The word "copilot," which used to just mean a person who helped fly an airplane, has been co-opted by a host of technology companies, from mainstream vendors like Microsoft, Google, Salesforce and ServiceNow to financial services software providers like Blend, Q2 and Intuit. (Some call the technology an "assist" or a "sidekick.")
These copilots use advances in generative AI to give employees a level of automated help that was not possible a year ago, by making information easier to find and by actually performing certain simple tasks. They typically use some combination of natural language processing, large language models, knowledge management, workflow and application programming interfaces to existing software to do this work. By calling these programs copilots, tech vendors take some of the scariness out of deploying advanced AI.
"Branding this as a copilot is kind of taking a step back," said Gilles Ubaghs, strategic advisor at Datos Insights. "It's giving banks much more sense of control, of human oversight."
When banks think about using artificial intelligence, they don't want the Terminator, they want Robocop, a client recently told Ubaghs.
"They want the human inside of the superpower suit. And that's where the copilot comes in," he said.
In banking as in most industries, people are not ready to fully replace humans with AI. A copilot works alongside a pilot, but the pilot is still in control. Many banks are testing this idea.
The trend "feels like something that is going to be around for a while, so we should explore it eyes wide open," said Andy Max, senior director, data enablement — analytics and insights at First National Bank of Omaha. "That's the approach that our bank is starting to take, one use case at a time."
At Valley Bank in Morristown, New Jersey, "we think about copilots as extremely useful because they are naturally human in the loop, enabling our relationship banking value proposition," said Stuart Cook, chief innovation officer.
In banks, copilots can help developers generate code, help call center agents answer customers' questions, help loan officers create loan offers, help investment bankers put deals together and help salespeople and financial advisors understand a customer's history, preferences and transactions and suggest next best actions.
Contact center copilots
At KeyBank, Jordan Olack, senior vice president of intelligent automation and contact center delivery, thinks about how to use copilot technology to help contact center teammates provide better service. The Cleveland bank has 1,200 contact center agents, most of whom have worked from home since the Covid pandemic began.
"We view copilots as not a replacement of those folks, but as another tool in their toolbox to help them do their job better and ultimately help our clients with their financial wellness with us here at KeyBank," Olack said.
Contact center agents have "a million and one different products and services and at any given moment, they have a customer on the other end that is in need of help and emotional about it," Olack said.
There are many different places the reps need to look to find information about credit cards, mortgages, deposits, bill pay and the like as well as policies and procedures, Olack said.
"If we could leverage the power of copilots, which is artificial intelligence to help them curate that knowledge or access information on these source systems, that is where we think copilots are really going to add a tremendous amount of value," he said. "That's sitting in the cockpit with the contact center agent, trying to help them in the moment deliver crazy good service."
The bank is working on risk management, compliance and model risk, to make sure the artificial intelligence in a potential copilot is safe and responsible, he said. KeyBank has set up an enterprise advisory committee to establish governance over this technology.
"I think this is going to be a big advantage for our teammates here at Key," Olack said.
KeyBank already provides some AI software to its contact center reps. For instance, when the Paycheck Protection Program launched, Olack's team created a virtual assistant to provide information, FAQ-style, to the customer service representatives. That rudimentary virtual assistant created in 2020 has evolved to be an assistant that can help with all business banking information.
Large language models and generative AI will only make such virtual assistants more useful, Olack said.
"That's where the copilot becomes a real advantage," he said. "In essence, you really do have two pilots flying the plane as opposed to an understudy that really doesn't know everything."
But his team is already thinking about the use cases for testing and learning. They are considering three main examples: knowledge management for contact center agents, generative coding for internal software developers and help in conducting anti-money-laundering investigations.
A wingman for loan officers
At First National Bank of Omaha, loan officers have been using PrecisionLender software to price commercial loans for about nine years, Andy Max said.
In 2017, FNBO began using Andi, a virtual assistant within PrecisionLender's software that makes suggestions and asks questions. Max thinks of it as a pricing coach.
In 2019, digital banking software provider Q2 bought PrecisionLender and, recently, Q2 added AI capabilities to Andi and made it more sophisticated and useful, Max said. He has not tried it yet, but he is intrigued and is considering testing the technology.
FNBO's commercial loan officers spend about half their time advising clients, the other half on loan-related tasks.
"We're looking for any opportunity where that type of work can be automated," Max said. "So, we're excited about this new wave and cautiously optimistic."
AI copilots could be used to make customer onboarding more efficient, Max hypothesized, eliminating the need to re-key information multiple times in different systems. It could make sure a loan application goes through the pipeline accurately, avoiding back steps or rework.
"It can just ensure and be an effective communicator that we have all the documents," Max said. "All of these simple things seem like they're easy, but they involve a human today. And these are things that are fairly low-hanging fruit that we believe, as we start to tip our toe into this, can definitely save our teams a lot of time and frustration."
He is still at the exploring stage with the technology.
