Generative AI in fintech earns its keep reading, not deciding

Generative AI in fintech does its best work as a reader. It summarizes advisor meetings, pulls figures out of loan files, drafts replies to complaints and checks marketing copy against a rulebook. Where it struggles, and where regulators get careful, is the moment someone asks it to decide: approve a loan, quote a payoff amount, place a trade.

That's the position this piece argues, and the evidence lines up behind it. The documented production results at Klarna, Morgan Stanley and Stripe come from narrow jobs done at volume. FINRA says the top generative AI use case among its member firms is summarization and information extraction. And the US rules that govern lending, advice and communications don't care whether a person or a model wrote the words.

Generating text and scoring risk are different jobs

"AI in fintech" covers two families of technology. Predictive machine learning has scored card payments for fraud and applicants for credit risk for years. Generative AI, mainly large language models, creates text, code, summaries or synthetic data. It reads unstructured documents, answers customers in natural language and drafts reports. It complements the predictive layer rather than replacing it. Most confusion about generative AI in financial services starts when the two get blurred.

The rows split along one line. If the output is a number someone will act on, it should come from a model you can test and explain, or from the system of record. If the output is words a person will review, generative AI is a fair candidate. Our guide to AI in digital banking covers the predictive side, including fraud scoring, AML and credit. This piece stays with the generative layer.

What the generative AI in fintech market numbers say

Ask how big this market is and you'll get several answers, sometimes from the same page. Market.us puts the global generative AI in fintech market at $1.9 billion in 2025, growing 31.1% a year. Further down, in its key players section, the figures become $1.95 billion and roughly 24.8%. Both sit in the August 2026 edition of the report.

Technavio's page has its own wobble, giving the forecast period as both 2024–2029 and 2025–2029. Neither public page breaks out a figure for the US generative AI in fintech market, so North America is the closest proxy. For a sector that prices risk for a living, that's a lot of uncertainty in one number.

Treat market sizing as a sign of direction, not a budget input. A business case should come from your own workflows: hours per loan file, chats per agent, days to close a compliance review.

Regulator surveys are more useful, because they describe what firms actually do. The Bank of England and FCA's 2024 survey of 118 UK firms found 75% already using AI, with foundation models behind 17% of use cases. In the US, FINRA's 2026 oversight report (December 2025) names summarization and information extraction as the top generative AI use case among member firms. FINRA says firms have started to implement it with a focus on efficiency gains. So the most common use of generative AI in financial services isn't a robo-advisor or an autonomous trader. It's reading.

Reading at volume is where the results are

Klarna is the most quoted example, and its numbers come from the company. In its February 27, 2024 announcement, Klarna said its OpenAI-powered assistant held 2.3 million conversations in its first month. That was two-thirds of its customer service chats, equal to the work of 700 full-time agents. Resolution times fell to under 2 minutes from 11, and repeat inquiries dropped 25%. Klarna also said customer satisfaction was on par with human agents. These are self-reported, first-month figures, so read them as an upper bound.

The CFPB's 2023 issue spotlight on chatbots supplies the counterweight. It estimated that about 37% of the US population interacted with a bank's chatbot in 2022. It warned that customers get stuck in loops with no route to a person, and that large language models can produce inaccurate information. Chatbots aren't new. What generative models changed is that the assistant answers in its own words instead of picking from a script, which makes the handoff more important, not less. Design the escalation before the greeting.

Inside the firm, Morgan Stanley Wealth Management offers the clearest example. Its June 2024 release says the AI @ Morgan Stanley Assistant was fully rolled out in September 2023. It gives financial advisors quick access to the firm's research, and 98% of advisor teams had adopted it. The follow-on tool, Debrief, takes notes in client meetings with client consent. Afterward it summarizes key points, drafts an email for the advisor to edit and saves a note into Salesforce. Notice what it doesn't do, which is give advice. The advisor still owns the recommendation and the conversation about a client's goals, and the tool clears away the paperwork around it.

