CBA's AI Companion is expected to face some hard questions

Commonwealth Bank is anticipating that its in-app AI assistant Companion will be privy to questions and information that customers may not ordinarily share through traditional support channels.

(L-R) Michael Baumann (CBA) and Alison Hobbs (AWS).

What makes the AI assistant useful places additional responsibilities on the bank and the cross-functional team building out Companion, and this is already reflected in some design decisions for it, executive general manager of customer, digital and AI Michael Baumann told the recent AWS Financial Services Summit in Sydney.

Companion was unveiled in late May and is currently in limited customer testing, both consumer and business.

The bank has broad ambitions for Companion, although initially the intent is to “bring together live spending and saving data into a single experience, so customers can be better informed and manage their money, and act with confidence,” it said back in May.

So far, about half of the questions that Companion has been asked align to this use case.

“Fifty (50) percent of the questions that we get in the Companion right now from customers are about spending and saving - and it’s better understanding how you spend your money and then also how you can save more money,” Baumann said.

Baumann noted that these types of questions are not typically the domain of branch or call centre staff - the clear imputation being that AI is not going to replace traditional support or interaction channels.

“[Companion] is doing things that branch or call centre staff would never have the time to actually go through or have the data on, or that customers would never ask [of] an external person, but they are very happy to engage with the agent to explore that more,” he said.

“I think it's really important that the customer has … their own little financial companion in their pocket to ask questions that otherwise you probably wouldn't ask. 

“Getting coaching on a savings goal is so important and you would not call the bank every day and ask them for guidance on how you can achieve that. But with a companion or with a conversational AI that you can speak to or type to, I think it's a much lower hurdle to actually get that advice.”

In addition to being a sounding board for customers, however, there is also some anticipation that Companion will be exposed to questions or information that would ordinarily not be posed to a call centre or branch employee.

Baumann said that “purpose-built guardrails” had been built for Companion to detect questions that could relate to sensitive topics such as problem gambling, financial abuse or hardship.

The guardrails are intended to offer customers breakout points from Companion to specialist human-run teams.

“We purpose-built guardrails to actually make sure that we can pick these things up,” Baumann said.

“When it comes to gambling problems, we would probably then try to actually connect [the customer] with our sensitive matters team to ensure that we actually take it forward. 

“It’s really important to actually take these signals from customers and make sure that we are actually reacting to them in the right way.”

Financial advice

Also likely to come up in broader rollout and use of Companion are tough questions that effectively come under the banner of financial advice - something the AI is restricted from providing.

Baumann said that being able to dispense financial advice through something like Companion would likely require legislative and regulatory intervention.

He noted that other AI models that customers could engage already returned responses that were borderline in terms of whether they could be considered financial advice or not.

“It’s clear that if you go into one of the other large language models right now, you are getting very, very close to financial advice [in what you get back],” he said.

“I think that we have to really consider whether it would be a good idea for a regulated ADI [authorised deposit-taking institution] to be able [to provide financial advice through an AI assistant], as long as we can demonstrate that it's actually best for the customer, that we can actually give them an answer that gives them some guidance.

“We have a lot of data about the customer, we have a lot of history and context. So I think we're actually well placed to be able to give them an answer or some guidance, but obviously we have to stay in our regulated methods, and so we will do what is allowed.”

For now, in the context of Companion, that will mean that “sometimes [we] can't give [customers] an answer because they are asking clearly a question around financial advice.”

“If there's one thing that we would love to do for our customers more is obviously exploring how we can actually close that gap,” he said.

Companion’s architecture

Scant detail of the technical basis for Companion was discussed, aside from it involving the work of multiple partners, of which AWS is one, and input from across CBA.

One thing Baumann did say about Companion’s architecture is that it is multi-agent, and that different teams and functional domains from across CBA were building primary or sub-agents that reflected their domain expertise.

“We have a federated model where we work centrally with those areas to actually build the key capabilities to be built into the common Companion,” Baumann said.

“We do not want to have one little area that is actually building something. We want to work across the different domains to bring this to life. 

“But we still [have] a centralised team to ensure that before we launch any sub-agent in the Companion, obviously it's well-tested, aligns to the design criteria that we have, the quality standards that we have, etc.”

Ry Crozier attended the AWS Financial Services Symposium as a guest of AWS.