Sam Sammane is the CEO of TheoSym.
getty
Where finance and banking are involved, there is one thing I am sure about: You cannot afford to use ChatGPT or any generic AI “as is.” It is tempting to be dazzled by the ability of these large language models, but when you’re dealing with interest rates, contracts and regulatory compliance as your business, that dazzle is dangerous. The future of AI in banking and finance is hybrid and custom. That is not an opinion. It is a necessity.
People like to imagine AI as a tidy, deterministic engine. It's not. The foundation of these models is probability. They've been shown millions of examples, and they're estimating outputs. Under the hood, it's a black box. We program it, but when it's running, we don't control the exact path it takes.
If you've experimented with "temperature," you've seen this firsthand. I like to think of it as similarity. Lower the temperature, and you get the same output every time because the model is looking for the highest similarity to the training data. Raise it, and you're inviting the AI to accept less similarity. This is where it gets "creative" or, to put it more bluntly, hallucinatory. Creativity can be fun in marketing copy. In banking, hallucinations are catastrophic.
I've witnessed AI give different answers to identical prompts, analyzing the same contract twice, claiming 70% compliance one day and 100% the next. Consider a calculator that gives you two distinct answers for the same multiplication problem. That's what we're dealing with.
That's why the future of finance and banking must be hybrid. We call it neuro-symbolic AI; combining the strength of neural networks with symbolic logic and exact mathematical modeling. Neural networks excel in language and pattern recognition. But when you need formal verification (i.e., rules, math, regulations), they aren't strong. Fortunately, humankind has decades of research in these exact sciences.
Hybrid AI is simply letting each tool do what it does best. Let the AI read and understand documents, but for the math, regulation and verification, hand that over to proven deterministic tools, libraries and engines written in Python, C++ and other precise programming languages.
In pharma, we talk about the need for custom AI. Finance takes it to the next level. Here, custom + hybrid is not a choice. When people get their interest rate wrong, they get annoyed. When their bank account is messed with because we put AI in charge, it is a catastrophe.
Practically, this looks like an agent: a layer of software that calls domain-specific tools automatically. The AI is a wrapper. It reads the contract, decides what needs to be verified and sends those things to external libraries for exact verification.
Think contract compliance. A bank's lawyers may draft a 70-page checking account agreement dense with fine print. With hybrid AI, the language model deconstructs the text, then feeds regulatory checks to formal verification tools. They flag missing disclosures or misleading language; work that once required deep, time-consuming research.
Calculations of interest rates are another case in point. The AI reads and structures the terms, but the calculations themselves are carried out by deterministic libraries that guarantee the right figures every time. The randomness disappears.
When I used to do modeling about two decades ago, these sorts of projects used to be long, tedious processes. Take AI and marry it with these same tools, and you can cut out 95% of the research and development time. Yet, the remaining 5%, how the agent functions, must be determined by human knowledge. AI does not need to guess any of this.
It’s no surprise that a recent MIT study found that 95% of AI pilot projects in large companies fail. Not because AI is useless, but because people ask too much of it. They expect a single chatbot to encompass all human knowledge and overlook the limits.
Instead of humbly saying, “This is a great tool for language and research. Let’s combine it with 50 years of human work,” many try to have the AI do everything. And then they wonder why the pilots collapse.
This new environment also changes how we value expertise. We no longer need armies of coders. We need experts who understand finance, regulation and your application and who can wrap AI around deterministic systems.
Don’t be fooled by the hype. The tools make it easy for someone to claim AI expertise and offer a “solution” on Fiverr that looks like what a serious firm charges $200,000 to build. But that cheap solution is garbage in, garbage out, and it can put your business in serious trouble.
If you are operating a bank or financial institution, you cannot afford to rely on generic AI. You need to customize it to your specific needs and hybridize it with exact tools for math, rules and regulations.
Indeed, these systems will save you time in data processing. However, the cost of lawsuits, regulatory penalties and reputational harm caused by a single mistake can dwarf any savings. AI will save you money, but only if you do it right.
In finance, custom plus hybrid AI is the only responsible way forward. The question is not whether you should build these systems. The question is whether you’re ready to demand the expertise and discipline it takes to do it right. Because when it comes to people’s money, “close enough” is never good enough.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。