Plaid has launched Agent Detection, a tool showing financial firms whether an interaction is human or performed by an AI agent.
The product is aimed at fintechs, banks, and other financial companies. Beyond flagging whether a session is driven by a person or an AI agent, it also identifies which AI provider the agent is associated with. Plaid describes the release as the first step in a broader effort to establish trusted agent interactions across the financial ecosystem.
The launch follows the arrival of several consumer-facing personal AI agents within weeks of each other, including Grok Bot, Meta Muse, OpenAI dots, and Claude Duet. Each can operate inside a browser to complete tasks on behalf of its user. According to Pedro Sanzovo, Head of Fraud at Plaid, these agents are already appearing on login pages and in onboarding and account-linking flows, making it harder for companies to establish who is acting in a given session.
How the detection model works
Because AI agents can navigate websites in ways that resemble human behaviour, conventional device and behavioural checks may struggle to separate the two. Agent Detection analyses thousands of signals spanning devices, behaviour, automation patterns, and network activity to classify each interaction.
The model has been tuned specifically for financial use cases. Plaid already runs device, behaviour, and network intelligence across hundreds of institutions for financial logins, which the company says allows it to distinguish between an agent logging in as a bank customer and one carrying out a shopping task. In side-by-side tests on live traffic, Agent Detection identified 2.3 times more agent sessions than competing solutions, with no increase in false positives.
In a practical scenario, a consumer could instruct an agent to pay an upcoming bill. The financial institution would then be able to recognise that an agent is behind the interaction and use that information to determine how it supports agent-initiated payments.
Extending consumer-permissioned access
Plaid already provides consumer-permissioned access to financial data for thousands of apps, including AI assistants such as ChatGPT, Claude, Perplexity, Grok, Meta Muse, and Google Gemini. Under this model, consumers authorise access without sharing bank credentials with the app and can see what data is shared and why. Access is read-only, runs through existing bank-approved APIs, and is subject to the same data use rules that apply to all Plaid customers.
Agents that act on a consumer’s behalf, rather than simply retrieving information, raise additional questions around transparency, authorisation, and control. Plaid intends to work with financial institutions and agent developers to extend the principles of consumer-permissioned access into a sanctioned route for AI agents, with security, visibility, and permission as core requirements.
Two approaches to agent traffic
Plaid expects financial companies to take different positions on how they support AI agents and is developing capabilities for each of them. For firms that want a dedicated channel for agents, the company is building on its consumer-permissioned access infrastructure towards a mechanism that lets consumers approve specific actions. The design aims to give consumers visibility and control over what they authorise, while giving institutions confidence in the interactions they accept.
For firms already seeing agents interact with their existing digital channels, Plaid is expanding its Trust Index, which currently scores sessions for risk, to assess agent-driven activity in context. Drawing on intelligence from across its network, the company aims to help these firms separate legitimate agent activity from potential risks and decide how to handle each interaction.