Artificial intelligence (AI) has revolutionized how businesses operate, making tasks more efficient, processes smarter, and interactions more seamless. At the heart of many of these advancements is a type of AI called a large language model (LLM). If you’ve used AI tools like OpenAI’s ChatGPT or Anthropic’s Claude, you’ve already encountered an LLM.
But what exactly is an LLM, and why does it matter for your business?
To understand LLMs, it’s first crucial to know what an AI model is and how it’s different from traditional software programs such as Microsoft Word. A software program contains instructions — code — for the machine to tell it how to execute tasks, such as performing math calculations. However, it can only do what it is explicitly programmed to do.
An AI model is different. Most, but not all, find patterns in the data you give it — a process called machine learning. In an AI model’s training phase, its architecture provides a framework for learning, ingests the data you give it and finds patterns. Then, it develops its own internal representation of these patterns. With these internal representations, it can look at new data and analyze it — a process called inference.
Think of an AI model as a student and his textbooks as the data. The student learns from his textbooks and internalizes this knowledge. The student gets hired in a new job and applies this knowledge (for example, accounting) to analyze or handle new data — such as quarterly sales or other business information the company provides.
A large language model is an AI model trained on vast amounts of text, such as the entire internet. It’s as if someone has read millions of books, articles, blogs and messages in a dataset. The AI model learns to find statistical relationships between words and phrases through this training.
The model’s knowledge is encoded in its parameters, or mathematical values learned during training. The more parameters an AI model has, the larger and potentially more sophisticated it is. The LLM determines how words fit together, how ideas are expressed and how to respond to questions or prompts.
For example, when someone starts writing “how are,” the model predicts “you?” as one of the most likely next words based on common language patterns. Since it generates “new” content, an LLM is classified as generative AI.
Initially, LLMs were mainly text-based models that could not understand images, audio or video. However, they can become multimodal LLMs that process text and images as inputs and outputs. The later series of OpenAI’s GPT LLMs, like GPT-4, is a multimodal model. Text-only LLMs become multimodal when their architecture is modified, and they are fine-tuned to integrate different modes of content (audio, video, images).
Generative AI is a fairly recent development in artificial intelligence, whose creation harkens back to the 1950s or earlier. Companies have been using AI for decades; these older AI systems are rule-based, expert systems that have predefined rules and logic to make decisions and solve problems.
They also include early natural language processing and classical robotics and vision systems. Organizations still use older AI today. A general way to differentiate old and new AI is this: Old AI uses predefined rules, while new AI is not explicitly programmed. Rather, new AI learns from the data (via machine learning, neural networks and pattern recognition) and it is more dynamic, flexible and adaptive than old AI.
But with its probabilistic nature, new AI brings new challenges. It can hallucinate, or make things up. It can introduce bias in its responses, since it may inherit them from real-world data. They can breach privacy, mimic copyrighted works, and be misused in cyberattacks or misinformation. Another issue is that AI models consume high energy, raising environmental issues. There are also concerns that new AI will lead to job losses.
Another thing to know: In new AI, foundation models are large models trained (or pre-trained in industry parlance, since it is the starting point) on huge datasets for a broad range of tasks for a variety of applications. Examples of foundation models are OpenAI’s GPT series, Meta’s open-source Llama family and Google’s Gemini family of models.
These foundation models are usually trained again for specific purposes, like analyzing medical X-rays. This further stage of training is called fine-tuning. Fine-tuning ushers in AI models that are useful to different industries, be it healthcare, finance, retail and others.
With their ability to learn from vast amounts of data at a scale no human can replicate, LLMs are transforming the way businesses operate. LLMs enable organizations to become more efficient, cut costs, and enhance innovation, whether it’s automating routine tasks, enhancing customer engagement, or unlocking insights from data. However, businesses must use LLMs strategically by choosing the right tools, training employees to use them effectively, and staying vigilant about ethical pitfalls.
Use cases for LLMs include:
Agentic artificial intelligence (AI) promises to improve operational efficiencies and the customer experience offered by enterprises.
The advanced technology is finding applications in loan underwriting and fraud detection, and now it’s moving across borders.
TerraPay Co-Founder and Chief Operating Officer Ram Sundaram told PYMNTS as part of the “What’s Next in Payments” series focused on exploring AI’s use in banking and by FinTechs that automated decision making and streamlined processes will continue to transform global money movement, especially as faster payments gain ground in cross-border transactions. That’s the inexorable trend, but as Sundaram put it, there’s still room, and a necessity, to have some human interaction in the mix.
In terms of global fund flows, TerraPay’s single connection ties more than 3.7 billion mobile wallets together across 200 sending and 144 receiving countries, touching 7.5 billion bank accounts. As one might imagine, coordinating and enabling the transactions is complex.
“Obviously, in the best-case scenario, everything goes smoothly, but when things are not going smoothly, that’s when the customer queries come in,” Sundaram said.
It’s no easy task to find out straight away where a transaction is, as analysts and representatives at the company have to look at logs and query partner systems.
“A lot of that work is done manually,” said Sundaram, who added that the agents “know the corridors and the markets that they are working in, but it still takes some time.”
TerraPay is using AI models with machine learning to bolster customer support and automate tasks as financial institutions (TerraPay’s client base) send payments in real time, and those payments are processed into local markets’ beneficiary banks.
“We still don’t trust [AI models] to let them respond to the customer straight away, but we can do the analysis, and then that gets reviewed by an agent who decides if [information] is accurate or not and then sends it off,” Sundaram said.
The same principles are guiding AI models and company practices to improve technical and security operations, analyzing and categorizing anomalous transactions and automating integrations with partner firms.
“Compliance is an issue where there is a lot of review needed of the alerts, and we are using [AI models] to speed up those processes,” Sundaram said.
Asked by PYMNTS about how agentic AI can be harnessed, he said: “In financial services, you can’t take chances on technology like this, which has the freedom to go wrong. You have to be careful about making sure that it’s 100% reliable before we can let things run entirely by automation.”
Agentic AI also remains pricey. For example, OpenAI is charging $20,000 a month for its specialized agents. However, Sundaram said the industry will become commoditized quickly, which will lower prices, and some open-source offerings are capable.
“There’s a fire hose of news about breakthroughs and new ideas and new ways of doing things that are coming out on a daily basis,” he said.
Data underpins it all, and Sundaram told PYMNTS that no matter what the application, the information fed into the models must be clean. Most organizations have a range of data sitting in different intra-company silos, and those silos need to come down.
In addition, the data must be structured so that it is accessible and can be synthesized by the models. Many firms may have more than 1,000 software-as-a-service (SaaS) resources to which they are subscribed but are not accurately tracked or monitored.
“Every database is separated, each one sitting somewhere else,” he said.
The days of stitching together those separate SaaS offerings to run an enterprise are ending, he said, and we’re headed to a future when data is collected in one place.
AI models and agentic AI “are extensions of what we’ve always valued at TerraPay, which means building the most efficient infrastructure possible in order to make sure that transactions are processed safely, quickly and affordably,” Sundaram told PYMNTS. “We see AI and [AI models] as powerful tools that help us scale all this very quickly while making sure we build more and more efficiency into the system.”
