Morgan Stanley creates bot that does junior analysts’ work — faster

By Penny Crosman, American Banker
16 February 2021

Technologists at Morgan Stanley have developed a virtual assistant that helps people throughout the organization plumb useful information from the 50,000 research reports the investment bank generates every year.

“Our research reports can be many, many pages long,” said Eden Kidner, global head of research technology at Morgan Stanley. “And now we’re getting to the ability to actually find specific charts and paragraphs within the reports that answer questions.”

The bot helps employees find information buried in proprietary research, making it easier for them write reports and serve clients. The technology incorporates a search engine, text messaging, natural language processing and machine learning.

It’s the kind of useful implementation of artificial intelligence that could be applied at any bank or organization that has troves of unstructured data. In fact, Morgan Stanley already uses AI to crawl through client information to provide financial advisers with “next best action” suggestions to make to clients.

“Software that helps knowledge workers be more productive, as opposed to supplanting them, is of value, in this case sifting through many pages of research reports to provide relevant, context-sensitive information that otherwise would take significantly more time to retrieve and synthesize,” said Steve Rubinow, director of the Institute for Professional Development at DePaul University in Chicago and former chief information officer of NYSE Euronext.

The AskResearch bot is an example of banks ramping up the use of AI in different parts of their business.

“What we are seeing on the AI side, especially in natural language processing, is really amazing,” said Brad Bailey, research director for capital markets at Celent. “One aspect is the ability to get content from all types of structured and unstructured data and leverage that in numerous ways is a competitive edge and a huge benefit to client service.”

When the team at Morgan Stanley created the virtual assistant, it had two goals: to help people in research and sales be more efficient and to make the information in the company’s research reports more discoverable to internal and external clients. Eventually, they want to make the bank’s research omnipresent — in other words, have it follow people around as they’re doing other things.

While someone is writing an email to a client in Microsoft Outlook, for instance, the research bot would be there “kind of looking over your shoulders and available to work with you to answer the questions that you may have,” Kidner said.

A virtual junior analyst

Morgan Stanley started out using a platform from the messaging technology company Symphony to create this research digital assistant.

The team fed the questions employees were asking in the search bar of the bank’s internal research portal into AskResearch. They also fed in the answers that seemed to resonate with users, according to Amit Singh, vice president of institutional securities technology and the main architect for AskResearch.

The bot uses machine learning and natural language processing to become better at answering questions over time, Kidner said.

The chatbot acts kind of like a junior analyst, said Jennifer Casas, executive director of research technology management, who leads client adoption of AskResearch.

“The chatbot is there to answer some of the more important but basic questions that your team or your clients have,” Casas said.

The bot can answer simple questions like, “Who covers Apple?” It will produce the name and contact information of the analyst, a biography and a list of all the companies that analyst covers and the reports the person has written.

The bot can retrieve earnings-per-share estimates for any company, or any of 50 other fundamental metrics the bank tracks or forecasts. It understands abbreviations like GDP (gross domestic product), so they don’t have to be spelled out, and synonyms. It understands multiple ways of asking questions.

The answers can come in the form of paragraphs and semistructured data from within research reports.

But while the bot can do some of the tasks of a junior analyst, Kidner said it is not replacing junior analysts.

“It is very important to our organization that we have junior analysts and that they move through the organization,” he said. “The way I think about the AskResearch bot, it’s kind of like a junior analyst, but it’s helping all of our analysts, junior and senior, be more effective in their roles and be more efficient in providing the information that one can get from different places together quickly in one place. It is not replacing anybody. It’s making them more effective at the job that they do.”

Casas was a junior analyst when she first started working at Morgan Stanley.

“Every day, with the amount of information available and the pressures of the market, there are more and more asks of our analysts, junior to senior,” she noted. “So if you free up the time a junior analyst will have to spend sending a model or looking for a report, then that junior analyst can dedicate that time to something else, something new, something that can enhance their platform.”

Challenge: Extracting useful nuggets

The hardest part of the project has been figuring out how to dig the right nuggets of information out of reports that are dozens of pages long, Kidner said.

“Trying to enable the bot to be able to find a specific paragraph, a chart or something that sits deep within that research and distill it up — that’s where the high value of this bot is,” Kidner said. “It’s also the biggest challenge to solve for.”

The data within the documents has to be tagged so the search engine can find it.

In the first phase, that has meant requiring analysts to tag each element of their reports as they write them. The team would like to automate that content tagging to relieve the analysts of that burden. It’s evaluating tools that could help.

The AskResearch bot is being used across all businesses at Morgan Stanley and is available to all employees.

“We have deep usage from within our sales, trading and research groups, and we’re beginning to see people use it on our fixed-income side,” Kidner said. People are mostly using it to help answer the questions they get from clients. In the past, it typically took 10 minutes to go to the research portal and find a piece of research, he said. With the bot, the task takes less than a minute.

The bot is answering more than 5,000 questions a month from employees. In the beginning, the most commonly asked question was, “What does [Morgan Stanley CEO] James Gorman eat for lunch?”

One question the team has grappled with, Casas said, is deciding whether to train the bot to answer irrelevant questions like the Gorman-lunch one or, “What’s the weather in San Jose?”

It’s possible to go to Google and retrieve such information, but so far the team has found answering such questions does not add value, Casas said.

In the second quarter, Morgan Stanley will begin rolling out the AskResearch bot to clients, so they can use it directly.

This will help clients find content they need, do their work more efficiently, which will help make their relationship with Morgan Stanley stickier, Kidner said.

Eventually, AskResearch may have voice activation, like Alexa or Siri. It may be embedded in other applications.

“We do see a future where we see AskResearch integrated into a broader interaction, a lifetime chat between our clients and our analysts, introducing AskResearch in other channels to be more effective and efficient,” Kidner said.