Banks Adopting AI that Don’t Rethink Processes Leave Productivity — and Money — on the Table
By Steve Cocheo, Senior Executive Editor at The Financial Brand
Simple Subscribe
Subscribe Now!
The banking industry is pouring money into AI adoption, moving rapidly from pilot efforts to broad usage. Recent research by Accenture indicates that 84% of banking leaders say they’ve seen at least moderate business value from the AI they’ve implemented so far.
However, there’s a big qualification to that statistic: Only 20% of those leaders say that their AI investments are producing broad and sustained value for the money.
Why this is happening: Too much of the industry’s AI investment has been governed by fear of missing out instead of a reasoned approach to getting the best out of AI technology, according to Michael Abbott, the firm’s global banking lead.
Abbott says leaders feel the pressure to keep pushing AI efforts forward, increase adoption, and ensure that their staff learn to use AI tools. As a result, he says, all but a very small minority of the business is overlaying AI on the same approaches they were taking beforehand.
Beyond that, many players attempt to use AI’s big guns, the latest large language models, for everything. Yet, other simpler AI tools would often do a better job with many tasks. In fact, Abbott says that recent experiments in which Accenture took part found, when AI is merely applied to individual processes within larger processes in banks rather than reconceived approaches, the result may be a drop in overall productivity.
Need to Know:
- Accenture’s Michael Abbott says most major players have moved beyond pilots to widespread application of AI tools, but they haven’t changed how their banks operate.
- The top reason banking leaders are ramping up AI adoption is to maintain or strengthen their competitive position, per the firm’s research. Many don’t consider the return on AI investment at this point.
- Many banks have leaned into the latest large language models in the belief that state-of-the-art tools are the best. But Abbott says smaller, more specialized models make more sense for many tasks.
To realize AI’s potential, banks need to pull must pull them apart and put them back together in configurations that take advantage of AI’s strengths and the way it best handles banking operations, says Abbott.
Abbott maintains that until banking institutions make the shift — and only a few that he’s worked with are doing so — productivity gains will be minimal.
“Maybe you’ll get to go home a little earlier,” says Abbott. But that’s about it.
How Banks Should be Approaching AI Implementation for the Long Haul
Bank processes have traditionally been linear, sequential in nature, according to Abbott, every step leading to a next step. That’s logical, and also very human — one foot in front of the other.
Why this matters: But Abbott says linear thinking is part of what’s been holding banking back. Many business areas need not be frozen into sequential approaches, he says. Often they can be re-arranged so that multiple parts of the overall process can be addressed at once, in parallel workflows.
What bankers must do: “AI is forcing the industry to think about everything in parallel,” says Abbott. “So organizational approaches are going to have to change pretty dramatically.”
Read more: How ‘Watson Era’ Thinking Is Holding Back Banks’ AI Benefits
Understanding the Parallelism Shift
Abbott has been trying to come up with an analogy that illustrates how the industry must evolve. He says he’s found one in the historical development of factory operations.
When water power was the best option for mechanization, an entire mill might be operating off belts or other forms of transmission from a single water wheel. When steam supplanted water power, taking off power from a central steam engine was much the same.
However, when factories could tap multiple electric motors, each work station could proceed independently using its own dedicated tools powered by dedicated motors. AI elements can follow a similar approach — enabling much more to happen simultaneously and truncating processing timelines by assigning each AI agent to a different task.
“When you rethink the work in new ways, you get productivity gains at a larger level,” says Abbott. “It’s all about the reconfiguration of the manufacturing line. You’re doing the same things, but the work is going to be accomplished in a radically different way.”
The implications: “I am becoming pretty convinced that it’s going to be a paradigm shift in how banks are run,” says Abbott.
How a Few Banks are Making Processes Work in Parallel
However, most banks aren’t up to the stage of rethinking their processes for AI-centric work. But he’s worked with a handful that are doing so.
Case study #1: The classic mortgage workflow is extremely sequential. The loan application moves square by square through the usual process as documentation is gathered, financial information compiled, credit evaluation conducted, and so forth.
With AI, the process doesn’t need to be that way.
“You can ‘parallelize’ the entire mortgage process and almost pre-approve everything all the way down,” he says, with holes being filled later in the flow, such as obtaining missing documents.
The key is that each part of the mortgage flow is controlled by a separate AI agent, according to Abbott. “Every part of the process can look forward and backward instantaneously,” he says, “and they can communicate immediately, because they are agents, not people.”
One Australian bank has already managed this shift. For this bank, creating a truly AI-driven process required breaking down walls between the separate parts of the bank that controlled various parts of the mortgage process.
Case study #2: A Latin American bank attempted to create an AI-based universal agent that could handle any call that came into the call center.
Along the way, the bank had a critical realization: Phone calls consist of a series of intents. Developing AI that can handle each individual intent creates a series of blocks that can assembled in multiple ways beyond automating a call center.
For example, consider a lost or stolen credit card as an intent. AI can serve that need through an AI-automated call center as well as other channels, including the bank’s mobile app, online banking service, and the desks of branch customer service representatives who can leverage AI assistance to handle in-person lost or stolen card queries.
“So there’s a complete reconfiguration of work with AI technology from journey to intent, and organizing around the intent of the customer,” says Abbott.
Key warning: Just breaking up a traditional process into sub-processes and applying AI isn’t the lesson to take from these examples. Research conducted by Accenture with G5 Labs.ai found that merely doing that doesn’t generate productivity gains. (The experiments used software development via AI as a test case.)
Read more: AI Can Help Banks Preserve Institutional Knowledge — Or Scale Your Worst Workarounds
Use the Right AI Tool for the Right Task
The AI tools that garner the biggest, most frequent headlines are the “frontier models,” the most-powerful AI models of the moment, trained on huge amounts of data and content.
However, for some tasks, they are not only overkill, but can result in errors, absolutely verboten in the banking business.
“You don’t want to wake up and have your bank balance be different every day,” says Abbott.
At play here are two sides of algorithms:.
• A deterministic algorithm that will consistently deliver the same result. In other words, two and two always equals four, no argument or variation.
• A stochastic algorithm that includes a degree of randomness. Large language models have this degree of chance in them.
For many tasks, small language models — special purpose-built tools — make more sense for specific tasks as banks revamp their processes to better match AI capabilities. Abbott says many of the banks that Accenture works with are taking large language models out of certain processes and replacing them with small, dedicated models.
“For example, if you need to ingest somebody’s tax return, you don’t need a frontier model,” he says. “In fact, it’s almost criminal to waste that amount of computational power on it. And because the large language models are so general-purpose-built, they can tend to hallucinate.”
Why this matters: “Banks live in a regulated world that needs to be deterministic, thoughtful, documented, controlled — and understandable to the regulators,” says Abbott.
How Humans Will Fit into the Picture
Even as AI agents take over more tasks, and especially as banks adopt teams of specialized agents, Abbott says human managers will need to manage the overall processes with connective intelligence and common sense.
Fulfilling this role requires an ability for critical thinking and the willingness to say that an AI agent is wrong.
In fact, Abbott knows of a firm that tests for such abilities when considering candidates for new positions. The company gives them a development project and a large language model to use.
The trick of the test is that the model is intentionally flawed.
How does a candidate succeed?
By spotting the mistakes the model makes and devising prompts that correct them.
“That’s the logical mindset that’s going to be needed,” says Abbott. “Otherwise, you will think it’s right all the time. And that’s not going to work.”
Read next: Who Will Protect Banking Consumers’ Rights in the Age of AI?
