74 Ways to Use AI | Special Edition

Banks are putting their biggest AI bets this year on internal tools that help employees automate routine tasks and easily retrieve information through chat and search, according to exclusive Evident analysis of banking AI use cases.

As the financial industry more widely adopts generative AI, Process Automation & Knowledge Access accounted for a third of all use cases and the most Gen AI activity.

Evident identified and analyzed the 74 use cases publicly released by the world’s top 50 banks tracked in the Evident AI Index from March through July of this year.

Also a big Gen AI focus: use cases that support developers in writing and reviewing code. Developer Augmentation accounted for the highest proportion of generative AI use cases (86%). That’s not surprising given the rise of tools such as GitHub Copilot (see “What AI Talent Makes”, The Brief, July 25).

While the Gen AI boom was reflected in the use cases that banks have rolled out this year, old legacy AI – the traditional kind that analyzes and predicts, not generates content – remains the standby for customer experience and fraud detection. But watch for financial institutions to deploy more Gen AI here too. Bank of America is experimenting with Gen AI on its Erica Chatbot but hasn’t unveiled anything yet.

Our Three Takeaways from 2024 Use Case Trends

OK, let’s go deeper. As banks focus their efforts on creating generative AI-powered tools to help staff get more efficient (see our lead item), we took a closer look at one recently unveiled Process Automation & Knowledge Access use case – DBS’s CSO Assistant.

What problem is CSO Assistant meant to solve?

The bank noticed their customer service officers spent as much as a fifth of their time searching for and collating information from across the organization. So it created this tool internally, without using an outside vendor, to improve efficiency and the quality of the customer service.

How does it work?

While the bank’s customer service officer talks to a client, the CSO Assistant transcribes voice to text (as below), using a Large Language Model (LLM) to summarize their asks.

Then it searches the bank’s sets of data and information – its so-called knowledge base – to give the customer service employee suggestions for what the client might want to do next.

After the call, CSO Assistant then auto-generates a summary of the call – including how the customer query was resolved.

How does this really work?

CSO Assistant taps into a vector database, which can handle complex data types and capture semantic relationships between words. It isn’t simply looking for the words themselves in internal documents – it understands their meaning. So for example a search for “daughter” could return a document about children, even if the document doesn’t mention “daughter.” It also uses a knowledge graph – a network of data that allows for retrieval of implicit information, linking concepts by their relationships (e.g. a daughter is a child of their mother). The assistant can then leverage this data through retrieval-augmented generation (RAG) (an approach that basically increases accuracy by fetching and integrating information from the available sources) and provide a better answer – even if the client uses imprecise words in their question.

Chief Data and Transformation Officer Nimish Panchmatia tells us DBS experimented with these approaches over several months to ensure accuracy and cut down on hallucinations.

Why didn’t DBS just go to an outside vendor?

Panchmatia says DBS wanted as an organization to learn about RAG, vectors and knowledge graphs.

Are employees actually using it?

All but two of the 20 CSOs involved in the initial pilot said the tool helped them work better, according to the bank, reducing call times by up to a fifth and delivering nearly 100% accurate results. CSO Assistant was then rolled out to every customer service officer.

What’s their return on this investment?

DBS says that narrow ROI is not most important at this stage and offers up this press release. They’re looking at how it reduces current and future costs, improves efficiency and changes work habits. There are also other benefits. What DBS learned from creating CSO Assistant is to be rolled out in use cases for different departments, like quality assurance and internal reporting.

"The risk of underinvesting is dramatically greater than the risk of overinvesting for us… Not investing to be at the front here has much more significant downsides."

—  Sundar Pichai, Google CEO, responding to an analyst question on their last earnings call

In every Brief, we look at the most interesting use cases put out by the top global banks. Here are the 20 that caught our eye so far this year, broken down by their functions.

