Citi and Bank of America AI: What Citi Stylus and Erica Change for Juniors

Citi and Bank of America have put AI in front of nearly every employee, by opposite routes. Citi wraps outside frontier models in tools with its own names, Citi Assist and Citi Stylus. Bank of America builds on Erica, the assistant it made in-house for clients in 2018 and has reused ever since.

For a junior banker, analyst, operations hire or wealth associate at either bank, the effect is the same: the first draft of the research summary, the client briefing and the pitch page now comes from a machine, and the job is to check it.

Citi Assist is Citi's internal assistant for policy and procedure questions, and Citi Stylus, which has since grown into Stylus Workspaces, is its tool for summarizing, translating and comparing documents; Citi launched both in December 2024. Erica is Bank of America's virtual assistant, launched in its mobile app in June 2018 and extended to staff as Erica for Employees in 2020.

Citi and Bank of America AI at a glance

The two banks started at opposite ends, Citi with its own staff and Bank of America with its retail customers, and much of the rest follows from that.

What Citi's and Bank of America's AI changes for a junior

Both banks describe the same shift from the desk's side: the machine prepares, and a person decides what reaches the client. Bank of America has been the more specific about what that means for a banker.

The first draft comes from a machine

At its November 2025 investor day, Bank of America's corporate and investment bank showed what its bankers now work with. A "GenAI Suite" pairs BofA Orchestra AI, a proprietary chatbot connected to the bank's research and market commentary, with Microsoft Copilot, which drew more than 250,000 prompts in its first month. A "Banker Efficiency Suite" adds a Company Overview Bot, a Working Group List Bot and Pitchbook Page Generators, which build automated pages inside the bank's own Pitchbook PowerPoint ribbon.

Those are pages a first-year analyst used to assemble by hand. On the bank's July 2026 earnings call, Brian Moynihan put the change in one line: "Our bankers automate the research and presentation materials."

At Citi, the everyday version is Stylus Workspaces, which staff use to summarize research, draft reports, manage projects and write emails. The bank's next step reaches further. Introducing its Arc agent platform in April 2026, chief operating officer Anand Selvakesari and technology head Tim Ryan described a banker in Wealth who "might spend many hours preparing for a client meeting, gathering portfolio data, analyzing market trends and modeling scenarios."

In the future, they wrote, a team of AI agents will be able to do that work, and "the role of the banker, therefore, evolves more decisively from coordinator to architect and advisor."

Preparing a client meeting by hand

• Pull portfolio data and recent activity from several systems
• Read research and market commentary to find the angle
• Build the briefing and pitch pages from templates
• Type up the notes and follow-up tasks afterwards

Preparing it with the banks' AI tools

• A meeting prep tool assembles a client dossier in seconds
• A research chatbot surfaces the relevant commentary
• Bots and page generators draft the pages
• An AI notetaker drafts the summary and the tasks

What it looks like by seat

The shift reaches every kind of junior seat, not only the banker's. Each tool below appears once, with the scale its bank has given for it:

What stays with the person

Neither bank describes the machine as the one that decides. Bank of America's wealth business says AI "cannot replace the valuable judgment, empathy, understanding and personal connection advisors bring to clients." Citi's head of Services, Shahmir Khaliq, made the same case with Citi's own payment volumes when investors asked him about risk in May 2026: even a model that is "99.9% effective" gets some payments wrong, which is why Citi keeps a person in the loop.

What are Citi Assist and Citi Stylus?

Citi's machine drafts come from two tools it launched together in December 2024, and it has widened both since.

Citi Assist, the policy assistant

Citi Assist started as a standalone application. By April 2025, as Citi's tools spread to 11 countries, the bank was building it into the chat apps its people already used, so the answer to a policy question sits where the question gets asked.

From Citi Stylus to Stylus Workspaces

Stylus has grown in three steps since its launch:

• April 2025: a new version for about 10,000 early adopters added follow-up questions, the ability to edit and store what the model produced, and a browser plug-in that summarizes web pages outside paywalls.
• September 2025: an agentic version, by then called Stylus Workspaces, linked to internal systems such as the global employee directory and common enterprise platforms, and to web search. Citi's example was one prompt that finds the top five branded-card businesses in the U.S., distills their strategic goals and translates the findings into Spanish. It opened as a pilot for 5,000 colleagues.
• February 2026: Spaces, a shared space for teams of up to 15 people, went to a 250-person pilot, with a rollout planned for all 182,000 employees with Stylus Workspaces access.

