Organizations have faced the challenge of deriving insights from their data for a long time. Some enterprises have the ability and resources to do this, but others are far behind. Artificial intelligence (AI) has the capability of catapulting data analysis into the future, allowing enterprise analytics to fit into the daily, general health and success of a company.
Billtrust has been at the forefront of using AI to build out analytics processes, especially within the payments landscape. In a recent PaymentsJournal podcast, Ahsan Shah, Billtrust’s Senior Vice President of Data Analytics, talked about the AI-fueled future of data analytics with Christopher Miller, Lead Analyst of Emerging Payments at Javelin Strategy & Research.
The Democratization of AI
Organizations can no longer say they are not looking at AI. The success for most is going to come with the democratization of generative AI as opposed to a top-down mandate.
“Some companies are more advanced than others, just by allowing people to try it in the form of their goals and their own self-training,” Shah said. “Some of our teams here at Billtrust are doing hackathons where they just learn how to do this amazing thing. I think it’s going to flourish organically, and I think that’s the right way.”
AI is poised to go from a foundational model universe to a large set of tools, tooling, infrastructure, and services. The technology advancements are moving much faster than the rate of adoption. OpenAI is already at the forefront of multi-modality.
“There has been an explosion in the number of different systems that are monitoring various parts of how a business operates, ranging from frontline customer success to the nitty-gritty details of actual payment processing or chargeback processes, all the way up to when is revenue recognized and how is cash managed,” Miller said. “One of the challenges for teams has been to figure out how to put together those different pieces.”
An Explosion of Data
Most companies ask someone to piece together various pieces of information or cut and paste some data in a spreadsheet. Maybe they have a dashboard that brings together different pieces, but even maintaining that dashboard, adding new data as it comes to the forefront, can be a challenge. The explosion of data creates opportunities for insight but also challenges in terms of the sheer scale, especially for organizations with limitations in teams and resources.
This idea of cross-functional analysis is a challenge not just because of the volume of the data but also because of its structure. “You have three different kinds of vectors happening here,” Shah said. “You have the insane amount of data, the urgency of trying to act on it, and the explosion of the different functions. Enterprises need a better way of synthesizing the data across the functions and to be able to get it to the right person who can act on it, which is often overlooked.”
Emerging generative AI technology may offer one way to solve some of these problems, such as a new way to create reports other than simply handing a definition to an engineering team that produces the report. Rather than being pushed from the systems, data can be pulled from the systems by precisely the people who are in a position to act on those insights.
The new term is generative BI, for generative business intelligence. You can simply ask a specific question in human language, such as “What anomalies are you seeing in my payment patterns for buyers in the West Coast?” That’s something that traditionally would have taken weeks of engineering analytics.
“It’s an exploding space,” Shah said. “Six months ago, there might have been one or two names that had LLM products in market that we could use. Everyone had written a poem in ChatGPT and experienced firsthand the power of the language model. But most people had also run headlong into the challenges of the data-gathering side of that model, which offers an interaction layer and doesn’t necessarily offer the insight. That’s the next step.”
Moving Beyond ChatGPT
Users of ChatGPT are limited to the context window. You can type in your question, but the tool doesn’t know about you, your enterprise data, your CRM, or your transactions. Integrating the data layer and the analytics layer into the LLM directly requires engineering and domain fine-tuning of the models.
There’s only so far you can go with a foundational model. How do you expose and make your data scalable and engineered in a way to take full advantage of generative AI? That is something Billtrust is actively working on.
“We are in the process of launching our Copilot product, essentially embedding a ChatGPT-like enterprise secure interface into it,” Shah said. “Rather than going back to the old way of hiring a data analyst and saying build me a report, you’re now going to Copilot and asking a specific question. We should not think of this as a profoundly transformative thing but rather a way of making what you do better.”
Some companies are already blazing through the capabilities. It’s not just Open AI, but also Facebook Meta and AWS and Claude Anthropic integration. You’re going to be hearing a term called agentic workflows.
“While this seems super forward-looking, I don’t think it’s that far ahead at all,” Shah said. “You’re going to see a universe where people are going to log into SaaS products or B2C products and simply ask it, “Book a trip for me and my family,” and it’s just going to do a multi-step flow to book your hotel. You could translate that to B2B now. Instead of booking a travel reservation, you might say run a campaign or target these customers.”
The Need for Governance
When systems act based on limited cues from human beings, the interoperability of those systems becomes critical. This suggests the need for standards and essentially another layer of API development.
“It’s important to have governance to avoid the problematic and even catastrophic implications of AI,” Shah said. “But it cannot be done in a way which impedes the ability of companies to innovate and build great products.”
One other concern is cost, which is high and still going up. The unit cost is slowly starting to bend, but the absolute cost is growing as the models exponentially add tokens, which creates additional computing demands to support them.
But the possibilities far outstrip the challenges. “You’re only limited by your imagination,” Shah said. “The best implementations on the agent level will create the biggest universe for that imagination to run wild. It’s almost like giving an artist the capability to focus on what they’re best at and removing the friction or the redundancy of other tasks. The technical capability will be there far before the implementations are there to support that kind of imagination.
“There’s going to be an entire knowledge of how to use different models effectively for different businesses. I see this explosion of options. It just might be a little bit of a zoo for a while till the dust settles.”