledge from hours to seconds, and the risk is low. A human verifies the answer anyway.
Document processing and summarisation. Applications, contracts, correspondence, claims documentation – the financial sector is drowning in paperwork. Summarisation, key data extraction and preliminary classification are areas where GenAI provides immediate operational relief. It is no coincidence that this is one of the most frequently cited applications in banking.
First-line customer service and support for advisers. The aim is not for the chatbot to replace a human, but to suggest a response, draft a letter or advise on the next step to the adviser. The adviser remains ‘in the loop’, service quality improves, and response times are reduced.
Support for IT teams. GenAI significantly shortens the path from idea to working solution. It speeds up documentation generation, code review, and prototyping. At Finture, we combine this with the methodology ai.pro. This allows us to move from concept to prototype in a matter of days, not weeks.
The common thread running through these four areas is simple: people remain the decision-makers, whilst GenAI acts as an overzealous, patient assistant.
What to deliberately put off until later
The greatest temptations are also the worst places to start. Systems that make autonomous decisions about a customer’s situation carry a high cost of error. These include scoring, credit risk assessment and setting insurance terms. For this reason, the AI Act classifies them as high-risk. This entails obligations regarding supervision, documentation and explainability. This does not mean ‘never’. It means ‘not as the first pilot project, and not without human oversight and a clear governance framework’.
This sequence also makes sound business sense. Quick wins build expertise, trust and data, on the basis of which one can then safely move on to more challenging, more heavily regulated processes.
A factor that is easy to overlook: data in legacy systems
GenAI implementations in finance often prove disappointing despite a brilliant concept. The main reason lies in the data and logic embedded in legacy systems. Key rules are embedded in applications written in Delphi, VB6 or old Java. The lack of up-to-date documentation means that these are a black box for the model. GenAI cannot work out how a commission is actually calculated or why a particular application is rejected.
That is why the first step is often not the model itself, but rather regaining knowledge about processes and systems. Automated, always up-to-date architecture documentation (S*. doc) and a thorough inventory of processes using the Event Storming method ensure that GenAI has data to work with—and that its responses can be verified at all. Without this foundation, even the best model will respond confidently, but not necessarily accurately.
How to get started without running a big programme
There’s no need to start with a three-year AI strategy. A sensible way to begin is with a single, well-chosen process from the list of quick wins, a short proof of value, and clear success criteria agreed in advance (e.g. halving the time taken to process a document). A pilot like this demonstrates real value within a few weeks, educates the organisation and – just as importantly – uncovers gaps in data, governance and documentation before they become costly.
If you want to tailor this sequence to your processes, that’s where we start during the GenAI workshop: together, we select the first process, assess data readiness, and classify regulatory risks so that the initial implementation is both fast and secure.
FAQ
Which process is the best place to start when implementing GenAI in a bank or insurance company?
Can GenAI be used for credit decisions and scoring?
How does GenAI differ from traditional process automation (RPA, BPM)?
Does the AI Act prohibit the use of GenAI in the financial sector?
What is needed for GenAI to work with data from legacy systems?
Are you wondering which process in your organization is best suited for your first GenAI implementation? Let’s talk—we’ll help you prioritize so that your first step is quick, measurable, and compliant with regulations.
