Test what you know about AI before a finance interviewer does

Answer: It was a research model that improved Terminal search, not a chat product

Bloomberg announced BloombergGPT in 2023 as a large language model built for finance, and its own machine learning team described it as working behind the scenes to improve search on the Terminal. Nobody opens it by that name. The conversational assistant Terminal users actually type into is a separate product, ASKB. Getting that distinction right tells an interviewer you have used the Terminal rather than read about it.

Answer: AlphaSense is an AI research search platform; AlphaSights is an expert network

AlphaSense is software: it searches filings, earnings call transcripts and broker research, and uses AI to pull out themes and shifts in tone. AlphaSights is a separate company, an expert network that arranges paid calls between investors and people who know an industry. The names are close and the businesses are not, and a junior on an investing team can easily use both in the same week. Mixing them up in an interview suggests you have used neither.

Answer: Microsoft sells several Copilots, from Excel to coding, and the line names none

Copilot is a family name at Microsoft rather than one product: there is Copilot inside Excel and the rest of Microsoft 365, and there is GitHub Copilot, a coding assistant for developers. A bare “Copilot” leaves the reader guessing which one and what you did with it. Name the tool and the task, such as a first-pass variance analysis built with Copilot in Excel and then checked by hand, and the line starts doing work. Banks such as Barclays have rolled Microsoft 365 Copilot out to most of their staff, so the tool alone sets nobody apart.

Answer: Built-in connections to licensed financial data such as FactSet and PitchBook

The finance editions are largely the same models, wired into the data an analyst already pays for, such as filings, market data, private company databases and research, plus templates for common finance tasks. That is why their launches name data partners rather than a new model. Nothing about them removes the need to check the output, which stays with the analyst. A model stripped of general knowledge would be worse at the reading and writing that makes these tools useful in the first place.

Answer: The software vendor or consultancy whose product the bank is putting in

The title comes from Palantir, which began embedding its engineers inside customers' organisations to make its software work on their data. In finance it is still mostly a vendor role: Palantir, OpenAI, Anthropic and the consultancies that implement their products place engineers inside banks and funds. The banks and trading firms do similar work in-house under titles such as AI engineer or machine learning engineer. Knowing that tells you whose interview you are really preparing for.

Answer: Turning models into fast, reliable systems that run on the firm's data

The researcher invents and tests methods; the engineer makes them run reliably, at scale, on the firm's data. Jane Street, for one, splits its machine learning hiring three ways: researchers who invent the models, research engineers who bring them to life, and performance engineers who make training faster. Which track a posting sits in tells you what the interview will weigh, so read that before you prepare.

Answer: It builds the machine learning tools and data pipelines the quants use

At large firms such as BlackRock, Fidelity and Schroders, data science builds for the investment teams rather than running money itself. The quants own the signals and how the portfolio is put together; the data scientists build the machine learning tools, document pipelines and data infrastructure those teams draw on. Fidelity describes its data scientists working with quants and portfolio managers to prototype and deliver those tools. Pitching yourself as a signal researcher for a tooling seat is the quickest way to show you have misread the job.

Answer: Model risk, which validates it independently of the people who built it

Banks run model risk management as a separate function whose job is to challenge a model before it goes live and keep testing it afterwards. US supervisors set that expectation out in guidance in 2011, and a machine learning model is still a model under it. It is why AI governance and model risk teams sit alongside the builders at banks such as JPMorgan. A candidate who can explain why builders cannot mark their own homework is ready for either side of that table.

Answer: Inside the bank's own approved AI tool, under the bank's data controls

Client information does not go into a public consumer app. The large banks license models from OpenAI, Anthropic, Google and others and run them inside their own tools and controls, which is what JPMorgan's LLM Suite and Goldman Sachs's GS AI Assistant are for. Deleting a chat afterwards does not undo having sent the data out. Saying so unprompted in an interview shows you understand why those in-house tools exist.

Answer: A named task it helped with, a result, and the finance skill beside it

A bare tool name reads as padding. What works is a line that names the task, the result and the finance work around it, such as using Python and an AI assistant to automate a monthly variance report and cutting the time it took. Keep the modelling and valuation first: the job is an analyst who uses AI, not an AI user who might pick up finance.

