Research
What can AI do in payments and finance operations?
Research on AI in payments and finance operations, published free: models tested on payments tasks, and what vendors offer for each process.
The research agenda
What
Which payments and finance operations jobs AI can do today, what its mistakes cost, and which vendors sell it for each process. The research covers work at merchants, acquirers, PSPs and fintechs under UK and European rules. It's written for the finance, operations and product leaders deciding whether to build or buy, and the risk teams who sign off the supplier.
Why
Payments firms are being offered AI agents for more and more of their operations, and new models and vendors arrive every month. Choosing between them means weighing what each gets right, what it costs at your volumes, how fast it runs, where your data goes, and what DORA and the EU AI Act ask of the supplier.
How
Each piece of research takes one process. A landscape scan finds the vendors that sell AI agents for it, compares them and gives a high-level view of how mature each agentic offering is.
Alongside the scan, a simple model test shows what a build could achieve and what it would cost to run, with frontier models tested next to open-weight models you can run on your own servers. Weighing the vendors against the effort to build points to one of three answers.
Vendors cover the process well, so the question is which one and what it has to prove.
The gaps are too big, and the tests show which models can do the work and at what cost.
A vendor covers most of it, and the rest is built around it.
Latest research
The first paper tested 34 AI models on reading UK card acquiring statements, scoring every fee they got wrong in pounds against what each model costs to run. With good instructions, a model at half a cent a statement matched one at 25 cents.
Read the paper online →Leaderboard: all 34 models →
- DeepSeek V4.1 Flash China · ZDR via third parties
- GLM 5.3 China · ZDR via Reka
- Mistral Large 3 France · ZDR direct
- Gemini 3.8 Flash US · ZDR direct
- Grok 4.7 US · ZDR direct
- Claude Opus 5.5 US · ZDR via Google
- GPT-5.5 US · ZDR via Azure
Seven finalists, round 2. Each arrow runs from the plain prompt to the expert prompt. The fees-wrong scale is stretched to the right of the 0.1% line. All runs used zero data retention (ZDR). Direct means the developer's own service qualified.
White paper · PDFAI economics in fintech: reading card statementsDownload the PDF →
Research by process
I'm working through five core processes, one at a time. Each gets a vendor landscape scan and a simple model test, run side by side.
Reading card acquiring statements
Pulling every fee out of a UK acquiring statement and classifying it correctly.
- Model testPublished34 models, scored on fees wrong in pounds against price. Read the paper
- Vendor landscapePlannedWhich vendors read and check acquiring statements with AI, and how mature each offering is.
Reconciling card income
Matching a merchant's settlement files, bank lines and ledger for a month, and finding and valuing every difference.
- Vendor landscapeNextWhich vendors sell AI agents for card reconciliation, how they compare and how mature each offering is. Each vendor checks its description before it's published.
- Model testNextA month of an invented merchant's files in one go, scored on money missed and false alarms against price per run.
Chargebacks
Deciding which chargebacks to fight and gathering the evidence.
- Vendor landscapeLater
- Model testLater
Merchant onboarding and KYB
Checking a new merchant's documents and risk before it's boarded.
- Vendor landscapeLater
- Model testLater
Payment exceptions and repairs
Fixing failed and returned payments, and answering the queries that come with them.
- Vendor landscapeLater
- Model testLater
Bespoke work
I can run the same research on one of your own processes, or advise on whether to build, buy or combine the two, which model to use, and the record DORA and the EU AI Act ask for on how each model was chosen and tested.
Writing on AI
AI Is Breaking the Zero Marginal Cost Assumption in Payments and Fintech
Token-based pricing puts real marginal cost back into software, and payments businesses are built on the assumption it is not there.Read → ArticleTwo AI Systems, Same Output, Completely Different Economics
LLM, ML or composable: the implementation you choose sets a cost curve, not a cost figure. What that means for payments and fintech at real volume.Read → ArticleFintech Moats in an AI World
Many fintech advantages were interface-layer rather than infrastructural. AI is compressing those, and moving defensibility underneath the product.Read →How to start
For custom research or advisory work, drop me an email and we can talk it through. Every piece of work is agreed in a fixed-price statement of work.
martin@scalepointpartners.com