Intent Intelligence: Why Fraud Prevention Needs a New Signal for Humans and AI Agents

For years, digital fraud prevention has revolved around a deceptively simple question:

Who - or what - is this?

Is this the legitimate account holder? Is this a recognized device? Is this a human or a bot? Are the credentials valid? Is this transaction coming from a trusted location?

Those questions still matter. But they are no longer enough.

Scams and social engineering have created situations where the person passing every identity check can be the victim themselves. At the same time, AI agents are beginning to browse, build carts, access services and make payments on behalf of customers.

In both cases, knowing who is interacting with your business doesn't necessarily tell you whether what happens next is safe.

The more important question is becoming:

What do they mean to do?

That shift - from identifying an actor to understanding their intent - is the foundation of Intent Intelligence.

Identity tells you who. Intent tells you why.

Traditional fraud stacks are typically built around checkpoints.

Bot management examines traffic at the perimeter. Device intelligence evaluates a login. Behavioral biometrics analyzes activity during a session. Payment risk systems assess the transaction.

Each can provide valuable information. But the weakness is structural: evaluating individual events or sessions can miss the meaning contained in the sequence of events.

Consider an Authorized Push Payment (APP) scam.

The customer may be genuine. They may be using their normal device, from their usual location, with valid credentials. Every conventional identity signal could indicate that this is a trusted customer.

But perhaps that customer is being coached by a fraudster.

What changes isn't necessarily their identity. It's their journey: perhaps a long period of otherwise normal activity suddenly leads to a new payee, an unusual transaction or an unfamiliar path through the application.

The risk is revealed by the transition between actions.

Darwinium defines Intent Intelligence as the continuous assessment of what a human or agent is trying to do, based on how they behave and the path they take - not simply the credentials they present.

AI agents make understanding intent even more important

Agentic AI introduces another challenge to identity-centric fraud controls.

AI agents can increasingly perform actions on a customer's behalf, including searching, shopping, interacting with APIs and progressing toward payment. Protocols such as Model Context Protocol (MCP) also create new ways for agents to interact directly with business services.

That doesn't mean AI agent traffic is inherently bad. Quite the opposite: legitimate agents have the potential to become an important source of customer demand.

The problem is determining which activity to trust.

An agent might be undeclared. A properly authenticated agent could be hijacked or manipulated through prompt injection. An agent might possess legitimate authorization but subsequently operate outside the mandate the customer intended. And malicious automation can potentially use similar interfaces to legitimate agents.

The whitepaper highlights the scale of this visibility problem: within Darwinium's network, only around one in four agentic transactions self-declares, while transactions involving agents are rejected nine times as often as other purchases.

Simply identifying an AI agent, therefore, solves only part of the problem.

Businesses also need to understand what that agent is doing.

What does Intent Intelligence actually require?

Understanding intent isn't about adding another point solution or another risk score at checkout.

It requires visibility and context across the journey.

Darwinium's Intent Intelligence framework identifies four fundamental requirements.

1. Capture every signal needed for a decision. Intent is inferred from context. That means bringing together device, behavior, identity and journey signals rather than evaluating each in isolation.

2. Join every surface into one journey. A mobile login, web account change, API interaction or an AI agent's MCP tool call may all form part of the same underlying customer journey. If they sit in separate security tools, the transitions between them can disappear.

3. Score the sequence, not simply the tranasction. The important question isn't just whether a session appears risky overall. It's whether this particular step, following that particular step, at this speed, makes sense.

4. Act when risk appears. Intelligence that arrives after a fraudulent payment has completed is useful for investigation, but not prevention. Decisions need to happen in-line, while there is still an opportunity to permit, verify, challenge or prevent an interaction.

Together, these capabilities move fraud detection away from isolated moments and toward understanding the shape and meaning of a digital journey.

The journey itself becomes a fraud signal

This is where Journey Transition Probability (JTP) becomes important.

Every digital business has patterns in its legitimate traffic.

Customers may browse before adding an item to their cart. Users typically authenticate before changing sensitive account information. A legitimate shopping agent might retrieve a catalog before building a cart.

JTP learns the shape of those journeys—including which steps normally follow one another, how frequently they occur and how long transitions usually take—and scores new activity against that behavior.

Darwinium's approach considers four types of journey signal: per-step probability, path probability, first-passage time and transition-rate change.

That creates another dimension of fraud intelligence.

