NLP (Natural Language Processing)

NLP, or natural language processing, is the branch of AI that reads and makes sense of human language. In payments it is put to work on the text a system gathers but cannot easily use. That means dispute notes, support chats, sanctions name lists, bad-press reports and free-text payment references. Much of it feeds aml work. A rules engine handles numbers well. It handles a sentence badly. That gap is what NLP is brought in to close. FATF makes the same point in wider terms: AI draws insight from data of many types, sources and levels of quality, both structured and not.

The clearest use is in financial crime work. FATF's report on new technologies for anti-money-laundering notes that NLP and fuzzy matching tools allow a better cut in false hits and misses. Its example is sanctions screening. It adds that such tools could bring real gains, taking in the knack of cross-checking bad press. Anyone who has watched a team work through name-match alerts will know the problem being set out.

Where It Gets Used In Payments

Four jobs come up most. Screening names against sanctions and watch lists, where spelling shifts from one alphabet to the next. Reading text to help build a suspicious activity report. Reading bad press to flag a customer worth a closer look. Sorting support and dispute text so a case lands in the right queue. And pulling meaning out of free-text notes that were not written for a machine to read. None of these makes the call. They shape what a person sees first. In a queue of thousands of alerts, that ordering is most of the value on offer, and it is why a risk engine tends to sit around them rather than under them.

Why Name Matching Is Harder Than It Looks

One person can show up a dozen ways across systems. Word order swaps. Initials stand in for names. Spelling shifts between alphabets, and typos pile up. An exact match misses most of that. A loose match floods a team with alerts. FATF frames the aim as building sharper rules by studying behaviour and patterns. That is a fair way to put the trade-off between too many alerts and too few.

How Much Of This Is In Use

The Bank of England and FCA survey of artificial intelligence in UK financial services gives some scale. It came out in November 2024 and covered 118 firms. It found 75% of firms already using AI, with 10% more set to within 3 years. On fraud, 33% were using it and 31% more planned to. Foundation models, the family behind current language tools, made up 17% of all use cases. These are UK figures.

The Third-Party Question

The same survey found a third of all AI use cases were built by third parties, up from 17% in 2022. The top 3 model firms made up 44% of those named, up from 18%. For a payments team that is a resilience point as much as a model one. If a screening tool leans on an outside model, the firm has taken on a reliance it does not control. The survey hints that plenty of firms lean on the same few models, which turns one outage into a shared problem.

Where The Law Draws Lines

The EU AI Act sorts systems by risk. The European Commission's page on the AI framework sets out 4 levels. They run from unacceptable risk, where a system is banned, through high risk and openness risk down to minimal risk. It names credit scoring as high risk, with the example of turning down a loan. The Act came into force on 1 August 2024, and its duties phase in over the years after.

The Carve-Out Worth Knowing

Payments teams should read one detail closely. The AI Act's own recitals treat systems that score a person's credit as high risk, because they decide who gets money. But they also say that AI set out in Union law for spotting fraud in financial services should not count as high risk. So fraud work and credit scoring are not treated alike, even where the same method sits under both. This is EU law, and other markets draw their own line.

Its Limits In Practice

These models work on odds, which is a polite way of saying they are at times wrong with great confidence. On a screening queue that shows up as a match that looks right and is not. So the sane pattern keeps a person in the loop on anything that hits a customer, and keeps a note of why each call was made. finera.'s overview of how intelligent fraud management works across transactions covers where machine signals fit next to human review.

What To Ask Of A Vendor

Useful questions are about proof, not features. What the false hit rate was on data like yours, not on a test set. Whether a call can be explained to a watchdog or a customer. Where the model runs and what data leaves your systems. And how changes are tested before they reach live traffic. finera.'s payment fraud detection and risk management capability is built round weighing signals. Its look at emerging payment technologies worth watching sets this in the wider picture.

‍