Chief technology officer David Griffiths described the gap Spaces closed: "Prior to Spaces, colleagues were running individual conversations within Stylus Workspaces and then sharing out the results or findings to their team outside the tool."

Which AI models does Citi use?

Citi's tools sit on Google's stack first. Under a strategic agreement signed in October 2024, Citi uses Google Cloud's Vertex AI platform "to deliver generative AI capabilities across the company," and Vertex AI is where it built Citi Assist and its coding tool, Citi Squad, with access to Google's Gemini models. A Citi spokesperson has said Stylus Workspaces runs on both Gemini and Anthropic's Claude.

The mix is deliberate: Griffiths has said he wanted options across several AI providers, with a strong preference for closed-source models.

The Google relationship now reaches clients too. Citi Sky, unveiled in April 2026 and built with Gemini and Google DeepMind's avatar technology, is an AI member of the Citi Wealth team for Citigold clients, due to reach U.S. clients from summer 2026. Andy Sieg, head of wealth, told investors it would "make our advisors and relationship managers vastly more productive."

Next: AI agents on Arc

The next phase is agents. In April 2026 Citi introduced Arc, a platform for building and scaling them, starting with its own developers on "specific, well-defined use cases" before access widens across the firm. The wealth banker's meeting preparation, above, is the example Citi chose to show what agents will be for, and a governance promise came with it: "Every agent will be monitored, auditable and governed."

What is Erica, and how did Bank of America build it?

Bank of America started six years before Citi, and from the customer's side. It launched Erica inside its mobile app in June 2018 to give customers an easier way around the bank's applications. Hari Gopalkrishnan, who worked on it as chief information officer of the consumer bank and became chief technology and information officer in 2025, says it was never a single model.

Erica's build started with open-source language models such as OpenNLP, on a platform designed to swap them out. "We always knew we were going to replace models as we go along," Gopalkrishnan said. "So it wasn't that Erica is a model. Erica is a platform that can actually leverage multiple models."

Build once, reuse

That design became the bank's AI strategy, which it calls "build once, reuse." Erica for Employees and ask MERRILL followed in 2020, then CashPro Chat for business clients, Capital Markets Insights and Ask Private Banking, all on the same architecture.

When large language models arrived, the bank tested the alternative. An AI council ran about 15 proofs of concept that deliberately skipped the Erica platform. "As the results came back, it was clear we were reinventing the wheel a lot," Gopalkrishnan said. In 2025 Bank of America rebuilt the Erica architecture as a second-generation platform for the new models.

The bank was also more careful with customers than with staff. It chose traditional AI for Erica's client answers, reading what a customer wants and sending them to the right place, and in April 2025 it still kept generative AI out of those answers. Gopalkrishnan has been candid about what generative output adds: "now you've got to worry about hallucinations."

How fast Erica has grown

Erica's usage has compounded. It took just over four years to reach one billion client interactions and 18 months to reach the second; the third was announced in August 2025, when Erica was averaging 58 million interactions a month.

First billion, June 2018 to October 2022

52

Second billion, to April 2024

18

Third billion, to August 2025

16

The first bar runs from Erica's launch to the October 2022 milestone; the others run between Bank of America's own milestone releases.

By March 2026 the running total stood at 3.2 billion, and in 2025 alone 20.6 million clients used Erica nearly 700 million times. Those are client conversations, but the same platform carries the tools Bank of America's own staff use every day.

What Bank of America's staff use every day

Beyond the desk-specific tools in the table above, three reach staff across the bank:

• Erica for Employees, launched in 2020 and widened in 2023, handles IT and HR help, from password resets to payroll and tax forms. More than 90% of the bank's 213,000 employees use it, and it has cut calls to the IT service desk by more than half.
• The Academy, the bank's training organization, uses AI conversation simulators so staff can rehearse client conversations with real-time feedback; employees completed more than 1 million simulations in 2024.
• Coding assistants serve 18,000 developers, with a 20% productivity lift in parts of the development cycle. They come mostly from a single vendor that Bank of America has not named.