Answer: Every figure in it, checked back to its source before anyone sees it

A draft from a tool becomes your work the moment you pass it on. Data feeds carry errors, peer sets get chosen badly, and a language model can produce a confident number that matches nothing in the filings. The analyst's value is moving from building the first version to catching what is wrong with it, and that habit is what separates a useful junior from a merely fast one.

Answer: Full-time jobs worldwide exposed to some degree of automation by AI

The 2023 research note estimated that generative AI could expose the equivalent of 300 million full-time jobs worldwide to some automation, and its headline was a lift to global growth. Exposure to automation is not the same as a job disappearing, and the note was about the world economy, not Goldman's own staff. Quoted as “AI will destroy 300 million jobs”, it became one of the most repeated misreadings in the debate. Getting it right marks you as someone who reads past the headline.

Answer: An in-house platform giving employees secure access to AI models

LLM Suite is JPMorgan's own generative AI platform for employees, released across the firm in 2024 inside the bank's secure environment. Staff use it to draft, summarise, analyse documents and write code. The stock-basket tool is a different JPMorgan product, IndexGPT, and a customer chatbot is the kind of assistant retail banks put in their apps. When a bank interviewer asks which AI tools you would use, the honest answer is usually the bank's own.

Answer: As one interface over several licensed models, run under its controls

Goldman's assistant routes work to models from several providers, including OpenAI, Google and Anthropic, and runs them behind the firm's own firewall. David Solomon described it on a 2025 earnings call as giving staff safe, secure access to firm-approved external large language models. One front end over many models lets a bank switch providers as the models improve. It is the same logic that led JPMorgan to build LLM Suite rather than hand its staff a public app.

Answer: The firm's own library of research and internal documents

When Morgan Stanley announced the assistant with OpenAI in 2023, it said answers would come only from Morgan Stanley's own content, with controls around it. At launch that meant a library of about 100,000 research reports and internal documents. The design matters more than the model: an assistant that answers from vetted firm content can be trusted with a client question in a way the open internet cannot. The firm later added Debrief, which writes up advisers' client meetings.

Answer: Bank of America's Erica, the assistant inside its banking app

Erica launched in 2018 as a virtual assistant in Bank of America's consumer app and has handled billions of customer interactions since. The other three are staff tools: a platform for employees, a firmwide assistant and a note-taker for advisers. Banks run the two kinds side by side, one facing customers and one facing their own people, and a new analyst lives in the second. Bank of America has an internal version too, Erica for Employees.

Answer: EQT's whole investment process, from sourcing to portfolio value creation

Motherbrain started in 2016 as EQT Ventures' system for finding companies before anyone else did. EQT has since made it a group-wide team and describes it as supporting the entire investment lifecycle, from identifying opportunities to creating value in the companies it owns. Calling it a startup-screening tool is a common and dated description. For a candidate the point is that sourcing is only the start: the same data work follows a deal through ownership.

Answer: The best price any company in the portfolio paid for the same product

Apollo's system compared purchasing contracts and invoices across more than 40 of its portfolio companies to find the best price paid for a given product. In one case it read thousands of software agreements in minutes and helped a company cut a procurement cost sharply. That is value creation at portfolio scale, something a single company cannot do alone. Procurement and AI deployment are both named levers for Apollo's portfolio performance team, APPS.

Answer: An in-house team building AI agent products with its software companies

Vista launched the Agentic AI Factory in 2025 as an in-house team of AI engineers and specialists that designs, builds and sells AI agent products alongside the enterprise software companies it owns. It sits within Vista's value creation work, the operating side of the firm. At a software-focused owner, AI becomes part of each company's investment case rather than just a tool the deal team uses, and that changes what an associate is asked to judge.

Answer: Every document can get a first pass, where teams once had to sample

Language models can read every contract and filing in a data room in the time a team used to spend sampling a few, so the first pass stops being the bottleneck. The judgement does not move: someone still decides which red flag matters, checks each figure and writes the case for the committee. TPG, for one, uses an internal AI assistant to turn documents into committee-ready notes while the decision stays with people. Advisers and the memo both survive; what changes is where a junior's hours go.