---
title: "AI Agent"
url: "https://mahouse.world/glossary/ai-agent"
type: "glossary term"
topic: "Agents"
author: "Ihor Vlasov"
reviewed: "2026-09-20"
published: "2026-08-10"
---

# AI Agent

An AI agent is a program built on a foundation model that pursues a goal over several steps — reading, deciding, calling tools such as a database, an API or a payment system, checking the result and continuing — rather than answering one prompt with one reply. In fintech it is the form AI takes when it is allowed to do things, not only to say them.

**Also known as:** Agent, Agentic AI, Autonomous agent, LLM agent, Autonomous agents, Personal AI agent, AI agents

## Key facts

- **Used at:** customer service, onboarding and KYC review, reconciliation, dispute handling, credit operations, coding
- **Components:** a model, a set of tools it may call, a memory or context, and a loop that plans, acts and observes
- **Autonomy:** from "suggests, human approves" to "acts within limits, human audits" — the limit is the design decision
- **Regulatory:** an agent that moves money or makes a credit decision is doing a regulated activity, and the firm is responsible for it
- **In deals:** acquirers pay for agents with a track record — logged actions, error rates, the controls around them — not for demos

## How an AI agent works in practice

A customer writes to a payments company: a settlement is missing. A chatbot would explain what settlements are. An agent looks up the merchant, finds the settlement batch, sees that the bank returned it because the account number changed, checks the merchant's KYC file for the new account, sees it was verified last week, re-submits the settlement and tells the customer when it will arrive — logging every step. Each of those actions is a tool the agent was given, with limits: it may re-submit a settlement below €50,000 without approval, above that it drafts the action for a human.

That last sentence is the product. The model is rented; the tools, the limits, the logs and the evaluation of whether the agent did the right thing are what the fintech builds, and what a buyer pays for. An agent that resolves 60% of support tickets end to end at a measured error rate below the human team's, with every action recorded, is worth the support team it replaces and the customers it keeps. One that has been demonstrated but never measured is a project.

For a regulated firm the autonomy question is the compliance question. An agent that approves a credit limit is making a regulated decision; one that moves customer money is executing a payment; one that reviews KYC is performing customer due diligence. Regulators expect the firm to be able to explain each decision, to have a human accountable for the process, and to show the controls that stop the agent doing what it should not. A buyer's diligence reads those controls the way it reads a safeguarding reconciliation.

## Example

A business-banking fintech deploys an agent for payment investigations. In six months it handles 38,000 cases: 61% resolved without a human, 34% drafted for approval, 5% escalated. The measured error rate is 0.4% against 1.1% for the human team, and every action is logged with the model version and the tools called. A buyer values the fintech at €120m and attributes €9m of that to the agent — the annualised cost of the 26 support staff it did not hire — after its own auditors sample 500 logged cases and find the controls hold.

## Common mistakes

- Giving an agent tools without limits, and discovering what "autonomous" means in a settlement file.
- Measuring the demo and not the error rate.
- No log, so neither the firm nor a buyer nor a regulator can say what the agent did.
- Treating an agent's credit or KYC decision as a suggestion when the firm acts on it as a decision.
- Buying an "agentic" company for the word rather than for a measured track record.

## Related terms

- [Foundation Model](https://mahouse.world/glossary/foundation-model.md)
- [Large Language Model](https://mahouse.world/glossary/large-language-model.md)
- [RAG](https://mahouse.world/glossary/rag.md)
- [Guardrails](https://mahouse.world/glossary/guardrails.md)
- [Coding Agents](https://mahouse.world/glossary/coding-agents.md)
- [Machine Learning](https://mahouse.world/glossary/machine-learning.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-20. https://mahouse.world/authors/ihor-vlasov