A credential-stuffing bot, for example, may arrive at login with virtually no normal pre-authentication journey. A scraping bot may systematically enumerate content without expected engagement. A coerced customer may suddenly deviate from a historically normal journey. A hijacked AI agent may start taking actions inconsistent with its declared mandate.

Different actors. Different attacks.

But all can expose themselves through the journeys they take.

One model for humans, bots and AI agents

Historically, businesses have often separated human fraud, bot management and other forms of automated activity into different technology stacks.

Agentic AI makes those boundaries increasingly difficult to maintain.

A customer may initiate a journey, hand part of it to an AI agent and then return to complete an action themselves. An agent may interact through MCP while the eventual payment happens through a conventional web experience.

From the customer's perspective, that's one journey.

Fraud controls need to see it the same way.

Darwinium's approach is to treat humans, bots and AI agents as actors moving through digital journeys and assess their activity using common journey context. Device, network, behavioral and identity signals - including agent credentials where available - can then contribute to the same decision.

That doesn't replace identity.

It makes identity more useful by putting it into context.

Extending Intent Intelligence to MCP

MCP creates a particularly important visibility challenge.

Traditional WAF, bot-management and identity technologies were generally designed around page requests, APIs and conventional sessions. They weren't designed to understand an agent's ordered sequence of tool calls, the arguments associated with those calls and how those actions connect to a subsequent human checkout.

Darwinium's MCP Protection brings agent activity into the same edge control plane as web, mobile and API interactions.

The whitepaper describes how an agent's prompt, ordered MCP tool calls, visible payload data and subsequent human activity can be stitched into a single journey. Because Journey Transition Probability scores events throughout that journey, tool calls can be evaluated using the same intent framework applied to human interactions.

This enables a much more nuanced response than simply "agent: allow or block."

A trusted and authenticated agent behaving consistently with legitimate intent can potentially continue with minimal friction. An unverified or higher-risk agent can be subjected to additional controls before a sensitive action such as payment.

Intent complements emerging agent identity standards

There is already substantial work underway to establish identity and authorization for agentic commerce.

Standards and initiatives including Web Bot Auth, Visa Trusted Agent Protocol, Mastercard Agent Pay and Google's AP2 address important questions around whether an agent is known or authorized.

But identity and authorization aren't the same as continuous risk assessment.

A legitimate credential tells you something important about the actor. It doesn't necessarily tell you whether that actor's current behavior is legitimate, whether its journey has been compromised or whether its actions remain consistent with what the customer intended.

As the whitepaper puts it, emerging standards focus on agent identity. Darwinium layers on intent.

Why architecture matters for Intent Intelligence

Intent Intelligence isn't simply a new label for fraud scoring. Delivering it continuously depends heavily on architecture.

Darwinium runs within a customer's existing CDN, including Cloudflare, AWS CloudFront and Akamai. This edge-based architecture provides visibility from the first pre-authentication interaction through subsequent web, mobile, API and agent activity.

It also places decisioning at the point where action can still be taken.

Darwinium can combine device, network and location intelligence, behavioral biometrics, identity information and journey context. Persistent Digital Signatures help recognize returning and related activity even when individual identifiers change or are spoofed.

The result is not just another fraud checkpoint. It's continuous context.

Are you ready for Intent Intelligence?

As humans and AI agents increasingly share the digital perimeter, fraud teams should reconsider what their existing controls actually understand.

Ask whether your current stack can identify the exact transition where a trusted journey became risky. Ask whether MCP tool calls and human web activity can be connected into the same journey. Ask whether the same model can evaluate humans, bots and agents. And ask whether decisions can be made in-line, before the risky action reaches the application.

Most importantly, ask whether your controls can answer both of the questions that now matter:

Who is this? and What do they intend to do next?

Identity isn't disappearing from fraud prevention. But identity without intent is becoming an increasingly incomplete picture.

Download the Intent Intelligence white paper

The shift from identity to intent has implications for fraud, risk, trust and safety, security and digital commerce teams.

In our new whitepaper, Intent Intelligence: From Identity to Intent, Understanding Humans and AI Agents Across Every Digital Journey, we explore the changing threat landscape, the four requirements for Intent Intelligence, Journey Transition Probability, MCP Protection and the questions organizations should ask when evaluating whether their existing fraud stack is ready for agentic commerce.

Download the Intent Intelligence whitepaper to discover how Darwinium helps reveal what's real - across humans and agents, at every step.