CUSTOMER EXPERIENCE

Use Case: Generative AI Knowledge Support
Bank: CIBC

Question & answer tool makes it easier and faster for employees to find info to address client concerns.

Use Case: Customer-facing chatbot
Bank: ING
Vendor: McKinsey

Chatbot improves the customer experience with more immediate answers based on the intent, resulting in 20% more customers assisted in early usage.

Use Case: Customer Service Officer Assistant
Bank: DBS

Gen AI-powered virtual assistant transcribes customer queries live and provides answers to customer support staff, reducing call handling time by up to 20%. (See Under the Hood above)

PROCESS AUTOMATION & EFFICIENCY

Use Case: Debrief
Bank: Morgan Stanley
Vendor: Open AI

Meeting assistant that records, transcribes and summarizes discussions, and drafts follow-up emails; reportedly saving an employee ~30 minutes per meeting.

Use Case: Microsoft Copilot
Bank: multiple Evident AI Index banks
Vendor: Microsoft

Improves efficiency on everyday tasks via LLM-powered support within Microsoft 365 software. CommBank reported 86% of early testers said they wouldn’t go back to working without it.

Use Case: Gen AI Copilot for Advisors
Bank: JPMorganChase

The tool provides advisors support on manual tasks around client meetings, with the pilot reported to save advisors a couple of hours a day.

Use Case: AI @ Morgan Stanley (AIMS) Assistant
Bank: Morgan Stanley
Vendor: Open AI

Internal chatbot for financial advisors promises increased employee efficiency.

Use Case: Bookkeeper
Bank: NAB
Vendor: Thriday

ML tool helps small traders automate administration tasks; made available to its small business customers and reduces admin time by up to five hours per week.

FRAUD DETECTION

Use Case: NameCheck x Liink
Bank: CommBank x JPMorganChase

Banking collab to tackle payment fraud and errors: CommBank’s AI-based NameCheck security tool reviews and verifies account details, piloted on JPMC’s B2B Liink payment network.

Use Case: Generative AI Fraud Detection
Bank: bunq
Vendor: Nvidia

The neobank is tapping into Nvidia’s computing power to train its fraud detection model faster and with a larger dataset.

Use Case: SaferPay
Bank: Westpac

NLP tailors specific prompting questions for customers throughout the payment flow process to protect customers against fraud; losses from scams fell 24% in last year’s rollout.

INSIGHTS & DECISION SUPPORT

Use Case: Q Assist
Bank: HSBC
Vendor: Quantexa

Adding Gen AI to its Decision Intelligence Platform, searches across company databases to yield data from a wider range of sources – enabling faster, more accurate decisions.

Use Case: Moneyball asset management support
Bank: JPMorganChase

Pilot project puts Gen AI into the hands of portfolio managers, flagging questionable decisions and helps portfolio managers better run their investments.

Use Case: IndexGPT
Bank: JPMorganChase
Vendor: Leveraging GPT-4 (OpenAI)

Users provide criteria for thematic investing, and the tool generates relevant keywords to scan media and identify emerging trends, outputting a selection of target companies.

Use Case: TIFIN AG
Bank: RBC
Vendor: TIFIN

Insight generator for wealth management advisors to detect significant “life events” to better track clients’ financial behavior and deepen customer relationships.

DEVELOPER AUGMENTATION

Use Case: Github Copilot for Engineers
Bank: multiple Evident AI Index banks
Vendor: Microsoft

The ChatGPT-like assistant can be prompted to create new lines of code for particular programming tasks.

Use Case: Ensayo AI
Bank: ANZ
Vendor: HCL x AWS

The tool streamlines every stage of development, reducing API testing time by 72% and integration testing time by 56%.

IMPROVING WORKFLOWS, MODELS & DATA

Use Case: FlowMind automatic workflow generation
Bank: JPMorganChase

This approach to automatic workflow generation with LLMs mitigates hallucinations and ensures confidentiality, with 90-99% accuracy depending on task complexity.