How many employees use AI at Citi and Bank of America?

Both banks now count their AI users in the hundreds of thousands, but they count different things, and each has changed its yardstick as the tools spread.

Citi

Read in order, Citi's own figures show its tools going from launch to nearly the whole firm in about 19 months:

Fraser points to engineering in particular: AI-driven code reviews passed one million in 2025, and she credits them with about 100,000 hours of developer capacity a week, a figure she repeated at the May 2026 investor day. Citi's Services business reports a 30% to 40% boost in developer productivity.

Bank of America

Bank of America's count has moved from one assistant to the whole toolkit. By November 2025 more than 130,000 employees were authorized to use its enterprise productivity tools, with access for everyone promised by the end of that year. By July 2026 Moynihan was reporting the wider picture:

200,000+

Teammates actively using AI tools

July 2026 earnings call

400,000+

Prompts a day

July 2026 earnings call

300+

AI use cases approved

July 2026 earnings call

114

Live generative AI use cases

July 2026 earnings call

Of those 114 live generative use cases, 34 were fully implemented, and Moynihan said new capabilities were arriving every week.

What have Citi and Bank of America executives said about AI?

Read together, the two banks' executives describe the two routes in their own words: Citi's about speed and reach, Bank of America's about foundations and control.

The jobs side of both stories is told in the analysis of whether AI will replace investment banking analysts: Fraser's January 2026 memo, Citi's mandatory prompt training for 175,000 staff and its restructuring of about 20,000 roles, and Moynihan's 200,000 applications for 2,000 places, his advice to "harness it," and Bank of America's intern and graduate classes held at 2,000 each.

How do Citi's and Bank of America's approaches differ?

Set side by side, the two banks answer the same design questions differently, and JPMorgan, the bank most candidates already link with in-house AI, gives a third answer.

Citi's route is the one most people would guess a bank would take: license strong outside models, wrap them in tools the bank controls, and put them in everyone's hands early, so the gains grow as the tools improve. Bank of America's is older and more patient: build the plumbing once, in-house, and put most new tools on it, even when that is slower. Gopalkrishnan's word for the discipline is "institutional patience."

JPMorgan sits closer to Citi. Its LLM Suite is a bank-built front door onto outside models, designed from the start to swap providers, and its story is told in full in JPMorgan's LLM Suite.

What should a candidate know about AI at Citi and Bank of America?

None of these employee tools is open to outsiders, so what a candidate can bring is understanding rather than experience. Four points are worth having straight:

1. Which route each bank took. Citi wraps outside models in its own tools; Bank of America builds on a platform it owns. Being able to explain the difference, and each bank's reason for its choice, says more than knowing a tool's name.
2. Three similar names, three different tools. Citi Assist, CitiService Agent Assist and EricaAssist are separate products at two banks, and mixing them up is an easy, checkable slip.
3. Governance is part of the design. Citi's agents come with the promise quoted above; Bank of America describes its approach as "human oversight, transparency, and accountability for all outcomes." A junior works inside those controls, not around them.
4. Prompting is taught, not assumed. Bank of America's academy runs AI training from basic prompt engineering to AI design and development, and the bank has filled 44% of its jobs in recent years through internal moves, partly on the back of that kind of upskilling.

How to prepare

1. Take a section of a public company's annual report, have any AI tool summarize it, then check every figure against the source. That is the draft-and-check loop both banks' tools create.
2. Build one client briefing or comparable-company page yourself from filings, so you know what a correct one looks like before a generator hands you one.
3. Read the banks in their own words: Citi's May 2026 investor day transcripts and Bank of America's newsroom releases are public and dated.
4. Rehearse describing a time you caught an AI tool's mistake with the interview readiness tool, and use the wider library of guides for the other banks.

The bottom line

Citi licensed the models and put its own names on the tools; Bank of America laid its own foundation first and has kept building on it. Which choice ages better is a question for the banks' technology chiefs.

For a candidate, the question the opening raised is already settled. At either bank, a first-year's work now begins with a machine's version of the page, and the job is to find what it got wrong before a client does.