AI agent detection is the process of identifying when an AI agent, rather than a person, is operating a digital session such as a login, account application, or payment. DataVisor dEdge detects AI agents by analyzing how a session is driven, not just what device it runs on. It scores keystroke timing, mouse paths, click placement, and press timing against how people physically use a keyboard and mouse, and it recognizes known AI agent environments. The result is an AI agent score from 0 to 100, plus attribution to the agent behind the session, delivered at the moment of decision.
AI agents have moved from answering questions to taking action. Browser-based agents can now sign in to accounts, complete applications, fill out forms, and initiate payments on a person's behalf. For banks, credit unions, and fintechs, that opens a new question at every high-value event: is a person acting here, or an agent acting for them?
The answer matters for more than risk. Some agent activity is legitimate and customer-approved. Some is a fraudster scaling account takeover or application fraud with automation that looks clean. Financial institutions need visibility first, so they can decide which policy fits each case. This article explains the trends behind the shift, how AI agent detection in dEdge works, and how fraud teams can put its signals to work in real time.
Why AI Agents Look Different From Traditional Bots
Traditional bot detection was built for scripts and headless browsers. Those tools leave obvious marks: automation flags like WebDriver, unusual browser builds, emulators, or device fingerprints that don't add up.
Today's AI browser agents are different. Products such as Claude for Chrome, ChatGPT's agent mode, and Meta's Muse operate through real browsers. In many cases:
- The browser is a genuine Chrome instance
- Automation flags are turned off
- The device fingerprint is clean and consistent
- Some run in their own cloud browser and leave no trace on the customer's device
A device check alone sees a trusted environment. What sets an AI agent apart is the way it drives the page. An agent types, clicks, and navigates with a precision and rhythm that a human hand does not produce. Keystrokes arrive faster than physical input allows. Clicks land on exact computed coordinates. The pointer reaches a button without traveling a natural path, or clicks happen with no mouse movement at all.
Those differences are measurable. dEdge measures them.
The Trends Driving AI Agent Detection
Trend 1: AI Agents Are Becoming a New Kind of Customer
Agentic AI is moving quickly from pilot to everyday use. Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024.
Payments networks are building for that future. In April 2025, Visa introduced Intelligent Commerce, with partners including Anthropic, Microsoft, OpenAI, and Stripe. Visa's announcement says people will soon have AI agents browse, select, purchase, and manage on their behalf, and that agents must be trusted by users, banks, and sellers alike.
The traffic is already arriving. The 2026 Thales Bad Bot Report describes agentic AI as a new class of automated traffic.
For banks, credit unions, and fintechs, the opportunity is clear. Customers will want to delegate routine tasks to agents they trust. Institutions that can recognize agent sessions will be the ones ready to support that convenience safely.
Trend 2: Automation Now Outpaces Human Traffic Online
The same capabilities that help customers also help fraudsters scale. The Imperva 2025 Bad Bot Report found that automated traffic reached 51% of all web traffic, surpassing human activity for the first time in a decade. Bad bots alone made up 37%. The report notes that AI has lowered the barrier to entry for attackers.
The financial impact is growing too. Deloitte projects that fraud losses enabled by generative AI in the US could reach $40 billion by 2027, up from $12.3 billion in 2023, a compound annual growth rate of 32%.
Regulators are paying attention. In November 2024, FinCEN issued an alert on fraud schemes using generative AI and deepfake media to target financial institutions. Among its red flags: a customer's device data that doesn't match their identity documents, and customers who decline multifactor authentication.
The takeaway for fraud teams is encouraging. Device and behavioral signals capture how a session is actually driven, which generated content alone doesn't change. That makes them one of the most durable layers in a modern fraud strategy.
How DataVisor dEdge Detects AI Agents
dEdge answers three questions at every high-value moment, from login to account opening to payment:
- Is an AI agent driving this session? An AI agent score from 0 to 100 measures how strongly the session's behavior looks like an agent at the controls.
- Which agent is it? Vendor recognition identifies known AI agent products active on the page.
- Is the session running inside an AI agent environment? Environment recognition, built into dEdge device intelligence, flags pages opened inside a known agent product.
Together, these signals give fraud teams both detection and attribution.
Attribution isn't proof. A person can browse by hand inside an AI-enabled browser. That's why vendor recognition informs the score but never decides it on its own.
No trace doesn't mean human. Some agents run in their own cloud browser and leave no product trace on the page. dEdge catches them through behavior.
How the AI Agent Score Works
dEdge already observes how people interact with a page. The AI agent score compares each session with how a person physically uses a keyboard and mouse. Agents give themselves away in ways like these:
- Typing at a pace or regularity no person produces
- Keystrokes arriving faster than physical input allows
- The pointer reaching a button without a natural path, or clicks with no mouse movement at all
- Clicks landing on precisely computed points rather than where a hand would land
- Clicks happening while the browser tab isn't even visible
- A known agent product's interface on screen while actions happen
The score is designed for precision:
- Agreement builds confidence. The score grows as more independent rules agree. No single behavioral rule can push a session into the confident range. Only direct evidence, an agent's own interface on the page, can do that alone.
- Short sessions can't score high by chance. Each rule needs a minimum amount of activity before it counts.
- Mobile users are protected. Rules that don't apply to a device type are switched off. A finger tap on a touchscreen is never mistaken for automation.
- Privacy by design. The rules use only the timing and geometry of input and the state of the page. They never read what was typed.
