We are in a strange place with regards to commercial use of Large Language Model (LLM) technology. Millions of people use models like ChatGPT and Claude every day for everything from shopping to therapy, and many companies are utilising it in customer service roles. What we don’t see nearly enough of is the use of their equally powerful abilities in data analysis, and how that can be paired with their ability to communicate in a way that humans understand.
According to Eurostat data published in December 2025, only 20 percent of EU enterprises with at least ten employees were using AI technologies in 2025, a jump of 6.5 percentage points from the prior year. Yet adoption inside merchant operations, where teams juggle payment processing, chargebacks, settlement reconciliation, and customer support, remains uneven.
Worries about ‘hallucinations’ are of course valid in business settings, but it is equally possible that LLMs have advanced enough – or will soon – that they can be genuine force-multipliers in business.
Payment ecosystems generate enormous volumes of structured and unstructured data, much of which sits unused inside dashboards or spreadsheets. That data will be visible as lines on a graph, but it will be lacking a why that LLMs can provide.
LLMs are well suited to bridging the gap between raw transaction logs and the people who need to interpret them. Eurostat also reports that text mining was the most common AI application in 2025, used by roughly 12 percent of European enterprises. That technique sits at the core of how language models surface meaning from log files, support tickets, and merchant correspondence.
Turning payment data into plain-language insights
One of the most immediate wins for merchants is using LLMs as a translation layer between raw payment data and the operations staff who need to act on it. Acquirers, gateways, and orchestration platforms produce decline codes, fraud flags, and authorisation patterns that often require specialist knowledge to interpret. A well-prompted language model or even AI agent connected to internal data can summarise weekly performance shifts, highlight unusual decline spikes by issuer or region, and flag concentration risks that a human analyst might miss in a hurried review. The output reads more like a written briefing than a dashboard, which makes it accessible to commercial and product staff alongside payments specialists.
Crucially, this isn’t about replacing analytics. It’s about reducing the cognitive distance between data and decision. McKinsey’s analysis of generative AI in financial services estimates that the technology could contribute between 200 billion and 340 billion dollars of annual value to global banking, with much of that gain tied to productivity in customer-facing and operational roles. Merchant operations sit firmly within that opportunity set.
Smarter merchant support and faster issue resolution
The second high-value use case is in merchant support. Support teams handle a constant flow of queries about failed transactions, settlement timings, refund processing, and KYC requirements, and the answers usually live across multiple internal systems. LLMs equipped with retrieval-augmented generation can pull from internal documentation, ticket histories, and live transaction data to draft accurate first responses, suggest next steps, or resolve simple queries autonomously.
McKinsey’s research on generative AI agents found that a study of 5,000 customer service agents using such tools saw issue resolution increase by 14 percent per hour while handling time fell by 9 percent.
Risk and fraud teams are seeing parallel benefits. Academic and industry research has shown that LLMs can act as a reasoning layer on top of traditional fraud models, providing analysts with plain-language explanations for why a transaction was flagged. A 2024 paper from Google researchers on scam detection in GPay India demonstrated how an LLM-based reasoning engine could augment existing machine learning classifiers, particularly for cases that would otherwise require manual review.
More recent work combining language models with graph-based fraud detection on ecommerce payment data has produced similar results, suggesting that hybrid architectures are emerging as a practical standard rather than a research curiosity.
A productivity multiplier across finance, support and product
The third pattern, and arguably the most underappreciated, is the use of LLMs as a productivity multiplier across the wider business. Finance teams running monthly reconciliation can use language models to summarise variances against forecast, draft commentary for management reports, and surface anomalies in fee structures or interchange categorisation. Product teams can use the same tools to digest user feedback, support ticket trends, and competitor release notes, turning what was a manual research exercise into a few minutes of structured reading.
Eurostat reports that, among EU enterprises already deploying AI, 23 percent are applying it to accounting, controlling, or finance management, showing the breadth of finance-adjacent adoption.
These benefits arrive with important caveats. Language models are confident generators of plausible text, which means that they can produce inaccurate outputs when grounding data is missing, ambiguous, or out of date. Sensible deployments use retrieval over verified internal sources, keep humans in the loop for any customer-facing or financially material output, and audit prompts and responses to detect drift.
McKinsey’s own field research suggests that only 11 percent of companies have scaled generative AI to enterprise level, with operations-related domains lagging behind, often because pilots are not paired with workflow redesign.
For merchants weighing where to start, the practical advice is to pick use cases where the cost of a wrong answer is low, and the time saved is visible. Drafting internal performance summaries, accelerating support responses with human review, and helping analysts navigate complex decline patterns are all candidates that meet that bar. Larger applications, such as autonomous fraud decisioning or fully automated merchant communications, deserve a more cautious posture and stronger guardrails.
The merchants likely to extract the most value over the next two to three years will be those that treat LLMs as augmentation for their existing teams rather than as a substitute for them, building incrementally on payment data foundations they already have.
