ChatGPT and Claude in Banking: How Analysts Use Them and What Stays Out

ChatGPT is OpenAI's AI assistant and Claude is Anthropic's. At a bank, staff mostly reach the same models inside a tool the bank controls, or through an enterprise licence it has approved, rather than through the public apps; several of the biggest banks restricted the public ChatGPT in early 2023.

What changes for a junior is the first draft. Research summaries, comps, model updates and pitchbook pages increasingly start as machine output. The analyst's job moves to checking that output, and to knowing what must never go into a prompt in the first place: client names, deal terms and anything that counts as material non-public information.

What do ChatGPT and Claude change for a junior banker?

The first draft moves to the model

OpenAI's own launch showed the shift. When it unveiled its banking product in September 2026, Nick Turley, its vice president of product, had the product analyse a potential M&A target, pull figures from industry data sources and build a formatted PowerPoint deck in a bank's house style. Then he explained why that was harder than it looked.

"It's very easy to make slides that look good, but it's much harder to make slides [that] actually make sense. To get here, ChatGPT had to choose the relevant peers. It had to pull the prices into a spreadsheet. It had to check the chart against the data, and it had to explain the sell-off and the rebound."

Those four steps are an analyst's afternoon, compressed into one request. Read them again as a checklist. Each is a place where a model can be confidently wrong, and each is now something a junior verifies rather than builds.

What the junior still owns

What stays with people is judgment at both ends of the task. Before the prompt, the junior has to know which approved tool the work belongs in and what information may go into it. After it, the junior checks the peer set against the one a senior banker would defend, traces each figure to its source, and makes sure the story on the slide is one the numbers support.

The building still has to be understood: an analyst who has never put a comps table together by hand cannot tell when a model has put together a bad one.

How banks give staff ChatGPT and Claude

Mostly through a door the bank built or bought. There are two main ways in, and neither is the app a student uses at home.

Inside the bank's own assistant

Several of the largest banks put models from outside labs, OpenAI and Anthropic among them, behind an interface they control. The bank decides which of its own data the model can draw on, which model answers, and what gets watched.

In every row the name on the screen is the bank's, and the model behind it comes from an outside lab.

A licensed enterprise seat

The second way in is a licence. OpenAI launched ChatGPT Enterprise, its business version, on August 28, 2023, and for a bank that does not want to build its own interface, the licence is the product.

BBVA shows how far that route can go. The Spanish bank started with 3,300 ChatGPT accounts in May 2024 and widened the rollout to 11,000 staff, who reported saving an average of three hours a week on routine tasks. In December 2025 it said it would extend ChatGPT Enterprise to its entire workforce of more than 120,000, calling it one of the largest corporate deployments of the technology in the world.

Claude also reaches staff through software they already use. Some of Microsoft's Copilot tools offer Claude models, and the Copilot in Excel page covers how that works inside a spreadsheet.

How do analysts use ChatGPT and Claude at work?

Inside those tools, the work looks less like chatting and more like handing over pieces of a deliverable. Both labs now sell finance editions of their assistants, built around the same short list of jobs that fill a first-year analyst's evenings, and the firms that train new analysts have built AI into how they teach those jobs.

Wall Street Prep and Financial Edge, which say they train analysts at nine of the top ten global investment banks, describe the sequence in their 2026 new-hire programmes: analysts "build the skills manually first, then apply AI to the same problem, and then compare outputs." What the tools produce, and what the junior then checks, looks like this.

One morning, illustrated

Put together, a morning might run like this. It is an illustration, not any one bank's workflow.

1. An associate asks for two pages on a possible acquisition target before an afternoon call.
2. The analyst asks the bank's assistant for a summary of the target's last two earnings calls, with citations to the transcripts.
3. The same tool builds a first cut of trading comps from the market data the bank already licenses.
4. It drops the results into the firm's template as a company overview page and a valuation page.
5. The analyst opens each cited passage, questions two of the peers, re-checks the multiples against the source data, and rewrites the summary so it says only what the numbers support.

Steps two to four can take minutes. Step five takes as long as it takes, and it is the step the associate will judge.

What do ChatGPT and Claude get wrong?

Step five exists because these models fail fluently. Three kinds of failure recur:

• Invention. A figure, a quote or a date that is not in the source, stated as confidently as one that is.
• Refusal. No answer to a question a person could answer from the page in front of them.
• Inconsistency. A different answer to the same question asked twice, because the models are not guaranteed to produce the same output from the same input.

What a filing test found

In November 2023 a startup called Patronus AI released FinanceBench, 10,000 questions and answers built from SEC filings, earnings reports and call transcripts, and ran a 150-question sample through the leading models. How much of the filing a model was given decided almost everything.

