How AI is used in banking and finance (2026)
By MacrosLM Team · Reviewed by Damira Baigozha, ex-PwC Valuation & M&A Advisory Expert
AI in banking and finance is the mix of statistical models, machine learning, and — more recently — generative and agentic systems used to detect fraud, price risk, execute trades, serve customers, and close the books. Some of this is decades old (rules engines, credit scorecards); some is new and unproven at scale (LLM agents that draft memos or reconcile accounts). Treating those two generations as equally mature is a common, costly mistake.
This is a practitioner survey: where AI in the finance industry is deployed today, how mature each use case is, and what it still can't do.
Two generations of AI, not one
Most talk of AI in financial markets blurs two different technology generations. Rules-based and classical ML — credit scorecards, gradient-boosted fraud models, statistical arbitrage signals — has been in production for 15–30 years, heavily validated, usually owned by a model risk management framework such as the Federal Reserve's SR 11-7 guidance on model risk management. Generative and agentic AI — LLMs that draft memos, summarize filings, answer customer questions, or chain multi-step tasks — is newer, less standardized, and prone to fabricating plausible but wrong output if unconstrained. Our explainer on what agentic AI actually means in a finance context covers where "agentic" is real versus marketing.
Fraud detection and anti-money laundering (AML)
Fraud and AML are the oldest, most mature applications here. Banks have used supervised ML — trained on labeled transaction data — to score fraud risk in real time for years, looking at velocity, geography, and device signals alongside rules engines. AML adds entity resolution, network analysis, and alert triage, where ML ranks alerts so investigators work the highest-risk cases first. Generative AI is now layered on to draft suspicious activity report (SAR) narratives from structured alert data — a real time-saver, but one a human investigator must still verify before filing with FinCEN. Related: KYC risk scoring, which feeds many of the same signals into onboarding.
Credit underwriting and lending decisions
Credit scoring was one of the first large-scale statistical modeling applications in finance, and modern underwriting extends it with ML that incorporates cash-flow and alternative data — under much tighter fair-lending and explainability rules than fraud models face. A model that can't explain why it declined an applicant creates real exposure: the CFPB has made clear (Circular 2022-03) that "the algorithm was too complex to explain" is not a defense under adverse-action rules. The newer layer is generative AI drafting underwriting memos and pulling structured data out of tax returns and bank statements — it accelerates information-gathering, it doesn't replace the credit decision. See our roundup of AI tools for credit analysis.
Feed it the real documents and demand a page reference for every figure.
Prompt
From the attached financials and tax returns, spread the borrower into a standard format and compute leverage, DSCR, coverage, and liquidity. Give a page reference for every figure and don't estimate anything not in the documents.
What good looks like: a standard spread where each ratio traces to a page.
In MacrosLM: the credit agents spread the file and run the ratios, each figure source-traced.
The memo is a first draft — the credit decision and policy control stay with the lender.
Prompt
Draft an underwriting memo: business overview, financial trends, ratio analysis, covenant and debt-capacity read, and the key risks. Cite every figure back to its source document.
What good looks like: a committee-ready draft with citations, not prose you can't verify.
In MacrosLM: the Credit Memo Generator drafts it with examiner-ready citations on every figure.
Ask for hits with sources and a rating — a triage list, not a decision.
Prompt
Screen this entity and its principals for sanctions and PEP exposure and adverse media. Summarize each hit with its source and assign a risk rating, flagging anything that needs a human review.
What good looks like: hits with sources and a rating — a triage list, not a decision.
In MacrosLM: the KYC Risk Scoring Engine and Sanctions & PEP Screening pull from regulatory sources with each hit traced.
Re-run the spread and the screening on a cadence so a slipping covenant or a new hit surfaces early rather than at renewal.
In MacrosLM: scheduled re-runs (daily or quarterly, per your instructions) re-score the file and flag what changed.
Or run the whole first pass at once
MacrosLM runs the spread, the memo, and the screening in one workspace — source-traced and re-runnable on your cadence. The credit decision, the explainability, and the sign-off stay human; every regulator still holds a person accountable, not the model.
Algorithmic trading and market analysis
AI in financial markets spans a wide maturity range. Execution algorithms that slice large orders to minimize market impact, and statistical arbitrage strategies, are long-established and heavily quantitative — classical ML and optimization, not generative AI. What's newer is generative AI applied to research — summarizing earnings calls and analyst notes, drafting first-pass commentary. This is assistive, not autonomous. See how AI is used in stock analysis for where generative tools sit versus where quantitative models still do the heavy lifting.
Customer service and front-office support
Chatbots in retail banking started as rules-based decision trees and have moved toward LLM-based assistants that handle open-ended questions and draft responses for human agents. These assistants are genuinely useful for routing, summarization, and first drafts, but banks generally keep a human — or a hard guardrail — between the model and any action that moves money, since generative models can still produce confident, wrong answers.
The back-office finance function: close, reconciliation, reporting
This is where AI in financial services meets corporate finance and accounting teams directly. ML has automated large parts of transaction matching for years — invoices to purchase orders to receipts is a natural fit for pattern-matching, underpinning processes like the three-way match. More recently, generative AI drafts reconciliation explanations and variance narratives, helping close teams triage exceptions faster. See machine learning applications in finance for the wider set of ML techniques across these workflows.
