Fintech has a peculiar relationship with AI. The industry that runs on models, credit scoring is older than the transistor, and machine-learning fraud detection has guarded card networks for decades, keeps announcing AI as if it arrived last quarter. What actually arrived is a new capability class: models that read documents, follow written policy, and produce a reasoned narrative. Where that lands productively is mostly not where the press releases point.
The showpieces, advice chatbots, autonomous agents moving customer money, are the hardest possible place to start: maximal regulatory exposure, minimal customer trust, and a failure mode measured in someone's rent money. The returns are concentrated somewhere much less cinematic: the operational back office that every fintech runs and none of them show on the landing page.
Follow the ops headcount
A useful exercise for any fintech leadership team: list where operational headcount actually goes. For most, it's some mix of onboarding review, transaction monitoring alerts, disputes, and support escalations. That list is the AI roadmap, because each of those is a high-volume queue of semi-structured judgment calls made against written policy, which is precisely the shape of work current models handle well under supervision.
- KYC and onboarding document review. Passports, proofs of address, business registrations, source-of-funds letters, cross-checked against application data and watchlist results. A model can extract, reconcile, flag mismatches, and draft the reviewer's summary, cutting review time from hours to minutes for the clean majority and reserving human attention for the genuinely ambiguous. Onboarding drop-off is a revenue problem, not just a cost problem: every day of verification delay loses signups to whoever verifies faster.
- Transaction categorization and enrichment. Raw merchant strings like "AMZN MKTP US*2K4" become merchant, category, and recurrence data. Rules and embeddings handle the bulk cheaply; an LLM resolves the long tail the rules miss. Unglamorous, but it compounds: categorization quality feeds budgeting features, lending decisions, and fraud signals alike. It's the data layer everything else stands on, the way a normalized feed from forty exchanges was for Allmarkets.
- Disputes and chargebacks. A dispute is a document assembly problem with a deadline: gather the transaction record, the customer's claim, the merchant's evidence, and the network's reason-code requirements, then produce a representment package. Models draft it, humans file it, and win rates go up because the packages are complete and on time, every time. Few queues convert model capability into recovered dollars this directly.
- The fraud investigation layer. The detection models mostly exist and work. The bottleneck is what happens after the alert fires: an analyst assembles account history, device signals, linked accounts, and recent behavior into a decision. An LLM can do that assembly and draft the case narrative in seconds, which matters twice, once for analyst throughput, and again because false positives that freeze legitimate accounts are among the costliest customer experiences in the industry. Faster case resolution is retention, not just efficiency.
- Compliance reporting. Suspicious-activity narratives and periodic-review summaries follow strict templates over case facts. Drafting them is exactly the structured-writing task models excel at, with the compliance officer editing and signing. Teams reclaim a startling fraction of investigator time from writing alone.
The common architecture across all five: the model reads and drafts, the human decides, and the system records everything. In a regulated industry that last clause is not hygiene, it's the product.
Explainability is a feature requirement
What separates fintech from most AI deployments is that the explanation is often legally load-bearing. Adverse-action rules mean a declined applicant is owed reasons. Model risk management frameworks, and increasingly consumer protection regulators, expect institutions to demonstrate their models are governed, monitored, and fair. "The model said so" fails that bar categorically.
This constraint should shape the engineering from day one:
- Log the full decision trace. Input documents, retrieved context, model version, prompt version, output, and the human's final action, replayable months later when an auditor or a complaint asks. If you can't reconstruct a decision, in this industry, you didn't make it, it happened to you.
- Keep humans on every adverse path. Approvals can be fast-tracked; declines, freezes, and offboarding decisions get human sign-off. This is where regulator attention concentrates, and where a wrong call does real harm to a real person.
- Evaluate like it's a credit model, because regulators will treat it like one. A golden set of past cases with known outcomes, scored on every prompt or model change, with fairness slices across customer segments. Drift monitoring in production. It is the bar credit assessment logic has to clear on a core banking platform like the one we helped build for Five Degrees. This is the eval discipline we apply everywhere, with the stakes turned up: a quality regression here isn't a worse sentence, it's a wrongly frozen account.
- Route by stakes, not by task. A cheap model for merchant-string cleanup, a frontier model for the source-of-funds narrative, deterministic code wherever the rule is expressible as a rule. Fintech margins are thin; per-decision cost discipline is what makes the unit economics work at transaction volume.
What about the customer-facing ambitions?
They're real, eventually. Support automation is already working where it's scoped honestly: card reissues, statement questions, limits, with clean escalation and without the bot ever guessing about money. The fully agentic vision, software that moves funds and makes financial decisions autonomously, is where the industry is heading and where almost nobody should start. Trust in fintech is asymmetric: years to build, one wrongly blocked payroll to lose. You earn the right to automate customer-facing decisions by first proving, in the back office, that your systems are evaluated, audited, and boring.
That's the strategic argument in one line: in financial services, boring is the moat. The companies compounding real advantages from AI right now aren't the ones with the flashiest demo. They're the ones whose onboarding clears in minutes, whose disputes get filed complete, whose fraud team closes cases before lunch, and whose auditors leave satisfied. Every one of those is achievable with today's models and last decade's engineering discipline.
Start where the ops headcount is. The returns are hiding in plain sight, in the queues. It is where every fintech engagement we take on starts too.