No filing, from memory

9% (14 of 150)

With a retrieval system

19%

Nearly the whole filing

79%

Pointed to the exact passage

85%

Patronus AI, November 2023. The retrieval bar is 100% minus the 81% failure rate Patronus reported for that setup. Models have improved since; the test shows what the source document is worth.

Even the best setting left gaps. Pointed to the exact passage, GPT-4 Turbo still gave a wrong answer 15% of the time. With nearly the whole filing in the prompt it was wrong 17% of the time, and Anthropic's Claude 2 was wrong 21% of the time.

Rebecca Qian, Patronus AI's co-founder and chief technology officer, set the bar higher still: "especially in regulated industries, even if the model gets the answer wrong 1 out of 20 times, that's still not high enough accuracy."

Newer models, the same habit

The models have improved since 2023. Vals AI, which runs a finance benchmark of 537 questions built with Stanford researchers, a global systemically important bank and industry experts, says the test covers "tasks expected of an entry-level financial analyst" and reports significant improvement on it. Anthropic said its Claude Sonnet 4.5 led that benchmark at 55.3% accuracy in October 2025, and that Claude Opus 4.7 led it at 64.37% in May 2026.

Even the higher score leaves more than a third of those entry-level analyst questions answered wrongly or not at all. Leaderboards move with every model release, so a score is only worth repeating with the name of whoever published it and the date.

One exhibit, worked through

Error rates sound abstract until they meet a real page. Take a table of six peer companies with five figures each: 30 numbers.

• At the 17% wrong-answer rate GPT-4 Turbo showed with nearly a whole filing to read, about five would be wrong (30 × 0.17 = 5.1).
• At the one-in-20 rate Qian called too high for regulated work, one or two would still be wrong (30 × 0.05 = 1.5).

Neither rate describes a tool a bank runs in 2026, and the 17% came from answering questions, not filling tables. The arithmetic is the point: a page that is 95% right still carries an error, and the person who has to find it is the analyst.

What banks allow, restrict or block

The 2023 shut-off

Wrong answers were not what worried banks first. Data was. Within three months of ChatGPT's public release in late 2022, several of the largest banks had restricted staff access, mostly by treating it like any other unapproved outside website.

Bank of America's move came as banks tightened control of staff communications after US regulators levied more than $2 billion in fines over unmonitored messaging apps such as WhatsApp. The cautionary tale of that spring came from outside banking: Samsung reported in early April that employees had accidentally leaked confidential source code and meeting recordings by uploading them to ChatGPT.

What replaced the block

After the block came the two routes above, and with them controls on what staff type. Goldman Sachs's chief information officer, Marco Argenti, described Goldman's to American Banker in July 2025:

"There are controls that are looking at what type of information people put in the prompts, and then we apply filters, and we flag and so on and so forth. But that's normal in a regulated industry. That's what you do for every form of communication."

For a junior, the phrase that matters is the last one. At a bank that works this way, a prompt is watched and filtered like an email or a chat message, not treated as a private scratchpad.

The rules that already applied

For US broker-dealers, the existing rulebook already covered chatbots. FINRA, their regulator, reminded firms in June 2024 that its rules are "intended to be technology neutral" and apply to generative AI "just as they apply when member firms use any other technology or tool." The standards for communications with the public apply "whether member firms' communications are generated by a human or technology tool."

In practice, a model's draft that reaches a client is the bank's communication, reviewed like anything else the analyst sends.

What can you put into ChatGPT or Claude at a bank?

If a prompt is a communication, what goes into it matters, and the first line is drawn by the vendors' own defaults.

OpenAI says that when people use its services for individuals, such as ChatGPT, "we may use your content to train our models" unless the user opts out. Anthropic changed its consumer terms in 2025 so that users of Claude's Free, Pro and Max plans choose whether their chats train its models, with the data kept for five years if they agree.

Business versions sit outside both rules. OpenAI says it does not train on business products such as ChatGPT Enterprise and its API by default, a promise it has made since Enterprise launched: "We do not train on your business data or conversations." Anthropic says of its finance edition: "By default, your data is not used for training our generative models."

Material non-public information, briefly

The harder line is legal. Insider trading, as the SEC's investor education site defines it, is trading "on the basis of material, nonpublic information" in breach of a duty of trust. A junior on a deal team is surrounded by that kind of information: which company is about to be bought, at what price, and when.