Documents are the quieter win, and possibly the most practical one. Fintech companies run on bank statements, pay stubs, tax returns, business filings and identity documents. A model can read them, extract the fields an underwriting or onboarding team needs and flag what's missing. That cuts manual data entry and the errors that come with it. It's also the safest place to start, because a person checks the extracted fields before anything gets decided.

Compliance teams read for a living too: new regulations, internal policies, marketing copy awaiting approval, alert narratives. A model can compare a policy against a new rule, draft a first version of a report or check copy against a checklist. FINRA's Regulatory Notice 24-09 says communications "generated by a human or technology tool" must meet the same Rule 2210 content standards. A compliance officer, not a model, still signs off.

Investment research follows the same pattern. Generative AI reads earnings call transcripts, filings, research notes and news, then summarizes market trends for analysts. Firms already use machine learning to inform trading strategies and portfolio optimization, as the US Treasury's December 2024 report describes. Letting a language model place trades is a different matter. Keep its output as an input to the strategy, not the execution engine, until you can test and explain it like any other model.

Two uses stretch the reading idea without breaking it. Stripe trained a payments foundation model on tens of billions of transactions and announced it in May 2025. Stripe says the model raised its detection rate for card-testing attacks on large businesses by 64% "practically overnight." Its previous models had cut card testing by 80% over two years. The transformer architecture behind chat models can learn patterns in transaction sequences, so fraud detection is moving toward larger models trained on more historical data. That raises the bar on data volume. It doesn't mean a chatbot should score payments.

Synthetic data is the other stretch. Generative models can create artificial transaction or customer records that mimic the statistical shape of real ones. Teams use them to test systems without exposing customer data, or to add rare fraud patterns to training sets. They still need privacy testing, since records generated too close to real customers can leak what they were meant to hide.

Fraudsters use generative AI too. Deepfake documents and cloned voices make identity checks harder, so layered verification matters more than any single check.

Keep generative AI away from the decision

Credit is the clearest line. US lenders must give specific reasons for adverse action. Under Regulation B, those reasons "must be specific," so telling an applicant they missed a qualifying score on a credit scoring system isn't enough. A language model that can't explain its own output can't meet that bar alone. Let predictive models and credit policy make the call. Use generative AI to summarize the file for the underwriter, draft the risk memo or turn reason codes into clear customer letters.

The same goes for any figure that must be exactly right every time. Payoff amounts, interest calculations, tax figures and regulatory capital numbers belong to the deterministic systems that run core operations. An assistant can display them. It shouldn't produce them.

Agents raise the stakes because they act instead of answer. On April 30, 2025, Visa announced Visa Intelligent Commerce, which opens its network to developers building AI agents that find and buy for consumers. Each consumer sets the spending limits. Named partners include Anthropic, IBM, Microsoft, Mistral AI, OpenAI, Perplexity, Samsung and Stripe. FINRA's 2026 report lists the risks: agents acting without human validation, acting beyond the user's authority, and reasoning that's hard to trace. If you're building for agents, start with limits, logging and an easy way for the customer to undo.

Even a reading job can be the wrong first project. If customer data sits in five systems that disagree, a model just summarizes the inconsistency faster, so fix the data first. And a small lender with modest volume may find that the generative features already in its core, CRM or contact center platform cover the need. Buy before you build.

US rules already cover it

No single US rule governs generative AI in financial services. Firms work under existing rules, and regulators keep saying so. Treasury's December 2024 report says generative AI can broaden opportunities while amplifying risks around data privacy, bias and third-party providers. It asks institutions to review AI use cases for compliance with existing laws before deployment. FINRA's Notice 24-09 says its rules are "intended to be technology neutral" and continue to apply. Rule 3110 still requires a supervisory system tailored to the business, and the 2026 oversight report adds a section on AI agents.

Banks got new model risk guidance on April 17, 2026, from the OCC, Federal Reserve and FDIC. The OCC leaves generative and agentic AI outside its scope, calling them "novel and rapidly evolving" technologies. Out of scope isn't unregulated. Consumer protection, fair lending and third-party risk rules still apply.

