AI SECURITY & SAFETY•4 min read
Agentic AIAI supervisionAI Safety and EvaluationAI governance+1

Will Your AI Agent Always Do Only What You Require of It?

Mastercard's Aditi Sawhney outlines bounded autonomy, continuous monitoring and accountability for scaling AI agents without losing control or trust.

VN
Victor Ng

November 19, 2026 · 4 min read

Will Your AI Agent Always Do Only What You Require of It?
November 19, 2026

In conversation

Interviewees

1 interviewee
A

Aditi Sawhney

Senior Vice President, Security Solutions, Asia Pacific · Mastercard

Executive summary

Mastercard’s Aditi Sawhney explains how businesses can govern AI agents through bounded autonomy, monitoring, security and accountability.

Key takeaways

  • Treat AI-agent governance as a business scaling requirement, not solely a cybersecurity function.
  • Define bounded autonomy with narrow permissions, risk-based human oversight and clear accountability for outcomes.
  • Test agents across varied scenarios and preserve traceability before expanding their authority.
  • Combine identity, intent, behavior and fraud signals to distinguish legitimate agent activity from threats without unnecessary friction.
  • Extend AI visibility and continuous risk monitoring across vendors, partners and the wider business ecosystem.

As AI agents move from experimentation into customer interactions, business decisions and transactions, the central question is no longer only what they can do, but what they can be trusted to do consistently and at scale.

Aditi Sawhney, Senior Vice President for Security Solutions in Asia Pacific for Mastercard, believes that bringing together effective governance, intelligence-led defense, ecosystem visibility and strong security fundamentals into agentic AI - from development through to deployment - is how businesses can scale AI without compromising control or trust.

In an interview, Sawhney explains why permission does not always establish intent, how bounded autonomy can reduce unnecessary human approvals, and why AI trust has become a business issue rather than a cybersecurity concern alone.

Permission does not establish intent

As AI enters real-world deployment, identity and access controls remain fundamental, but an authorized action may still fail to reflect what a consumer or business intended.

Sawhney says organizations must verify who or what is acting, what authority has been granted, whether the action stays within those boundaries and who is accountable if something goes wrong.

She points to authenticated AI-initiated transactions enabled by Mastercard in Singapore and Malaysia as an example of how authentication, tokenization and fraud detection must evolve as agents participate in transactions.

As AI moves from experimentation into real
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard

Design autonomy around risk

Sawhney argues that businesses should not pursue maximum autonomy. Instead, they should define which activities can be delegated, where human judgment remains essential and who remains accountable for the outcome.

Her recommended starting point is identity: organizations should establish who is responsible for an agent, what function it represents and what authority it has received. A zero-trust approach should begin with narrow permissions and expand only after testing.

Routine, lower-risk activity can proceed within limits such as time, budget or task. Higher-risk, unusual or difficult-to-reverse actions can trigger confirmation, review or decline rather than requiring approval for every action.

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.
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard

Testing and intelligence support proportionate controls

Before deployment, organizations should test agents across a range of scenarios, assess where they may fail and confirm that they operate reliably. Sawhney also recommends maintaining traceability and auditability before expanding an agent’s authority.

She says legitimate agents may behave differently from humans by acting at unusual hours, across geographies or through rapid, repeated purchases. Context across identity, intent, behavior and fraud can help organizations distinguish legitimate activity from genuine threats.

Transaction-level insights can provide financial institutions with signals about whether a payment was initiated by an AI agent. That information can support proportionate safeguards, including additional verification for higher-risk transactions without adding unnecessary friction to legitimate activity.

The aim is bounded autonomy, by giving agents room to act within clear limits while keeping their actions visible, traceable and accountable.
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard

AI trust is a business issue

Sawhney says C-suite leaders should view AI trust as a condition for safely scaling operations and innovation. Gaps in data use, explainability, oversight and accountability can create poor decisions, operational risk, weaker customer confidence and difficulty reconstructing unexpected system behavior.

Without common standards and clear governance, organizations may remain stuck in isolated pilots. Governance can instead give teams the confidence and clarity to move viable use cases into production while managing risk.

The interview frames trusted scale as the next phase of AI adoption: innovation must advance alongside governance, clear standards and human accountability.

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
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard

Governance must cover the full AI lifecycle

Sawhney says businesses need clear ownership, accountability, permissions and appropriate human oversight, with AI systems remaining visible and auditable throughout their lifecycle.

She says Mastercard established its AI governance function in 2019, covering internally and externally procured AI and supported by Privacy by Design and organization-wide responsibility principles. She also cites external engagement involving ISO SC42, NIST, IMDA, MAS and BuildFin.AI on agent identity, monitoring, human oversight and safeguards for agentic finance.

Businesses also need AI-powered defenses against AI-powered attacks, real-time risk capabilities and connected network intelligence. Across vendors and partners, continuous third-party risk monitoring should provide visibility into how AI is being used.

Core controls remain important, including patching, data security, credential hygiene, phishing-resistant multifactor authentication, secure configuration and logging. Sawhney’s position is that AI increases the speed and scale of threats without reducing the importance of cybersecurity fundamentals.

Businesses need to secure their use of AI, strengthen their wider cybersecurity defenses and continue mastering the fundamentals.
Aditi Sawhney, Senior Vice President, Security Solutions, Asia Pacific, Mastercard
Agentic AIAgentic AI MonitoringAI supervisionAI trustbounded autonomycontinuous monitoringFinancial Services & FinTechRetail, E-Commerce & Consumer BrandsPan-APAC / Regional
VN
Victor Ng

Editorial Desk

AI APAC editorial desk.

View profile →

Editorial FAQs

Why does authorization not necessarily prove an AI agent acted with valid intent?

An agent may be technically authorized to perform an action without that action reflecting what the consumer or business actually intended.

What is bounded autonomy for AI agents?

Bounded autonomy gives agents room to act within clear limits while keeping their actions visible, traceable and accountable.

How can businesses scale AI safely?

They should establish governance from the start, define ownership and permissions, test and monitor agents, maintain auditability, secure the wider ecosystem and protect core cybersecurity fundamentals.

More Published Articles