ChatGPT, now nearly two years old, made obvious a pent-up demand for a new way of working. This new way is more intelligent. It provides answers instead of links. It gives back precisely what you are looking for, and, crucially, it eliminates so much of the noise that search engines or other tools provide.
ChatGPT demonstrated great potential. We saw in it a way to use natural language to change the way humans interact with data. We envision LLMs as the new, preferred method to essentially act as translators between a person and a data system, and even stubborn or complex databases are now open to non-technical users through our tool Conversational Finance.
But our vision did not end there, rather we sought to build more than a new tool, an agent, which enables a new, simpler way to work — this agent can use Conversational Finance to answer far more complex queries, perform significant analysis and come back within a few minutes with a full answer.
Today, we’re unveiling this agentic approach to FP&A.
About the agent
Introducing the reasoning agent: You can get analysis from financial data with a few simple questions. The reasoning capability builds on top of the components of the hila Platform and hila Conversational Finance, including our anti-hallucination technology, various fine-tuned models, and the domain-, company- and user-specific knowledge we have.
Using our Conversational Finance system, the agent gets accurate answers for finance and business users to even very complex questions. For example, when asking for office expense variance, the agent breaks it down automatically into five questions, including:
Each of these answers comes with an accompanying graph and the steps that are used to solve the problem
Generative AI may have a lot of promise, but…
A new paper from researchers at Google, University of Waterloo and Salesforce (among others) reveals that most systems, even when getting hundreds of examples, cannot achieve more than 17 percent success overall, and fail nearly nine out of ten times for complex questions. This performance gets worse with key analysis, such as DBT tasks, inclusion of external documents, nested schemas, or handling different SQL dialects.
What becomes clear immediately is that one has to have the kind of differentiated platform such as Conversational Finance, with significant engineering to understand the user’s intent, disambiguate the answer, run a confidence check on the answers, have domain knowledge, run many fine-tuned and purpose-built models, and so much more.
Time, Accuracy and Cost
Our agentic approach builds off of Conversational Finance, which is deeply optimized. As such, it costs less than 1 cent per query. Running cost including infrastructure and LLM cost can be less than $10 a day.
It uses the tools that we’ve honed through several customer engagements, including our patented technologies, to eliminate hallucinations. It utilizes our multiple LLMs that ensure financial, company and user knowledge are integrated into the answers. And it provides the security on the data to ensure that there’s no customer data that leaves their firewall.
Get in touch to see a demo of this new capability in action.