- Calibrated against real traffic. Rules are calibrated so a person can't reach 60 from behavior alone. The 60 threshold is a convention, and teams can set their own cutoff.
DataVisor tuned these rules on real sessions from leading AI agents, including Claude for Chrome, Claude Desktop's built-in browser, ChatGPT's extension and desktop browser, and Meta's Muse. Human baselines came from desktop Chrome, Android and iOS browsers and WebViews, and a customer's production login page. Agent sessions landed in the confident range. Human baselines stayed well below it.
In testing, agent sessions from Claude and ChatGPT browser extensions scored between 85 and 100, whether the agent worked in the active tab or a background tab. A real person, including one with erratic input, scored 0.
Built on dEdge Device and Behavioral Intelligence
AI agent detection is the newest layer in dEdge, which already analyzes more than 100 device, browser, network, integrity, automation, and behavioral signals across Android, iOS, and web. The new signals inherit that full context:
- Device risk profile: Device, browser, OS, hardware, and graphics attributes, plus integrity flags such as incognito mode, WebDriver, and suspicious browser or OS changes
- Behavioral signals: Typing speed, paste and autofill use, mouse movement, click timing, page duration, navigation, and session switching
- Automation and tampering detection: Bots, browser automation, emulators, cloud phones, app cloning, hooking, and tampering
- Network intelligence: Persistent device IDs that link devices, accounts, identities, and IPs to reveal shared infrastructure and coordinated activity
dEdge generates 99% consistent and unique device IDs across Android, iOS, and web, even when attackers reset or manipulate devices. Its edge computing filters out 80%+ of evident bot traffic before it reaches the backend, keeping the experience fast for real customers. And dEdge doesn't collect personally identifiable information.
Putting AI Agent Signals to Work in Real-Time Decisioning
Detection is only valuable when it shapes a decision. dEdge signals flow directly into the DataVisor platform, where they're evaluated alongside rules, supervised and unsupervised machine learning, identity, transaction, and Knowledge Graph data. Real-time decisioning runs at 15,000 QPS with 30 ms latency, so agent signals inform the outcome before the event completes.
Because the score is a risk feature, not a verdict, teams can set policies that match their risk appetite and customer experience goals:
- Account takeover and credential stuffing: Combine a high agent score with a new device, unfamiliar IP, or recent credential change to trigger step-up authentication at login.
- Account opening and application fraud: Use agent signals with identity and velocity data to spot applications completed by automation at scale, then route them for review.
- Payments and transfers: Add a verification step when an agent initiates a payment to a new payee, while letting low-risk, recurring activity flow.
- Fraud ring and mule detection: Link agent-driven sessions through persistent device IDs and graph analysis to surface coordinated activity across accounts.
- Customer-approved agents: Use vendor and environment attribution to recognize legitimate agent use and design a smoother path for it.
Example: one money transfer, three outcomes. The DataVisor decision engine weighs device trust, the AI agent score, and the transfer's amount and payee history together:
The mid-range score is useful too. A score between 1 and 59 adds context to models and rules without forcing a hard decision, which keeps false positives low for real customers.
When it's time to write or tune those policies, Vera, our AI agent, can help fraud teams create and refine strategies that put the new signals to work.
Lead With Visibility Into Every Session
AI agents will play a growing role in how customers bank, borrow, and pay. Financial institutions that can see agent activity clearly will be ready to welcome the legitimate use cases and stop the fraudulent ones, without adding friction for the people they serve.
dEdge gives fraud and risk teams that visibility at the moments that matter: a clear AI agent score, vendor and environment attribution, and the full context of more than 100 device and behavioral signals, all feeding real-time decisions on the DataVisor platform.
See how dEdge detects AI agents in your own traffic. Request a demo.
Frequently Asked Questions
What is AI agent detection?
AI agent detection identifies when an AI agent, rather than a person, is operating a digital session. It analyzes input behavior and the page environment to tell agent-driven activity apart from human activity at high-value events like login, account opening, and payment, so financial institutions can apply the right policy to each.
How does DataVisor dEdge detect AI agents?
dEdge compares how a session is driven with how people physically type and click. It checks typing rhythm, keystroke speed, mouse paths, click placement, press timing, and clicks in hidden tabs, and it recognizes known AI agent environments. These signals combine into an AI agent score from 0 to 100, plus the agent identified.
Can traditional bot detection catch AI agents?
Not reliably. AI browser agents often run in real Chrome browsers with automation flags off and clean device fingerprints, so device checks see a trusted environment. Behavioral analysis of how the page is driven, combined with recognition of known agent environments, is what separates AI agents from people.
What does a high AI agent score mean?
A score of 60 to 100 indicates strong evidence that an AI agent or similar automation drove the session. Scores of 1 to 59 show some automation-like signals and work best combined with other risk data. A score of 0 means no automation-like behavior was observed.
Will dEdge flag real customers as AI agents?
The rules are calibrated so a person can't reach a score of 60 from behavior alone. Rules that don't apply to touch devices are switched off, short sessions can't score high by chance, and in testing a real person with erratic input scored 0.
Does dEdge read what users type?
No. The AI agent rules use only the timing and geometry of input and the state of the page. They never read typed content, and dEdge doesn't collect personally identifiable information.
Which AI agents can dEdge detect?
dEdge was validated on sessions from Claude for Chrome, Claude Desktop's built-in browser, ChatGPT's extension and desktop browser, and Meta's Muse. Agents that leave no product trace on the page are still caught by the behavioral score.
