"At first it was a tool, and now it's becoming a partner, to be candid, because of how frequently I use it," he said.

Morgan Stanley partnered with OpenAI, the maker of ChatGPT, to roll out this chatbot in September to its 16,000 financial advisors,most of whom work in small teams. The bank said 98.5% of teams now haveat least one member who uses the product once a week or more. Morgan Stanley plans to expand AI projects firmwide, promoting Jeff McMillan, the head of wealth-management tech, in March to lead the effort.

The chatbot,called AI@MS Assistant or AIMS for short, is decidedly less sexy and fun than ChatGPT. It can reference only the bank's internal content, can't use metaphors or analogies, and answers questions related only to wealth management. When asked if Donald Trump would make a good president, AIMS declined to answer, McMillan demonstrated to Business Insider.

"We spend as much time teaching it the things it cannot answer that we do what it can answer," he said.

This is by design, to prioritize accuracy and prevent "hallucinations," he said. Despite being a self-described "AI guy," McMillan believes AI won't change the world in the next several years but will allow people to free up "crap time" and focus on more important tasks.

"I actually am hoping that AI will make us more human because it's going to allow us to spend more time talking to clients, more time solving complex problems, more time having lunches as opposed to being dragged into the morass of bureaucracy," he said.

AIMS can theoretically answer anything, but it doesn't

Biggs was one of the first advisors to try AIMS as a demo user in the summer of 2022. He recalled that it got easily confused by acronyms and questions that weren't complete sentences.

McMillan acknowledged that its accuracy was "quite low" at first and that some answers were too long or too short.

But these bugs are easier to troubleshoot than with traditional software. Instead of fiddling with code, AIMS can be updated by typing commands in the system, such as giving more details on alternative investments when asked about them. Advisors can rate answers with a thumbs-up or a thumbs-down, and the generative-AI team reviews the incorrect or incomplete answers.

The bank declined to specify the percentage of accurate answers by AIMS, but McMillan described it as more accurate than other approaches, such as searching the company's intranet, asking colleagues, or drawing on employees' own knowledge. Biggs said AIMS replies had gotten quicker since the September launch.

Before AIMS, Morgan Stanley had a product called FAST that could answer about 5,000 questions. AIMS, in theory, can answer infinite questions. Its expertise runs the gamut from market-outlook digests and stock comparisons to estate-tax tactics and compliance procedures. Drawing on 100,000-plus internal documents, AIMS can address highly technical topics, like using a charitable trust for a client about to sell a business.

But AIMS is capable of answering far more questions than it's allowed to, McMillan said. For instance, it declined to specify Morgan Stanley's most recommended investment products when this reporter asked. McMillan said this is to prevent AIMS from making anything resembling a suggestion to clients without knowing their finances.

"You could be critical of the guardrails that we put in. People would say, 'I want to know if I should buy IBM stock,'" he said. "We just have to be very thoughtful about how we do this."

To prioritize accuracy, AIMS doesn't use information outside Morgan Stanley's database. (OpenAI also doesn't train on AIMS data logs, which are stored in a private cloud.)

"Sometimes ChatGPT makes stuff up. We're not having that problem because we're controlling the inputs," McMillan said. "The more information you put in place into this system, and the more disparate that information is, the harder it is to get things right."

What AIMS is (and isn't) for

Advisors frequently ask for AIMS to adopt ChatGPT-like features such as the ability to have extended conversations, McMillan said. While AIMS will be updated over time, it will never write poems in the style of Dr. Seuss as ChatGPT can, as that isn't relevant to the chatbot's purpose: helping advisors serve clients better and faster.

AIMS is meant to be a time-saver. While the bank declined to specify the estimated time savings per advisor, Biggs said it added up to a meaningful amount on a weekly basis. For instance, reading a US economic forecast would take him about five or six minutes, while AIMS can generate a bullet-point summary in under 30 seconds. The answers have references so he can check the sources.

Fewer advisors are contacting their designated call center after using AIMS, indicating the answers are more reliable, McMillan said. Instead of contacting the center, advisors can handle tasks like processing wire transfers in the Cayman Islands by themselves.

This chatbot is just the beginning for AI at Morgan Stanley

McMillan expects to use the underlying tech for AIMS with other parts of the bank, as search tools are relevant firmwide, but he hasn't committed to any specific projects.

The wealth-management head Jed Finn said in February that two AI product releases were in the works. AIMS Debrief, already in pilot mode, summarizes Zoom conversations with clients and generates emails with next steps for the client and the advisor. AIMS Plus helps advisors send emails, conduct research, create newsletters, and more.

Morgan Stanley's ultimate goal is to have a voice-activated bot that can execute complex tasks like drafting proposals, rebalancing portfolios, transferring money, and creating performance reports.

Biggs isn't concerned about his role being replaced with technology.

"When the ATM launched in the '70s, there was a thought of, OK, all of these branches are going to go away because you now have ATMs. There's now more people in branches than ever before," he said. "I think it's going to just free up a lot of time to engage more with clients."