Use Case: ML Pipeline for Debt Recovery
Bank: BBVA

The BBVA AI Factory develops predictive models for debt mitigation, anticipating customer challenges and offering solutions.

Use Case: Restructuring Open Banking Data
Bank: Mastercard

Using AI to structure a huge amount of unstructured data and help banks get more from messy data.

Don’t rush your data teams, says the Reliable AI Survey from data governance platform Monte Carlo: 89% of data engineers feel “some” or “a great deal” of pressure from leadership to implement Gen AI and 90% think their bosses have unrealistic expectations. Their message: Speed can make you sloppy, which can be costly – 70% of companies lost over $100,000 in a data incident, the survey found – and undermine trust in AI.

Resist the lure of an AI ‘czar’ at your company, writes Verizon Consumer CEO Sowmyanarayan Sampath in Harvard Business Review. This isn’t one person’s job. It can’t be imposed top down. Let ground teams lead you, along with “empowered stakeholders,” who work closely with the C-suite.

Don’t think about AI as just a cost-saving information technology, argues Ethan Molick, co-director at Wharton’s generative AI Lab. The steam engine made river travel faster, but revolutionized the world only when others found new uses for it. Same with AI. Ask the people who are experts in what they do – like your employees – to figure out new ways to use AI.

Use a product mindset with LLMs, argues Benedict Evans. There is a good and a bad way to use an LLM. Don’t rely on them for your facts. Learn how to use them. LLMs are multi-purpose motors.

So what do use cases really tell us about where the industry is at with this emerging technology?

Compared to others, banks are ahead of the curve. Goldman Sachs projects it’ll be the sector with the largest increase in AI adoption. Banks are investing as much as any sector besides technology, according to a McKinsey survey. Not doing so is not an option. Not testing is not an option. Those are beyond debate within this industry or most any industry (as this chart shows).

The big issue in 2024 is something else: How “integration ready” are you – how quickly are you able to move use cases from idea to production.

We’re seeing the most successful banks we track push hundreds, sometimes thousands of use cases through their pipelines. The efficiency gains should be meaningful. JPMorganChase, which tops the Evident AI Index, expects AI to deliver $1.5 billion in value this year.

Harder questions are and will be asked about the costs. CEOs have already had to dance carefully around market expectations. This investment won’t pay off immediately, they all insist. “This is a ‘26 and out benefit,” said Robin Vince of BNY. Or listen to Lloyds’ Charlie Nunn: “It’s going to take a few years for this to really be used, generative AI that is, on top of the existing AI tools.”

A little more sobering talk comes from David Solomon at Goldman Sachs: “Adoption rates will lag, the most fascinating use cases are in their early stages, and a lot of work still needs to be done in data security, regulatory frameworks and ethical considerations for the technology to reach its full potential.”

That’s where we think we are. But is this where we should be?

AI adoption isn’t going as fast as many expected it to be. The market selloff this week, propelled downward by the big seven tech stocks, is one reflection that the early hype hasn’t matched reality (yet).

We’re not seeing banks push consumer-facing AI; these are still in the “experimentation” phase. We’re also not seeing much automation. When ChatGPT first launched in 2022, there was great fear and excitement at the prospect of mass job loss to AI-powered automation. Current systems lack the reliability to be entrusted with sensitive or complex tasks. Agentic AI and other approaches are aiming to solve these problems, but these are still firmly in the domain of research and pilots.

But remember: AI adoption has been happening for years. Leading banks were setting up AI centers of excellence and hiring AI leaders back in 2018. For many, the journey started with digital transformation efforts dating back to the early 2010s. Last year brought a post-ChatGPT flood of exciting news about Gen AI and other revolutionary features. 2024 is the year of the grindstone – the work of testing and adopting AI by business. It’s slower, less sexy. But all the same, based on what we’re seeing, transformative.

Do you know or run an event that you think should be featured? Let us know at [email protected].