Glossary

Find a term

No term matches that search.

AI models and agents

Agent

A model the firm has given tools to reach specific systems, a loop to work a case step by step, and limits on what it may change. A person usually approves the result before money moves. See AI models, copilots and agents.

Audit trail

The log of every lookup, result and draft an agent made, so a reviewer or auditor can follow a case later.

Context window

The most text a model can handle in one call, counting the instructions, the files and its own answer.

Copilot

A model built into a product that suggests an answer or a draft for the case on screen. A person checks the suggestion and does the work.

Fine-tuning

Training an existing model further on a firm's own examples so it does one task better.

Function calling

Also called tool use. The model asks a piece of code to run by sending a short structured request with the tool's name and inputs, such as get_settlement_lines with a batch number. Code runs it and passes the result back.

Hallucination

A model stating something wrong with confidence, such as a fee that isn't on the statement.

Human in the loop

A person reviews and approves the model's or agent's work at set points, usually before money moves.

LLM

Large language model, the kind of AI model behind GPT, Claude, Gemini and Mistral. It takes text in and writes text out.

Machine learning

Models trained on labelled history to score or rank at volume, such as fraud scores or likely matches. They give a number without a reason. See Rules, machine learning or a model.

MCP

Model Context Protocol, a standard way to connect a model to a system's data and tools. Anthropic published it in November 2024, OpenAI and Google adopted it in 2025, and since December 2025 it has been run by the Agentic AI Foundation, part of the Linux Foundation. A vendor that offers MCP access lets a firm's agent read its data without a custom integration.

Model

A large language model called through its API, the way a business builds a model into its own systems.

OCR

Optical character recognition, which turns a scanned page into text a model can read. Test 1 used Tesseract on scanned statements.

Open weights

Anyone can download the model and run it on their own servers, so documents never have to leave the business.

Prompt

The instructions and material given to a model for a task.

RPA

Robotic process automation: software bots that repeat fixed steps on screen, such as copying data from one system to another.

Rules engine

Software that applies rules written in advance, such as pairing a settlement line with a bank receipt on amount, date and reference.

Temperature

A setting for how much a model's answers vary from one run to the next. The ScalePoint Research tests run at temperature zero.

Thinking

Working notes some models write before answering. The user doesn't normally see them, but they're billed as output.

Token

The unit models are priced in, roughly three quarters of a word. A long settlement file can run to tens of thousands of tokens.

Tool

A small piece of code an agent can ask to run, such as fetching the settlement lines or searching the ledger. The firm writes the tools and decides what each one can do.

How the tests work

Agent test

A test where the model works a task in steps with tools, deciding what to look up next. ScalePoint Research plans agent tests after its first round of one-step tests.

Agent workflow

A process map rating: the work comes as a queue of cases, each case needs several lookups across systems, and a person can check the result.

Agent-assisted

A process map rating: an agent can gather the facts and draft, and a person makes the decision.

Cluster

A group of processes on the process map that share their steps and their vendors, such as reconciliation or disputes. The map has 11.

Developer

The company that built and trained the model, such as Anthropic, Google or Mistral AI.

False alarm

An item a model reports as an error that isn't in the answer key.

Harness

The software that sends each task to a model, records the reply and passes it to the scorer. The tests use Inspect, from the UK AI Security Institute.

Host or route

The company whose servers ran the model for a test, such as Google Vertex for Claude or Microsoft Azure for GPT.

Money wrong

The value of everything a model missed or got wrong on a task, such as money missed in a reconciliation or fees wrong on a statement.

One-step test

A test where the model gets all the files at once, with no tools, and answers in one reply. ScalePoint Research's first round of tests is one-step.

OpenRouter

A service that sends requests to models from many developers through one connection. The tests call models through it.

Planted error

A mistake put into the test files on purpose, with a known value.

Process map

ScalePoint Research's map of 51 payments and finance processes across six types of firm, each rated agent workflow, agent-assisted or single step.

Run

One attempt by one model at one task, from sending the files to marking the answer.

Single step

A process map rating: the work is one read or one calculation, or it already runs on an engine. A model may do the step, and there's no case for an agent to work.

Worst run

A model's poorest result across its runs of the same task. The leaderboards show it next to the average.

Zero data retention

The host keeps no copy of the files sent in or the answer once it has replied.

