AI—whether it turns out to be the best thing since sliced bread or the beginning of the end for humanity—is here to stay. The most practical thing we can do is embrace it. However, like any new tech, there are some risks with adopting it.
When it comes to using AI within a finance team, security risk is one of the first things that come to mind. For good reason; finance professionals aren’t too excited to hand over their cash flow data to the internet. How many times have we heard about another massive data leak?
Then, there’s the risk of the output being inaccurate. In finance, there’s just no room for being wrong.
In this article, we’ll address some of the biggest questions we’ve received from treasury and finance departments about Trovata AI, the first generative AI tool (think ChatGPT) for finance. We’ll uncover the inner workings of Trovata AI, showing how it was architected for scalability, flexibility, security, and accuracy from the outset.
- At Trovata, we understand the importance of keeping all of your financial data private and help you through every step of the InfoSec process during onboarding.
- As a SOC II, Type II Certified organization, we are committed to remaining compliant.
- Trovata keeps your data private and does not share it with any AI company, like OpenAI, to reap the benefits of its technology. Through the use of tokens, Trovata has developed a secure translation layer to ensure that your data stays within your Trovata bank data lake.
- Instead of letting GPT4 access your data, we have taught the large language model (LLM) how Trovata is structured so it can query it and then present you with the information you request. As such, it does not do any calculations on its own; they’re done by Trovata, which means they remain accurate.
- Trovata AI is not limited to working with only one large language model (LLM). Currently, it runs on GPT4 but may change as new models become available.
Are Fears Around AI Unfounded?
First of all, given the risks mentioned above, are the fears around AI overblown? Of course not!
Concerns about sharing your sensitive information are completely warranted. That information has to be protected. The world is coming off a wave of tech and digitization, and certain fraud and cybersecurity issues have been unraveling, in the most public of ways. Think of the scam where victims are duped into wiring money to fake escrow companies, the Equifax data breach, or the Cambridge Analytica data scandal.
Security was one of the biggest issues our leadership and engineering team discussed before setting out to build a generative AI tool for finance.
Our Founder and CEO, Brett Turner, and CPO/CTO, Joseph Drambarean, recognized that to commission the project, we’d have to do it in the right way. How could we use a large language model (LLM) like OpenAI’s ChatGPT while keeping privacy and accuracy at the forefront?
Here’s how we did it:
The Journey to Develop a Safe AI That Finance Teams Could Actually Use
Check out this live stream session where Joseph answers questions about Trovata AI – live.
First of all, we wanted to maintain the novel and delightful user experience of ChatGPT.
But, we realized we had a clear division: The AI would need access to clients’ financial data, but the data cannot leave their secure instance of Trovata, our cash management platform.
“The data that our customers have should never leave their instance. That was a rule that we set from day one. But we wanted ChatGPT to be beneficial in a wide range of contexts, and the only way to do that is by being able to communicate with it in a way that is not providing private access to data, but is giving enough context to say, ‘Hey, you have knowledge of how Trovata works as a system and within the constraints of that knowledge, give me what I need to look up the information to solve this user’s query.'”– Joseph Drambarea, Trovata CTO and CPO
We knew we wanted to take advantage of ChatGPT’s natural abilities. ChatGPT is exceptional at taking any sentence that you throw at it and then condensing it and inferring what you mean. It figures out what the user seeks to understand and what’s their motivation for asking. Then, it comes up with a solution.
Knowing this, the first step we took to using ChatGPT (actually, GPT4) was effectively give it knowledge of how the Trovata database is structured—not the data within it— which gave it a superpower.
It now knew that Trovata has access to transactions and balance information, for example, and the fact that Trovata knows how to create a cash flow forecast, etc.
Since we taught it that, it would be able to combine that knowledge with the query, or the question you ask of it, and then it says, “All right, now that I know that I can look it up, how would I look it up?”
Once we got this far, enabling the LLM to understand how to navigate within Trovata, the next step was to find a way for it to translate the output back without sharing any private information along the way.
The answer? Tokenization.
Using Tokens to Keep Data Private
To keep information private, we translated it all into tokens. This way, the LLM will never really know the language of what you’re communicating back and forth with it. All it knows is that there is a string that you’re looking up in a database that is structured in this specific way.
Then, the final little piece of the puzzle was, “Well, we don’t want to ever give it the ability to look anything up in Trovata, so how could we do that? How could we make it such that it only gives us instructions?”
Digging a little deeper, we realized that GPT4 is actually exceptional at programming. It’s excellent at producing code, and one of the languages that would be useful within the context of our database would be a query language, SQL.
This revelation led to our breakthrough moment.
Now we had the context, plus knowledge across the entire internet, and now it had the ability to give us instructions in a meaningful way, in SQL.
Finally, we put it all together and created a dynamic interface where you can type anything you want, and it then sends that in tokenized form, with the context that you need, to GPT4.
It determines, “Oh, this is what you mean. I know how to look that up in Trovata,” and then tells Trovata how to look it up. We then look it up on your behalf and display it within their Trovata AI interface.
“The data never leaves Trovata, it’s never given to OpenAI. That’s because the whole interaction is only one direction. It’s, ‘Hey ChatGPT, just tell me how to look it up.’ They never see anything else from there.”– Joseph Drambarean, Trovata CPO and CTO
How Did We Ensure Accuracy?
Okay, so we know that Trovata AI is secure and private. But that’s only half the battle, of course. In finance, if you’re wrong, you just might get fired. You need information that is reliable since the top level of the organization depends on it for making critical business decisions.
It’s not a secret that AI can hallucinate and come up with inaccurate information in its responses. Since we architected the tool in such a way that it is instructed to use Trovata in a very strict way, it is not able to come up with its own answers.
In addition, GPT4’s strength is not math. That’s why the calculations are all done via Trovata. Trovata AI is merely querying the database and then giving you the answer to that query in a very convenient way, skipping the steps that you’d have to take to get that answer yourself.
Scalability and Flexibility to Access the Best LLM
Last but not least, when building Trovata AI, we took into deep consideration how it could scale. We didn’t want to be constrained only to ChatGPT.
For example, if tomorrow, Bard were to release the best AI experience on the market (and a big part of it might be multilingual), we’d be able to make the jump. Since we have several international clients, this could very well be a reality.
“We’re keeping a close eye on the available LLMs because at the end of the day, there may be models that emerge that are much better for specific contexts. We happen to be using ChatGPT. It’s impressive as it stands, but let’s say that a year from now, there is a company that creates the world’s best finance-based LLM. Then suddenly, we can plug that in and have another breakthrough for our customers, still built on the foundations of our own AI technology. So that’s why that scalability really matters to us.”– Joseph Drambarean, CPO and CTO, Trovata
We have seen the tides change like this in the past with other innovations. For example, it might be hard to imagine nowadays, but Google wasn’t the first to search. First came the pioneers at AltaVista and Yahoo.
At the end of the day, we’re experimenting and innovating alongside everyone else in this wild, wild west of AI. We’re here to pay close attention to what’s developing in the space so we can help treasury and finance teams take advantage of it like never before.
P.S. – We just updated Trovata AI to be able to help you create forecasts. Book a demo of Trovata AI to see it in action today!