Why Most Financial Services Businesses Won’t Maximize AI Without Outside Help

AI is likely to have a significant impact on financial services and many enterprises within the sector have been among the technology’s early adopters. But most still aren’t properly equipped to take full advantage of AI without outside help, according to Invisible Technologies’ Head of Business Development, Financial Services, Addison Chu.

“We talk about AI-readiness or how equipped enterprises are to deploy resources into the AI space,” he says. “We see finserv businesses at all ends of the spectrum. But even the best ones won’t be able to integrate to an AI-first approach without strategic guidance and infrastructure support.

Becoming an AI-First Enterprise 

A 2021 McKinsey survey revealed the financial services sector was one of the most knowledgeable and engaged of any sector when it came to AI. Already by then, one-third of global financial services enterprises declared they had incorporated AI into their operations. 

This year, a subsequent McKinsey study revealed that as many as 24% of financial services employees regularly used AI for work, placing them second only to employees in the ‘tech, media and telecom’ sector. Another 18% of financial services workers said they regularly used it outside of work.

But for all their AI exposure, McKinsey also found that few financial services enterprises were deploying AI successfully.  Many lacked a clear and comprehensive AI strategy, had “a weak core technology and data backbone”, and “an outmoded operating model and talent strategy”.

McKinsey concluded that the financial services companies that were likely to survive and thrive over the next decade would be the ones that adopted an ‘AI-first’ approach to everything. This meant “adopting AI technologies as the foundation for new value propositions and distinctive customer experiences”. 

“Those that use AI already often apply it to siloed parts of their business,” Chu says. “They might use it for fraud detection, underwriting, or virtual assistants for customer service. But it doesn’t yet underpin everything they do.”

The Missing Ingredients in Financial Services

Chu argues that, despite their highly regulated operating environment, financial services businesses should really be looking to the examples of AI-first businesses outside of their sector to see what lessons they can draw.

For instance, platforms such as TikTok, Pinterest, and Spotify have all successfully embedded AI and machine learning (ML) into the fabric of their recommendation platform, allowing them to provide exceptional levels of personalization. This, in turn, has helped them grow rapidly, disrupt existing players, and become dominant in their respective markets. 

Chu says financial services enterprises could conceivably use AI in the same way - reimagining the customer experience to provide deep personalization and a seamless experience, especially in personalized financial management. 

“This has the potential to truly democratize financial service products and reduce today’s ‘financial information asymmetry’,” he says. 

“They could also use it to integrate all aspects of their internal operations, ensuring the knowledge and information from one part of the business is immediately understandable, available, and actionable by other parts and team members.”

But for that to happen, Chu says financial services businesses need some key ingredients in place. This includes having a data infrastructure in place that is ready to ingest a large language model (LLM). It also includes having a skilled machine learning team to clean that data and carry out the next steps.

“For those finserv enterprises with an existing machine learning team, moving to AI really is the next step. But for those without these existing and established resources, it will be more difficult,” Chu explains. 

How to Train Your Data

And yet, having a rich bank of good, clean data really is just the beginning, according to Chu. Enterprises next need to prepare their data for training, a process that begins with data annotation.

This involves subject matter experts labeling or tagging the data so that it’s understandable to machine learning algorithms and can be used to teach them how to make accurate predictions or decisions.

Only then is the data in a fit state to train an AI model.

The Machine Learning/AI lifecycle

To become AI-first, financial services enterprises need to go through four key steps, according to Invisble Technologies' Addison Chu and Adam Haney. These will need to be carried out by the machine learning team or an outside provider. 

1. Data collection and cleaning. Organisations with a rich bank of clean data will have a headstart on competitors. 

2. Data annotation. It’s only when that data is tagged that large language models (LLMs) can understand it. 

3. Model training/fine tuning. These intensive processes require the input of subject matters experts (SMEs), who make sure the model’s outputs are fit for purpose and accurate.

4. Model evaluation and deployment. Finally, the model must be rigorously assessed to make sure it aligns with the enterprise’s specific goals and needs.

Adam Haney, VP of Engineering at Invisible Technologies, says that enterprises have two real options to carry this out. The first is to start from scratch, developing their own AI model in the same way as Pinterest, Spotify and TikTok - usually a time-consuming and costly process. 

The second and more cost-effective option is to take what essentially amounts to a shortcut by tapping into existing large language models (LLMs) such as OpenAI’s ChatGPT, or Meta’s LLaMA 2, and refining them for their own use. 

“You may not be able to realistically get to 100 like a TikTok without significant investment, but you can get to 60 or 70 much more easily.”

To do this, the model must first be fine-tuned - a process in which a generalist model is taught the features and patterns specific to a task or domain. This involves SMEs instructing the LLM on the right response to questions involving the data.

“In the context of FinServ, the data that AI needs to understand and respond to could include anything from understanding financial statements and regulatory documents that need to be filed. But it can also really apply to any kind of document-heavy workflow,” Haney says. 

Beyond fine-tuning, there also needs to be a period of reinforcement learning with human feedback (RLHF), during which the SMEs stack rank five different responses. In doing so, they teach the model which answers align best with human preferences. 

“Without this step, the model will provide answers that are sometimes inaccurate,” Haney says. “When it’s done, you can have the model understanding exactly what matters to a financial analyst or customer service agent, or anyone within the business.”

“Invisible can help with the technical aspects of model selection and model evaluation. We can also capture the initial data sets needed to fine-tune those models, and ensure that, as they’re being fine-tuned, they still meet with company’s requirements,” Haney explains.

The Threat of Doing Nothing

As AI evolves and consumers’ habits change, Chu believes the pool of competitors for financial services enterprises will widen significantly.

“It’s conceivable that some other platforms will begin to offer financial products, and they will have the advantage of coming from a different risk profile to traditional FinServ enterprises, especially when it comes to AI.”

“Financial services is a ‘zero error environment’. Having AI make a mistake is not an option,” Chu says. 

“You can’t have a model ‘hallucinating’ or giving an incorrect answer because it could lead to a breach of regulations or privacy and serious reputational damage.”

“The writing is on the wall that FinServ enterprises need to move to AI-first. But it’s also clear that they need to do it in a way that minimizes the risk of getting anything wrong.”