Use case maturity at a glance
| Use case | What AI actually does | Maturity |
|---|---|---|
| Card/transaction fraud scoring | Supervised ML scores transactions in real time against fraud patterns | Mature — production standard for years |
| AML alert triage & SAR drafting | ML ranks alerts by risk; generative AI drafts narrative text | Mixed — triage mature, drafting emerging |
| Credit scoring (consumer/SME) | ML/statistical models predict default risk from structured and alternative data | Mature, tightly regulated |
| Underwriting memo drafting | Generative AI summarizes financials and drafts first-draft analysis | Emerging — human review required |
| Execution & stat-arb trading algorithms | Quantitative models optimize order execution and short-term signals | Mature |
| Research summarization | LLMs summarize filings, calls, and notes into first drafts | Emerging |
| Customer service chat | Rules-based intents plus LLM assistants for open-ended queries | Mixed — mature for routing, emerging otherwise |
| Reconciliation & close support | ML matching plus generative AI for variance narratives | Mixed — matching mature, drafting emerging |
| Agentic multi-step workflows | LLM-driven agents chain tasks with limited human checkpoints | Early — pilots, not standard practice |
Classical ML has 15–30 years and heavy validation behind it; generative and agentic tools are newer and need careful human review. Knowing which is which matters more than the "AI" label.
What AI in banking and finance still can't do reliably
Model risk doesn't disappear because a model is newer — every ML or generative model in a regulated decision needs a governance owner, documented validation, and drift monitoring, the discipline SR 11-7 established for classical models and now extending to harder-to-validate LLM tools. Generative models hallucinate — a fluent, confident summary can contain a wrong number, easier to miss than an obvious error. Explainability is not optional: adverse-action rules and examiner expectations mean a model that can't produce a defensible rationale is a compliance liability, and new regimes like the EU AI Act classify much credit and insurance decisioning as "high-risk." Accountability stays human — no regulator accepts "the model decided." And data quality is still the binding constraint: most failures blamed on "the AI" trace back to incomplete or poorly labeled underlying data.
Where MacrosLM fits
Disclosure: MacrosLM is our own product. MacrosLM sits on the analysis-heavy side of this picture rather than in real-time fraud scoring or trade execution, which need purpose-built infrastructure. It runs agents like the KYC Risk Scoring Engine, Sanctions & PEP Screening, and Credit Memo Generator, and pulls from regulatory sources through built-in connectors such as FFIEC bank Call Reports, OpenSanctions, and CFPB complaints — with every figure tied back to its source. Same principle as the limitations above: it accelerates first-draft work; a qualified professional still reviews and owns the output.
Bottom line
AI in banking and finance is really two things under one label: mature, well-governed statistical and ML systems that have run fraud, credit, and trading decisions for years, and a newer wave of generative and agentic tools that speed up drafting, summarizing, and triage but still need careful human review. Knowing which category a given tool belongs to matters more than whether it's labeled "AI."
Sources
- Federal Reserve — SR 11-7, Guidance on Model Risk Management — the model-risk governance baseline for regulated models.
- CFPB — Circular 2022-03, adverse-action notification and complex algorithms — why 'too complex to explain' is not a defense.
- European Commission — Regulatory framework for AI (EU AI Act) — high-risk classification for credit and insurance decisioning.
- FinCEN — suspicious activity reporting.
This article is educational and reflects the AI-in-finance landscape as of 2026, which changes quickly. It is not legal, compliance, or investment advice; regulatory obligations should be confirmed with the applicable authority and qualified counsel.
Frequently asked questions
- Is AI already widely used in banking?
- Yes — rules-based and machine learning systems have been standard in fraud detection, credit scoring, and algorithmic trading for well over a decade. What's newer and less standardized is generative and agentic AI, still mostly in assistive roles like drafting and summarization rather than autonomous decision-making.
- What's the difference between traditional AI and generative AI in finance?
- Traditional ML models (scorecards, fraud classifiers) are trained on structured historical data to predict a specific outcome, like default or fraud probability. Generative AI produces novel text or analysis from prompts and can hallucinate — better suited to drafting and summarization than to standalone numeric decisions.
- Can AI replace credit analysts or auditors?
- Not currently. AI can accelerate data gathering, summarization, and first-draft analysis, but judgment, explainability, and accountability for regulated decisions still require qualified human professionals. Firms remain liable for AI-assisted outputs the same way they would for any staff work product.
- Is AI in financial markets mostly used for trading?
- Trading is one major use, but far from the only one — AI in financial markets also covers research summarization, risk modeling, and portfolio analytics. Execution and statistical arbitrage algorithms are the most mature; generative AI applied to research is newer and more assistive.
- Does using AI reduce a bank's compliance obligations?
- No. Regulators hold institutions accountable for AI-assisted decisions the same way as any other decision — documentation, explainability, and human accountability requirements still apply, and in practice AI adoption usually increases the model governance workload rather than reducing it.
Reviewed by Damira Baigozha
ex-PwC Valuation & M&A Advisory Expert. Written by the MacrosLM editorial team.
View profile →
