When the Fraud System Has to Start Trusting Bots

The Inversion Problem

For more than two decades, Mastercard trained its fraud systems to recognize one category of threat above all others: automated actors attempting to transact without human involvement. The network built layered defenses, risk scoring algorithms, and pattern recognition models that flagged bot behavior as inherently suspicious. Last year, those systems evaluated 175 billion transactions, each scored in under 100 milliseconds on a scale from zero to 999 for fraud likelihood.

Now the company faces an architectural challenge that flips that entire framework. The bots are no longer trying to steal. They are the customers.

Greg Ulrich, Mastercard's chief AI and data officer, described the shift during a recent industry presentation in California. The company must rewrite risk rules that were engineered specifically to stop automated purchasing, because those same rules now block legitimate commerce conducted by AI agents on behalf of consumers and businesses. The technical debt of two decades of anti-fraud work has become a bottleneck to a new transaction model.

Fraud Detection at 175 Billion Scale

Every tap, swipe, or digital authorization on the Mastercard network triggers a real-time decision. The company generates a fraud probability score and passes it to the issuing bank within a tenth of a second. That speed requirement has held constant even as transaction volume grew and the definition of "normal" shifted.

Generative AI has already changed what those scores can detect. According to Mastercard, newer models can identify 300 to 400 percent more fraudulent transactions in high-risk bands without increasing false positives or adding friction for legitimate users. The company's Safety Net system has blocked more than 70 billion fraudulent transactions to date, and Mastercard is now building its own transformer model trained on transaction data to power future safety, security, and personalization features.

At DailyTechWire, we have tracked how large-scale payment networks handle inference latency under regulatory and commercial pressure. The Mastercard case is notable because the same infrastructure built to stop bots must now be reconfigured to serve them, without sacrificing the fraud detection that underpins trust in the network.

Services Revenue and the AI Wedge

Roughly 40 percent of Mastercard's business now comes from services rather than transaction fees alone. That includes marketing services, fraud and security offerings, and business intelligence products. One-third of that services revenue is built on AI, and those segments are growing faster than the rest of the company, Ulrich noted.

The shift matters because it repositions Mastercard not only as a payments rail but as an identity and trust layer for agentic commerce. The company is betting that as AI agents gain autonomy to transact, the limiting factor will not be agent capability but the trust infrastructure that allows humans, institutions, and merchants to delegate purchasing authority.

Five Layers for Agentic Trust

Mastercard has designed a five-layer architecture to secure transactions initiated by AI agents rather than humans. The first layer is identity, which extends beyond knowing the consumer to knowing the agent itself. Ulrich described this as "know your agent" or KYA, a framework for registering, validating, and authenticating the software acting on a user's behalf.

The second layer is verifiable intent, a cryptographically signed, tamper-proof record of the original instruction set that travels with the transaction. If a user requested black Nike sneakers in size 12 and the agent purchased a non-returnable item on final sale, the verifiable intent log provides an objective record for dispute resolution. This solves what Ulrich called the "wrong-Nikes" problem, where delegation introduces ambiguity into what was actually authorized.

Controls form the third layer, defining merchant whitelists, spending limits, and transaction constraints. The fourth layer is execution, delivered through Mastercard Agent Pay, which embeds tokenization, authentication, and acceptance protocols. The company has launched Agent Pay with Microsoft, OpenAI, and Google, among others.

Intelligence is the fifth layer. It includes updated risk rules, insight tokens that grant permissioned access to personalized recommendations, and monitoring through Recorded Future to identify threat actors operating inside the agent ecosystem.

B2B Procurement as the Bigger Play

Consumer-facing agentic commerce has dominated early headlines, but Ulrich pointed to business-to-business procurement as the larger commercial opportunity. His example was a manufacturer that runs continuous production and delegates inventory management to an agent. The agent monitors stock levels, triggers replenishment orders automatically, stays within budget, and only transacts with approved suppliers.

Enabling that use case requires the same five-layer trust stack, but with added complexity. Multiple agents must interact: a procurement agent, a supplier agent, and a banking agent, each representing different organizations with different risk tolerances and compliance obligations. That multiplies the number of parties that must trust each other and requires interoperable standards for identity, intent, and execution.

At DailyTechWire, we have observed similar multi-agent coordination challenges in supply chain automation and cross-border logistics. The difference here is that Mastercard is positioning itself as the trust broker in a system where no central authority has full visibility, and where delegation creates principal-agent problems at enterprise scale.

New Models, Familiar Motions

Mastercard participated in early access programs for Anthropic's Mythos model and OpenAI's GPT-5.5-Cyber, both designed to identify security vulnerabilities. Ulrich characterized the models as powerful new tools rather than a fundamental shift in security practice. The company's chief security officer leads a dedicated team that prioritizes critical assets, runs them through the models routinely, and tracks findings by severity level. The same technology is used to automate patching.

Mastercard has extended this approach beyond internal use and is working to make the architecture and patching capabilities available to others in the payments ecosystem. The implication is that vulnerability scanning and remediation, once manual and resource-intensive, are now being industrialized through generative models.

The Agentic Factory

After 14 months of building AI agents internally, Mastercard concluded it needed a different architecture. The company built what Ulrich described as an agentic factory, an operating system where compliance, observability, and guardrails are embedded from the start rather than added per agent. Model drift, once tracked manually by dedicated teams, is now automated into that factory.

Mastercard deployed a series of agents last year for its 4,000 consultants, covering deep research, text-to-SQL, Excel, and PowerPoint automation. Those tools did not exist at the level the company required, so it built them. Ulrich was candid about the cost of that approach. If Mastercard were starting today, it would build them fundamentally differently. The company did not anticipate, 14 months ago, that it would already be rethinking the fundamental architecture.

The lesson he emphasized was that guardrails, security, and compliance must be embedded at the front end. Trying to retrofit them after an agent is built leads to failure. Scalability and observability matter as much as the intelligence of the agent itself.

The Identity Gap

Ulrich expects the market to move next on agent identity standards. Inside Agent Pay, Mastercard authenticates the consumer using traditional e-commerce methods and binds the agent to that person. Outside that framework, he anticipates open standards will emerge to identify agents and link them to consumers, combined with verifiable intent.

Industry research suggests the gap is real. A recent enterprise survey found that only 32 percent of respondents give every agent its own scoped, managed identity, and just 12 percent include an agent-identity product in their consideration set. Mastercard has been expanding its identity capabilities organically and through acquisition for six or seven years, and that work now includes agentic identity alongside traditional know-your-business and know-your-customer frameworks.

Rewriting the Rules

The risk rules that kept bots off the Mastercard network were the product of more than 20 years of applied AI. Those rules encoded a simple premise: automated actors are adversaries. That premise no longer holds.

The rewrite is already underway on the same network that scored 175 billion transactions last year. Mastercard is not dismantling its fraud defenses. It is adding a new category to the trust model, one where delegation is legitimate, intent is verifiable, and the agent is a known entity rather than an unknown threat.

The technical challenge is significant, but the commercial and strategic stakes are higher. If agentic commerce scales, the payment networks that can authenticate agents, verify intent, and manage multi-party trust will control a new layer of the transaction stack. Mastercard is betting it can rewrite two decades of anti-bot logic fast enough to claim that position before open standards or competitors do.