For a working structure, the voluntary NIST AI Risk Management Framework and its Generative AI Profile (NIST AI 600-1, July 2024) are the usual reference. If you serve EU customers, the EU AI Act (Regulation (EU) 2024/1689) adds a layer. Its Annex III lists AI used to evaluate individuals' creditworthiness or set their credit score as high-risk, with an exception for fraud detection.

Five ways generative AI goes wrong in financial services

The NIST Generative AI Profile names the risks. Five matter most in fintech.

1. Confabulation, NIST's term for "hallucinations," means confidently stated but false content. In finance, a wrong fee, rate or due date becomes a complaint, a refund or a regulatory finding.
2. Data privacy. Prompts and outputs can carry sensitive customer data, so know where every prompt goes and how long it's stored.
3. Information security. Assistants connected to account systems can be manipulated through crafted inputs. Limit what each one can see and do.
4. Human-AI configuration, which covers automation bias and over-reliance. A reviewer who approves 200 AI drafts a day stops reading them.
5. Value chain risk. Your model, hosting and data often come from different vendors. Treasury's report notes that generative AI increases smaller institutions' dependency on a few large companies.

Cost isn't on NIST's list, but it catches pilots out. Large models need significant compute to run, so price inference per transaction before you scale. Savings on paper can shrink once every customer chat carries a model bill.

Where to start

Pick one document-heavy workflow with a clear cost: loan file review, complaint drafting, KYC document checks or advisor meeting notes. Measure today's time and error rate first. If you can't put a number on the current process, you won't be able to prove the model helped.

Then build the controls around it. Run the model on an enterprise platform where your contract covers data retention and training use. The major clouds each offer one, in Amazon Bedrock, Microsoft Foundry and Google Cloud's Gemini platform, and model providers also sell enterprise tiers directly. Ground the model in approved policies, product terms and customer records through retrieval, so answers cite a source. Numbers come from systems of record, never from the model's memory.

Decide what a reviewer sees, what they approve and how corrections feed back. Then sample reviewed outputs weekly, so the review doesn't turn into a rubber stamp. Write the governance down early: purpose, data sources, owner, vendor terms, monitoring thresholds and the FINRA, CFPB or bank guidance that applies. Examiners will ask, and so will your board.

Say your first workflow is a loan file queue or a complaint inbox. Our generative AI development team would start by timing it as it runs today, then design the review step. Brainvire is an AWS Partner and a Microsoft Solutions Partner, and our fintech software development and AI and machine learning teams can connect the model to your core systems. For customer-facing assistants, see conversational AI for customer contact. Once you have a workflow picked, send us the details.

FAQ

Is there a US-only figure for the generative AI in fintech market?

Not on the public report pages we checked. Market.us gives North America more than a 36.5% share, about $0.69 billion in 2025. Technavio attributes about 31% of global growth to North America. Treat North America as the closest proxy and say so in any deck that uses it.

Is there a ChatGPT for finance?

Not one product, but many firms run assistants built on the same kind of model. Morgan Stanley's Debrief tool for advisors and Klarna's customer service assistant both run on OpenAI technology, according to the companies. Most financial firms run models through enterprise platforms with data controls rather than public consumer apps.

Will generative AI replace fintech staff?

It changes what they spend time on more than it removes them. Klarna said its assistant did work equal to 700 full-time agents in its first month. Even so, the CFPB warns that customers need a clear route to a human. The work that grows is review, exception handling and AI governance.

Related reading

Databricks Cost Optimization: 10 Tips From a Partner's Playbook to Stop Burning DBUsOctober 20, 2026 · 8 min readBeyond the Add-to-Cart Button: A Builder's Guide to Commerce Agents on ClaudeOctober 12, 2026 · 8 min readWhy Enterprises Need Databricks Consulting Partners to Navigate the Agentic AI EraOctober 12, 2026 · 8 min read