What happens when an authorized agent does something its user never intended for it to do? We’re already seeing the risk emerge – and increasingly so.
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard, believes that bringing together effective governance, intelligence-led defense, ecosystem visibility and strong security fundamentals is how businesses can scale AI without compromising control or trust.
We find out more from her this interview.
What changes when AI moves from experimentation into real-world customer interactions, business decisions and transactions?
Aditi: As AI moves from experimentation into real-world deployment, the conversation shifts from what AI can do to what it can be trusted to do, consistently and at scale.
This is particularly important as AI agents begin acting on behalf of consumers and businesses. Identity and access controls remain fundamental, but permission alone does not necessarily establish intent.

Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard
An agent may be technically authorized to perform an action without that action reflecting what the consumer or business actually intended. Organizations therefore need to verify not only who or what is acting, but also what has been authorized, whether the action stays within those boundaries and who is accountable if something goes wrong.
We are already seeing this shift in practice. At Mastercard, we have enabled authenticated AI-initiated transactions in markets such as Singapore and Malaysia. As adoption grows, established safeguards such as authentication, tokenization and fraud detection must evolve to address the risks associated with AI agents. This includes ensuring transactions remain secure, traceable and consistent with consumer authorization.
As AI moves into real-world deployment, governance cannot be layered on after deployment. Security, privacy, testing, monitoring and accountability have to be built in throughout development and deployment for consumers and businesses to adopt this technology with confidence.
Businesses are reportedly moving faster on AI than their security and governance frameworks can keep up with. With the emerging gap between permission and intent, how can businesses automate more with AI agents without still relying too much on human approvals?
Aditi: The goal should not be maximum autonomy. The sweet spot is clarity about what can be delegated to an AI agent, where human judgment remains essential and who is accountable for the outcome.
Start with identity: who is responsible for the agent, what function does it represent and what authority has it been given? Then take a zero-trust approach to its actions, beginning with narrow permissions and expanding them only after testing.
Human oversight does not mean approving every action. Routine, lower-risk activity can proceed with clear limits such as time, budget or task. Higher-risk, unusual or difficult-to-reverse actions can trigger confirmation, review or decline.
Organizations should rigorously test and validate agents before deployment, assessing how they perform, where they may fail and whether they operate reliably across a range of scenarios. Before expanding an agent’s authority, they should also maintain clear traceability and auditability so that any issues can be quickly understood and addressed.
Intelligence becomes equally critical. Legitimate agents may behave differently from humans, for example by acting at unusual hours, across geographies or making rapid, repeated purchases. Organizations need context across identity, intent, behavior and fraud to distinguish legitimate activity from genuine threats without introducing unnecessary friction.
For example, transaction-level insights can provide financial institutions with additional context, including signals indicating whether a payment was initiated by an AI agent. That additional intelligence can support more proportionate risk decisions and safeguards, such as applying further verification to higher-risk transactions without introducing unnecessary friction for legitimate activity.
The aim is bounded autonomy, by giving agents room to act within clear limits while keeping their actions visible, traceable and accountable.
Why should C-suite executives care about agentic AI beyond cybersecurity? What is the business cost of getting AI trust and governance wrong?
Aditi: Leveraging AI and agentic AI will be critical not just in defending against growing fraud, scam and cybersecurity threats, but also multiple other functions and domains. AI trust is ultimately a business issue because it determines what an organization can safely scale and whether it can innovate ahead of the competition
When capability moves faster than governance, gaps emerge around data use, explainability, oversight and accountability. As AI moves deeper into business decisions and transactions, these gaps can lead to poor decisions, operational risk, weaker customer confidence and difficulty understanding what happened when a system behaves unexpectedly.
There is also an opportunity cost. Without clear governance and common standards, organizations may struggle to move beyond isolated pilots. Strong governance gives teams confidence and clarity to move viable use cases into production and scale them responsibly.
Trust is therefore a business enabler, not just a risk control. Organizations that get AI governance, data, security and resilience right can innovate faster with greater confidence, while those that do not risk slower adoption, weaker customer confidence and higher operational risk.
The next phase of AI will be defined by trusted scale, not simply the speed of adoption. Organizations that succeed will pair innovation with strong governance, clear standards and human accountability.
What do businesses need to put in place to scale AI safely and effectively?
Aditi: Businesses need to secure their use of AI, strengthen their wider cybersecurity defenses and continue mastering the fundamentals.
Governance must be built in from the start. Organizations need clear ownership, accountability, permissions and appropriate human oversight, with AI systems remaining visible and auditable throughout their lifecycle.
At Mastercard, we established our AI governance function in 2019. It operates globally across both internally and externally procured AI, supported by frameworks including Privacy by Design and organization-wide responsibility principles. Our external engagement – including ISO SC42 and work led by NIST, IMDA, MAS and BuildFin.AI – also helps advance practical approaches to agent identity, monitoring, human oversight and safeguards for agentic finance.
Businesses must also be prepared to use AI-powered defense against AI-powered attacks. As threats move at greater speed and scale, security teams need real-time response capabilities and connected network intelligence that helps them make real-time risk decisions and proactively defend against fraud and scams. They also need to prioritize the software vulnerabilities and risks most likely to be exploited, rather than simply dealing with more alerts and more data.
That visibility also needs to extend beyond the organization. As vendors and partners adopt AI, businesses need a clear and continuous view of how that technology is being used across their ecosystem, supported by continuous monitoring of third-party risk.
Finally, the fundamentals remain critical. Patching, data security, credential hygiene, phishing-resistant multifactor authentication, secure configuration and logging continue to prevent and contain breaches. AI changes the speed and scale of threats, but not the importance of getting the basics right.
