“In the financial world, the accuracy and reliability of its responses to prompts depend on more than just processing power. This is where advanced technologies like machine learning and natural language processing (NLP) play a key role, helping to contextualize complex and heterogeneous data”.
Juan José Gómez
CMO of Coinscrap Finance
Generative AI is not infallible, and in finance, this can be a risk
As a marketing professional and technology enthusiast, I cannot deny that the arrival of generative AI has been a game-changer in the way I work. It has become an incredible enhancer for creating written content, translating, analyzing data, generating Excel formulas, or writing commands in various programming languages.
But –and here’s the heart of the matter– it doesn’t get things right every time. And even worse, you’re not always aware that it’s happening.
The reason is actually quite complex, and there is a technical barrier beyond my knowledge, as a non-technical person. But I’ll try to explain the reasons and what we can do to fix it, from a business perspective.
Context in Generative AI: The key to accurate responses
Let me set the stage by interacting a bit with ChatGPT. I wrote the following prompt: “I earn 30,000 euros a year, what house can I afford?”
ChatGPT responds: “With an annual income of 30,000€ (equivalent to 2,500€ monthly), here’s an estimate of what kind of house you can afford“
Mortgage monthly payment calculation with ChatGPT
“It is recommended that the monthly mortgage payment does not exceed 30% of your income. So, the calculation would be as follows:”
“You could afford about 750 euros per month on your mortgage.”
You’d agree that this is –more or less– a good answer… Based on the information and context it has at that moment, of course. Because, if I were already paying a similar amount on a mortgage for another property, canceling a loan I took out for a master’s degree, or simply had unstable income, taking on this new monthly expense could seriously endanger my financial health.
If you’re interested in understanding how AI makes decisions, check out our CTO Óscar Barba’s article on AI explainability.
Why doesn’t ChatGPT give me a better answer?
ChatGPT and other generative AI tools are very useful, but without enriched context, they generate output based on the information provided and the training they’ve received. In essence, the response may be inaccurate because it lacks the necessary data to fully understand and interpret my complete financial situation.
Now, imagine it knew exactly where my income came from and how frequent it was. Or even where my expenses were going and which of them fell into categories like “Leisure” and “Travel.” The same would apply if I shared accounts with people who also had income.
All that information would actually be a curated context. If ChatGPT had access to that, its response would have been of higher quality and relevance. It could personalize the answer based on my current situation or even make a projection for the future.
So, how do we provide generative AI with the context it needs to deliver hyper-personalized responses? Let me tell you about it, because this is precisely what we do at Coinscrap Finance.
The role of transactional data enrichment in creating curated context
Since our beginnings in 2016 as a B2C fintech focused on micro-savings, we’ve been working with customers’ payment data.
We quickly realized that if we were able to structure, categorize, and enrich transactional data, we could understand users’ financial habits. And in an increasingly digital environment, where cash usage continues to decline –at least in this part of the world– this meant understanding almost everything about them.
Over the years, we developed the technology and created our COCO algorithm, which uses machine learning and natural language processing to turn raw banking data into actionable, useful information for decision-making.
COCO adds attributes like category, subcategory, type, specialty, merchant, geolocation, recurrence and more, to the millions of transactional data points generated by a bank every day.
In this way, we help banks build the necessary context so that their AI-powered virtual assistant can provide flawless, personalized answers to questions like: How much can I afford for a mortgage? Can I go on vacation again this summer? or
“What car can I afford?”
Generative AI is an incredible tool, but its effectiveness in the financial industry depends on creating and maintaining curated context that makes sense of raw data. In an area where every decision has a real impact on our financial health, the ability to transform unstructured data into precise and useful information is not just an advantage; it’s a necessity.
About the Autor
Juan José Gómez is the CMO of Coinscrap Finance and leads the Marketing team in defining strategies and implementing communication and branding actions. Through close collaboration with Sales, he continuously works to improve lead generation and potential customer processes.
He holds a degree in Advertising and Public Relations and several postgraduate degrees in Digital Marketing, Inbound, Social Media, and Analytics. His career spans the creation of projects for multinational companies, SMEs, and startups in various sectors. Thanks to his experience in both B2B and B2C, he is able to approach the growth of our scaleup from multiple angles.