Cost

Baseline

Today's cost per case, time to close and error rate, measured before a pilot so there's something to compare the AI against.

Cost per task

What one run of one task cost in model fees at list prices: one statement in Test 1, one month of files in Test 2.

Escalation rate

The share of cases an agent hands to a person instead of finishing.

Full cost per case

The model fees, plus the checking time, plus the mistakes that get through, plus fixed costs spread across the cases. See The unit economics of AI.

Full review

A person checks every case the model works. In Test 1, the level for models with up to 2% of fees wrong.

Override rate

How often reviewers change or reject an agent's draft. A rising rate is an early sign that something has drifted.

Price per million tokens

The list price for text sent in and text written back. Text written back usually costs more.

Spot checks

A person checks a sample of cases, such as one in ten. In Test 1, the level for models with up to 0.5% of fees wrong.

Use as is

The model's work goes through with no checks. In Test 1, the level for models with 0.1% of fees wrong or less.

Payments and finance operations

Acquirer

The bank or payment company that processes card payments for a merchant and pays the money into the merchant's account.

APP scam

Authorised push payment scam, where a customer is tricked into sending a payment themselves.

Blended pricing

One rate per card type covers all of the acquirer's fees, with interchange and scheme fees bundled in.

Break

An item that reconciliation software can't match, which a person or an agent has to trace to a cause.

Cash application

Matching incoming payments to the invoices they pay.

Chargeback

A card payment the cardholder has disputed, which the acquirer takes back from the merchant.

EMI

Electronic money institution, a firm licensed to issue e-money and make payments, such as many payment apps and cross-border providers.

Interchange

The fee the cardholder's bank receives on each card payment, paid by the merchant through its acquirer.

Interchange++

Pricing where the acquirer passes interchange and scheme fees through at cost and adds its own margin on top.

KYB

Know your business: checking a business customer's registration, owners and controllers before taking it on.

KYC

Know your customer: checking a customer's identity and risk before taking them on, and again at set intervals.

MLRO

Money laundering reporting officer, the person responsible for a firm's anti-money laundering controls and for reporting suspicions.

Nostro account

An account a bank or EMI holds at another bank, often in a foreign currency.

Payment repair

Fixing a payment rejected for missing or wrong data, such as a missing country code, and sending it again.

PSP

Payment service provider, a firm that lets merchants accept payments, often across several methods.

Reconciliation

Checking that the money a firm expected matches what arrived, across files from acquirers, the bank and its own ledger, and explaining every difference.

Safeguarding

Keeping customers' money separate from the firm's own, as UK and EU rules require of EMIs and payment institutions.

Sanctions screening

Checking customers and payments against sanctions lists, then clearing or escalating each hit.

SAR

Suspicious activity report, filed with the national authority when a firm suspects money laundering. In the UK it goes to the National Crime Agency.

Scheme fees

Fees the card schemes, such as Visa and Mastercard, charge on each payment, on top of interchange.

Settlement file

The acquirer's record of what it paid a merchant for a batch or a day: the sales, minus fees, chargebacks and refunds.

Transaction monitoring

Reviewing payments for signs of money laundering or fraud, then closing, investigating or escalating each alert.

Rules and regulation

Consumer Duty

The FCA's rules requiring UK firms to deliver good outcomes for retail customers. The FCA relies on it, among other rules, for AI.

DORA

The EU Digital Operational Resilience Act, in force since January 2025. It covers ICT risk at EU financial firms, including software and AI bought from vendors.

EU AI Act

The EU's law on AI. It classes some uses as high-risk, such as assessing the creditworthiness of individuals, with requirements for human oversight that are due to apply from December 2027.

GDPR automated decisions

GDPR limits decisions with legal or similarly significant effects made about a person by automated means alone. The UK version, amended in 2025, still gives people the right to have a person review such a decision.

SM&CR

The UK Senior Managers and Certification Regime. A named senior manager is accountable for each function, including what an AI agent does in it.

SR 26-2

US model risk guidance from the Federal Reserve, issued with OCC Bulletin 2026-13 in April 2026 to replace SR 11-7. It leaves generative and agentic AI out of scope for now.

Three lines of defence

The usual control set-up in financial firms: the business owns its risks, risk and compliance set limits and test them, and internal audit checks the controls work.