Understanding what AI language models and RAG can do for financial professionals
As a financial and technology professional and founder of Creative CFO, I’ve spent years studying how to access information efficiently, interpret that information and produce coherent answers to contextual questions posed by business owners and managers of high-growth SMEs.
I recently came across a podcast by Lex Fridman with Aravind Sriniva (the CEO of Perplexity), which gave me several lightbulb moments and sent me down a tech rabbit hole to see how I could test the concept called RAG (retrieval augmented generation) for use in questions often asked of financial professionals and tax practitioners.
First, the basic concepts of how AI language models and RAG work:
A digital assistant takes a question, references it against the specific documents you have uploaded (the ‘RAG’ part), and passes on the question + relevant document extracts to a language model (e.g. ChatGPT) that gives you an answer to your question using natural language.
It sounds simple enough, although the underlying technology is clearly pretty amazing!
What is the role of the financial professional in setting up and maintaining an AI-based RAG assistant?
At Creative CFO, we’ve always sought to use the best tools for the job, from the laptops we work on to the software we use.
Using language models and RAG-based assistants is just another step in this path. There are four key areas where I see major value adds:
- Identifying the best models and assistants for use with financial data and related questions
- Ensuring the data are structured in a way that leads to the most efficient and accurate retrieval
- Providing system instructions and background training data so the models perform in a way we would expect a professional to
- Setting up and making these tools accessible to our customers in a cost-effective manner
Beyond that, technology moves quickly, so monitoring and enhancing the functionality on an ongoing basis is also part of the role.
Since the quality of the analysis is still driven by how well the data are structured and whether they are up-to-date, a professional finance team plays a key role.
Most businesses still have manual data collection points (e.g., capturing slips, doing stock counts, ensuring new bank accounts are loaded onto Xero, etc.), and that means ensuring those data processes are run and monitored for quality. The finance team’s weekly bookkeeping and control checks are needed to ensure the reports are up-to-date and accurate for analysis.
Of course, many of these tasks will become automated, but for now, there are still many manual steps in gathering and processing business data, and the quality of reporting heavily depends on the checks and balances put in place by professionals who understand business and financial processes inside out.
At Creative CFO we’ve always built from the ground up, ensuring high-quality data capture, streamlined financial processes and reporting that financial professionals and business owners can use for strategic insights, and these tools only enhance our model for building the world’s best finance teams.
Podcasts I found interesting on the future of AI, searching for knowledge, and concerns over safety and security
Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434
The podcast was Lex Fridman’s interview with Aravind Srinivas. It will help you understand the history of search, what Google Search did (and still does) very well, the development of AI and language models, and what is impacting the speed of that development to answer questions and solve the problems we put to it.
I recommend you look at Perlexity AI. It’s a really good example of how search and answers can be improved using language models, specifically it:
- Takes your search prompt and looks at a few ways to phrase the question to get the best results (if on Pro*)
- Reviews publicly available websites or any documents you upload (if on Pro)
- It provides an easy-to-understand answer using a language model (which you can choose, e.g. Claude, ChatGPT, etc)
*The Pro setting is currently available free for five searches a day.
Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
In this Podcast, Lex talks to the CEO of Anthropic and two of the team members, who focus on the personality of Claude’s models (Amanda Askell) and how to ensure the security and integrity of the models through mechanistic interpretability (Chris Olah). It really opened my eyes to the way people are weighing up the trade-offs of the model’s ability to help or harm and how it’s more of an organic process than you might think.
An Interactive, Short Course from Google Cloud on Vertex AI
Even if you manage just the intro section of this free course, it will teach you a lot about AI in general, as well as terms like multimodal, prompt design, and model tuning, such as the ‘temperature’ to control the degree to which the model sticks to the script (or data provided).
We work primarily with the Google ecosystem, and they are putting together a lot of really interesting content on how AI and language models work and how to leverage Google Cloud to create applications.
- Access the course here: Introduction to Vertex AI Studio
A Book from Ethan Mollick called “Co-Intelligence”Co-Intelligence: Living and Working with AI”
Ethan Mollick is from the Wharton School and has written extensively on AI, followed its development, and considered how to educate professionals on how to constructively engage with AI and successfully integrate it into their field of expertise.
His basic principles in the book are:
- Always Invite AI to the Table: Consistently use AI for various tasks to understand its capabilities and limitations.
- Be the Human in the Loop: Maintain active human participation, providing oversight and addressing ethical considerations even as AI capabilities improve.
- Treat AI Like a Person (But Tell It What Kind of Person It Is): Interact with AI as if it were human-like, but define its role and personality to enhance interactions.
- Assume This Is the Worst AI You Will Ever Use: Prepare for rapid advancements in AI technology by assuming current AI is the least capable you’ll encounter.
I found this a positive way to move away from the hype and engage with the technology, using your expertise to test its level and where it can start to add value in your field.
Demo Video – an OpenAI assistant using SARS guides to answer a tax question
We built an AI portal so the team could test the latest AI models and assistants. Here is a short video of what an OpenAI assistant can do with some SARS guides and a standard tax question.
Thanks for reading
Get in touch with moc.ofcevitaerc@pihsrentrap if you’d like us to help you enable the finance function in your business to enable a high growth trajectory.