Financial services adopted machine learning long before the LLM wave — fraud scoring and credit models have been in production for decades. What's changed in 2025–2026 is the second layer: language models and agents that sit on top of those classic models, turning scores into decisions, alerts into investigations, and regulations into checklists. That combination is where the real leverage is, and it's also where teams get into trouble if they treat a regulated domain like a chatbot playground.
Here's how we think about the four problems fintech teams bring to us most often.
Real-time fraud: the model is half the system
Card-not-present fraud, account takeover, synthetic identities — the attack surface keeps widening, and fraudsters now use generative AI too. The detection side is a classic streaming problem: behavioral features (device, velocity, session patterns) feeding a low-latency model with a decision budget measured in milliseconds. The newer piece is what happens after the flag. LLM-powered case agents assemble the evidence — transaction history, device fingerprints, linked accounts — into a readable investigation summary, so a human analyst clears or escalates in minutes instead of an hour. Fraud teams don't need fewer alerts; they need each alert to arrive pre-investigated.
Credit risk: explainability is a feature requirement, not a nice-to-have
Lenders are required to tell applicants why they were declined. That makes uninterpretable models a non-starter for adverse-action decisions, no matter how well they score. The pattern that works: keep the decisioning model constrained and auditable (gradient boosting with monotonic constraints, reason codes derived from feature attributions), and use LLMs where language is the actual job — turning reason codes into clear, compliant adverse-action notices, summarizing bank-statement cash flow for underwriters, and drafting the credit memo a human signs. The model that decides and the model that explains don't have to be the same model.
AML and KYC: agents that do the document grind
Onboarding and anti-money-laundering work is drowning in documents — IDs, proofs of address, corporate registries, sanctions lists, transaction narratives. This is a natural fit for bounded agents: extract and cross-check entity data, screen against watchlists, flag mismatches, and assemble a suspicious-activity narrative for a compliance officer to review. Two guardrails are non-negotiable. Every extracted fact needs a citation back to the source document, and the final filing decision stays with a human. Done right, the same headcount clears a multiple of the case volume with a better audit trail than the manual process ever produced.
Compliance copilots: keeping up with the rulebook
Regulations change constantly, across every jurisdiction you operate in. A retrieval-grounded copilot over your actual obligations register, policies and regulatory feeds turns "did anything change that affects us?" from a quarterly panic into a standing daily answer. The critical engineering detail is grounding: the copilot must quote and link the source clause, and say "not found" rather than improvise. In compliance, a confident wrong answer is worse than no answer.
The governance layer that makes all of it deployable
None of the above survives a model-risk-management review without discipline: versioned models and prompts, evaluation suites with regulator-relevant metrics (false-positive rates by segment, fairness checks, drift monitors), immutable decision logs, and a documented human-override path. This is the same evaluation-first culture we push in every agentic deployment — fintech just makes it mandatory. Teams that build the governance layer first ship faster afterwards, because every new use case inherits the rails.
The short version
AI in fintech works when the boring constraints are treated as design inputs: millisecond latency for fraud, reason codes for credit, citations for AML, grounding for compliance, and governance underneath all of it. The black box isn't a model property — it's a system design failure, and it's avoidable.
Building in payments, lending or compliance and wrestling with any of this? Tell us what you're up against.