Tech roundup: How Wealthsimple’s AI chatbot works

This wasn’t the first time the company’s chatbot changed service providers; Wealthsimple initially built it using open-source software and maintained the chatbot itself.

The chatbot was “born out of necessity,” Talasila said. Wealthsimple experienced explosive growth during the latter half of 2020 and needed a more efficient way of handling the higher number of customer inquiries.

But their tech developers were needed in core business areas when interest rates shot up in 2022 and Canadians were looking for high-interest savings accounts, so they outsourced it, Talasila said.

“It’s a very typical build versus buy decision,” he said. “It was really important for us to take the data scientists and the devs that were working on the chatbot and figure out if we can reposition them into [the] business itself.”

The chatbot was initially trained with Wealthsimple’s public help centre documents, Talasila said. Since the files were written for humans, employees added clarifications to teach the chatbot what to do in certain situations.

To mitigate the risk of hallucinations, the AI can only draw information from those documents. “It has access to highly curated information,” Talasila said. “The chatbot is grounded in the truth.”

While the chatbot’s first generation required coding for every change, the AI now understands instructions in natural language; Wealthsimple feeds its standard operating procedures for humans into the chatbot to create a consistent customer experience. Managers can also tweak the instructions to the chatbot without needing to code.

If the chatbot doesn’t know the answer to something or takes too many steps to find out, it tells the customer it doesn’t know and connects them to a human. Customers can also ask to speak to a live agent at any time, Talasila said.

There are, however, limitations in what the chatbot can do. For example, it can’t read a customer’s account information for security reasons. Doing so would require a new AI architecture that preserves client privacy while providing personal information.

“LLMs are super powerful, the riskiest thing we could do is… copy and paste everything about [a customer] into the [chatbot] then have at it,” Talasila said.

Wealthsimple is exploring using AI to respond to customer emails, but Talasila isn’t comfortable with the technology’s quality yet. Many email inquires require access to customers’ personal information to resolve, and need to be processed by a human.

In addition, Talasila’s team is analyzing what sorts of questions the AI can’t answer and expanding the chatbot’s knowledge. The chatbot will be able to respond to a lot more queries in two months’ time, he said.

Notetaking software for advisors adds French language transcription

San Francisco-based notetaking AI software Focal began supporting French language transcriptions this month, John Connell, co-founder and CEO of Focal, said in an email.

Focal made the announcement one day after this reporter published a report on notetaking AI apps for Canadian advisors.

Other financial advisor notetaking software like Salt Lake City-based Jump and San Francisco-based Zocks already support French transcriptions.

While Focal can take audio input in French and transcribe it in French, the meeting summary and other outputs, such as draft emails, will still be in English, Connell said.