Philip K. Dick’s novel The Penultimate Truth, introduced a device called the “rhetorizor,” that the main character uses to help him write a speech. However, frustrated at the poor results, he quickly learns the importance of the correct prompts to deliver what he needs. Sound familiar?

What was once science fiction is now embedded in our everyday lives, largely due to the rise of large language model (LLM)-based chatbots like ChatGPT and other tools being integrated into software and apps we use.

Artificial intelligence (AI) is evolving from the first wave of predictive AI that analyses data and uses machine learning algorithms to forecast future outcomes, through to generative AI that creates new content like text, images, and music. Now, we’ve arrived at the Agentic AI stage – a model that generates content and is capable of being conversational and autonomously act and react.

But what exactly is Agentic AI, and how will financial services firms use it?

Agentic AI isn’t new. It's existed in practice since the 80s. Simply, it is AI that can complete actions in an environment – this could be a robot hoover or self-driving car to a chatbot ticketing machine. Recently, by which I really mean the last 12-18 months, the concept has been supercharged by LLMs (Large Language Models) and Generative AI.

This transformation is largely due to technological advancements that enable AI systems to process vast amounts of data more efficiently, resulting in enhanced decision-making capabilities. As a result, Agentic AI can now operate with greater autonomy, learning from its environment and applying that knowledge to perform complex tasks with minimal human intervention.

This has impacted two components: knowledge and action.

• Knowledge: AI models can handle a far greater variety of inputs, especially LLMs with unstructured text, and have also learnt to distil vast information sets and access what is appropriate for individual queries. DeepSeek-r1 is a great example of this.
• Action: Previously you would invest time defining in detail and managing how an AI model could complete actions. Now you can plug in pre-trained models that already know how to input data into cells, create formulas, iterate and improve them (in Excel for example).

You can also teach an agent what actions to take with reinforcement learning, a process where the agent is given access to an environment/system/programme and through trial and error with a reward when the correct outcome is achieved.

So, if you take a traditional definition of Agentic AI from something like Artificial Intelligence: A Modern Approach by Peter Norvig and Stuart J. Russell or modern players in the space like Tomoro.ai we can broadly define Agentic AI as follows. It is not just about data outputs it’s about taking actions and automating whole processes that might previously have been considered too complex due to the potential number of outcomes or the expert knowledge required to reach those outcomes.

Traditional AI, on the other hand, is very much focused on data outputs.

AI has always played an important role in financial services from liquidity and balance forecasts to credit decisioning and fraud prevention. In all these cases the AI is a focused element within a process rather than being able to manage and complete several different actions along the process or even end to end.

As an example, before really advancing Agentic AI and LLMs, the expectation of a chatbot would be to direct inquiries to the most appropriate teams or documents based on keywords and phrases. This is oversimplifying it a little as there were exceptions, but the general point stands.

With Agentic AI you can set up a multi-agent system that will be able to manage lots of different processes and domains of knowledge/expertise. Instead of being a part of the solution AI (with automation and assistant AI) can be the full solution for managing most queries.

Operational efficiencies will be the big focus for financial services, given the vast number of internally and externally facing processes that are managed by Operations teams. We will also see business units with repetitive tasks such as legal document reviews, copy for press releases and so on. Anywhere a consistent approach is required that can be described, AI can be trained to follow it.

Looking further ahead, once there is greater control over energy consumption and we see smaller, more focused and efficient foundation models, we will see AI used for full prototyping of products end-to-end. From market analysis to coding and code reviews, scenario planning and stress testing.

A common narrative in pushing back against AI, particularly in such a highly regulated industry as financial services is: can you trust the output? What if, for example, it provides guidance or action contrary to Consumer Duty rules?

Contrary to some commentators I believe we can trust it.

But just as you would a human there needs to be clear KPIs and monitoring in place and multi-agents improve what we can see and trust. This is because instead of one big model completing a huge task, if you break it down you can set up models with clear motivations and goals as well as expertise and expectations. You will be able to see the interactions between the agents and understand why an outcome has been reached.

We've been testing this approach within ClearBank as part of a wider program to understand how and where AI could enhance operational efficiency. Everyone had an agent that represented them in a multi-agent system. We set what motivations define us and then had them discuss things like what is a good holiday destination and what cheese is the best cheese. You can see full conversations and even arguments that are very insightful.

Another example is that many LLMs, if asked a mathematical question, will guess the answer based on a similar (or potentially the exact same) question they have seen in training data. An agent, however, can be trained to recognise a maths question and use a calculator. This overcomes a huge weakness with current LLMs.

What is vital is that the owners and users of AI models have a good basic understanding of how any model they are interacting with works. There needs to be a strong monitoring solution in place capturing as much data on the inputs and outcomes as possible. We are not in a place where there is standardised best practice yet, so we need to have enough data to take multiple approaches to find what works.

Finally, we need AI experts collaborating closely with the business teams using the AI to make sure the models are right for the job, don’t pick the latest and biggest model you can, set a clear set of criteria and pick the simplest model that hits them.

After its application and potential use cases, regulation is next big topic in AI. In the UK, we have seen a principle-based approach from the UK government and the fair outcome for customers requirements from the FCA drive the national view.

Critically it is the EU which is the driving force as GDPR and now the EU AI Act put requirements on the data firms can hold, how they use it for AI models. The EU AI Act takes a risk-based approach where the materiality of the outcome is more important that the complexity of the solution. There is no Agentic AI-specific regulation, but it will be affected by other AI regulation we are likely to see developing over the coming years.

Agentic AI is a developing part of the broader artificial intelligence space. So, it’s important for ClearBank to continue to invest time and efforts to explore its application, the best practices and stay ahead of any requirements rather than needing to retroactively bring them in under duress.

Based on what we have seen coming out of the EU, and what was coming out of the US, none of it is outside current best practices or what I wouldn’t expect from any good supplier of AI solutions.

All companies will need to have Agentic AI on their roadmap and the trend of focusing AI on operational processes to deliver greater efficiencies will continue as it’s the next logical step.

ClearBank is the same. We’ve had an AI Ethics policy in place since 2023 putting us ahead of the game. The focus of that policy is on accountability, transparency and fairness and, as such, it is applicable to Agentic AI as much as it is to GenAI and Traditional AI. We’ve already implemented a few exciting prototypes for us to test and learn and 2025 will be the year when we’ll see ROI on those.