"We don't want to get out in front of our skis," Max said. "We know there's a lot of value here, but we have to get comfortable with it from a risk standpoint, from a compliance standpoint, from a general bank appetite."
If the Andi software passes these hurdles, the bank may start by rolling it out to ten relationship managers for a few months of testing.
Q2's copilot offers a more intuitive way of letting humans interact with AI, said Corey Gross, vice president of Q2's AI Center of Excellence.
"A GPT-like interface can address real human needs so that humans don't have to spend as much of their time behaving like computers," he said. "They can do human stuff and let this GPT-like interface take on more of those tasks to reduce time and sweat."
Q2's copilot combines aspects of a virtual assistant — it can answer basic questions — with the ability to execute tasks in other software applications, an activity that used to be the purview of robotics process automation.
"It might say, hey, there's this thing you want to execute in an application over here, I can do that for you, and then I can do the next thing for you. Are you OK with that?" Gross said. "And then the copilot is confirming, yes, I'd like you to do that. Or, hey, wait, what happens if I do this? And it can have more of that contextual conversation with you so that when you do execute the next action, you were in the loop the whole way."
Q2's Andi copilot was built using OpenAI's GPT-4 large language model. According to Gross, the company has also been testing it with competitors Anthropic and Bard. The copilot might handle four or five tasks within the same workflow.
In addition to helping commercial loan officers price deals, Andi can help remind them of the bank's policies and procedures.
A generative AI-powered copilot is only going to be as good as the knowledge it builds up over time, Gross noted. With enough knowledge, it can predict the next best action with great accuracy, he said.
Gross thinks of large language models as akin to Ron Burgundy, the main character in the movie "Anchorman."
"He doesn't have any knowledge, which is why in the movie, when Christina Applegate changes the cards around, he says something regrettable on the air because he is not even really thinking about what he's saying," Gross said. "An LLM doesn't have the capacity to actually think about what it's telling you. It's predicting the next token in a chain that it believes you want to hear. And that's why it sounds so compelling, and that's why it's so easy for hallucinations to be believable. So in order for a translator like Ron Burgundy to be valuable, he needs a pretty strong source of knowledge. That's what Andi is."
Software company Blend also has a copilot for loan officers. It uses generative AI to understand the intent behind the questions customers ask in natural language, said Nima Ghamsari, founder and head of Blend.
Potential borrowers often have complicated questions, like, "If I wanted to go see a showing tomorrow for a $300,000 house, would I be pre-approved for that amount?" Or, "that showing went really well, I want to make an offer. Can you give me a pre-approval letter so that I can send it to my real estate agent?"
While the loan officer is getting coffee, the copilot could read the customer's question, interpret it, do the required research or tasks and provide an audit trail of its work, Ghamsari said. The loan officer can come back to his desk, review the work to make sure it makes sense and provide the answer to the customer.
"It's taking the existing workflows that are happening today and putting a mini assistant in between that's doing the first pass at all the work for the team internally, so that the consumer can get better answers faster," Ghamsari said.
Blend uses large language model technology from multiple companies for its copilot, including OpenAI's GPT and Amazon's Bedrock and Claude.
Help with general office work
Microsoft was one of the first companies to roll out AI copilots and popularize the idea. The company's recent product announcements have been dominated by copilots, including ones for Dynamics 365, Microsoft 365 and GitHub.
Such copilots are intended to save people time on general office work, too, for instance in summarizing text and pulling out bullet points.
"The thing will just spit out the first draft for you," Ubaghs said. "The letter that may come out, your draft sales email, might be terrible. It might be the blandest thing. But at least it gives you that framework. It can still save you a ton of time."
"You can't get through a topic of conversation without generative AI coming up," said Bill Borden, corporate vice president of worldwide financial services at Microsoft, who has had many conversations with senior executives at financial institutions about this.
"There's the excitement of wow, what this technology could mean in terms of the ability to actually have large language models that can be used and interface with natural language and what that can mean for the industry," he said.
Bank clients are asking for a controlled environment where they can learn fast and understand security and privacy controls, Borden said. Microsoft offers enterprise versions of OpenAI's ChatGPT and Meta's Llama large language models for this purpose.
Microsoft's Github copilot, which helps developers generate code, has achieved 40% to 50% productivity gains, Borden said.
"That was really the spark behind the idea of copilot and the creation of our generative AI approach that says, we're not looking at replacing workers, we're looking at augmenting workers," Borden said.
Intapp, a provider of software to law firms, accounting firms, consulting firms and investment banks, is partnering with Microsoft on an AI copilot that would help people put deals together and handle related tasks like compliance.
All such copilots are "like having a grad student next to you who is in your environment and learning about your company and what it can do and what products and services it can offer," Borden said. "And as a grad student, you have to tell it what it can learn, you have to tell it how to approach it. But you would never send that grad student out to your customer and say, 'Hey, here's what you need to do. I'm a grad student, I know what I'm talking about.' A knowledge worker is still going to work with that copilot to figure out how you can jointly bring value together for that customer to solve that problem."