Banks keep it behind information barriers between teams, and the vendors now build for that. OpenAI's banking edition says "protecting material non-public information and client confidentiality is critical" and lets firms create multiple workspaces "to enforce information barriers." Training firms treat it as day-one material too: Wall Street Prep and Financial Edge list "MNPI restrictions" and "the judgment required to know when AI is categorically off-limits" among the things new analysts learn.

So what stays out of a prompt is anything a client or a deal team told you, unless the bank has approved a tool for it and its policy allows it at all. Public material is fair to practise on anywhere, including the consumer apps.

Public by nature

• An annual report, a quarterly filing or an earnings transcript
• A press release, or research you are entitled to read
• A blank template or a formula question
• Your own writing with no client detail in it

Firm-only, and only where approved

• A client name, its numbers or its plans
• The terms or timing of a deal not yet announced
• Anything from a data room or a live pitch
• Personal details about a client or a colleague

Where the line blurs, compliance decides, not the analyst.

ChatGPT for Financial Services vs Claude for Financial Services

Both vendors speak to these concerns in the finance editions they now sell: versions of their assistants that put the model, licensed market data and a firm's own templates in one place. At a bank that licenses one, this is where the first draft comes from.

Claude for Financial Services

Anthropic moved first, in July 2025. Kate Jensen, then Anthropic's head of revenue, called the product "a tailored version of Claude for Enterprise" that is "specifically built for financial analysts." Anthropic pitched it for due diligence, competitive benchmarking, "financial modeling with full audit trails" and "institutional-quality investment memos and pitch decks."

The customers quoted at launch were a hedge fund, an insurer and a retail bank. Bridgewater's AI unit said Claude had powered the first versions of its Investment Analyst Assistant, which worked through complex financial analysis "with the precision of a junior analyst," and AIG described building Claude into its underwriting.

ChatGPT for Financial Services

OpenAI followed on September 10, 2026, with a tailored version of its enterprise product, ChatGPT Work, running on its newest model, GPT-6 Astra. OpenAI says two design partners shaped it: "Our early work with Morgan Stanley and Evercore has helped steer where we have started: investment banking and equity research." Reliable access to data and high-quality output, it added, "proved to be the biggest pain points for their teams."

OpenAI's answer to both is built into the product: market data arrives inside it, with "granular citations so that bankers can trace figures and claims back to their sources." It is available to eligible financial institutions. Side by side, as each vendor describes its edition:

How ChatGPT and Claude in banking are changing

The next step is already in the product announcements: from answering a question to doing a sequence of work across the software an analyst already has open. Anthropic's May 2026 description of its pitch agent reads like a first-year's to-do list: "Hand the Pitch agent a target list, and you can get back a comps model in Excel, a pitchbook drafted in PowerPoint, and a cover note ready in Outlook."

Anthropic says its add-ins carry context from one application to the next, so that "an analyst who's started a model in Excel doesn't need to re-explain it when that work moves to PowerPoint." OpenAI is pushing from the data side: it hosts financial data inside its product and has said, "We will post train our models to find, interpret, and use this data like we know the best analysts can."

However far the agents reach, Anthropic says users stay in the loop, reviewing and approving work before it goes to a client. The junior's two jobs survive the upgrade: checking the draft, and deciding what goes into it.

What to say about ChatGPT and Claude in an interview

Banks do not publish interview questions about these tools, so the aim is a credible answer if the subject comes up, not a scripted one. The bar shows in how banks train their own juniors: UBS builds an AI fluency track into its graduate programme, centred on real-world use cases, responsible use and sound judgment. A strong answer covers five things.

1. Name the route correctly. At a bank you would use the firm's approved assistant or licence, not the public app, and you would keep client information out of consumer tools. That shows you understand why the banks blocked it in 2023.
2. Have one real mistake ready. A figure that did not match the filing, a peer that did not belong, a summary that missed the one line that mattered, and how you found it.
3. Be specific about what the tools are for. First drafts, summaries, formula help and layouts, followed by the checks in the table above.
4. Claim only what you can have done. A bank's internal assistant is open only to its staff, so talk about what the bank has said publicly, not hands-on experience you cannot have had.
5. Show you could do the work without the tool. Building a summary or a valuation table by hand is what makes your checking believable.

To rehearse these answers, try the interview readiness tool; the investment banking guide goes deeper on each bank's own system.

The bottom line

The ChatGPT and Claude a candidate knows from a phone are not what a bank hands a new analyst. What arrives is the same models inside an approved tool, wired to licensed data and the firm's templates, watched like other communications, and able to produce a first draft of much of what a first-year used to build by hand.

That changes the job at both ends. After the draft exists there is less building and more checking; before it exists there is a new responsibility for what goes into the prompt at all. A candidate who can explain both ends is describing the job as it now works.