An agent booking a hotel or managing your finances isn't a human impersonator. It's software legitimately acting on someone's behalf. Yet traditional authentication systems weren't designed to distinguish an authorized agent from an unauthorized one.

"Identity shifts from who you are to what you've been authorized to do," Anurag explained.

That distinction changes how product managers should think about AI agent security.

Rather than simply verifying identity at login, businesses must determine what an agent can access, which actions it can perform, and when those permissions should expire.

Consider an AI assistant helping someone shop for home insurance. The customer might authorize it to compare policies, contact insurers, and summarize coverage options.

But purchasing a policy?

That's another decision entirely.

Anurag recently used an AI agent to research insurance options for his own home. He was comfortable delegating the research, but wanted to make the final purchasing decision himself.

The lesson for product teams is straightforward: Permission to help isn't necessarily permission to act.

Security shouldn't feel like airport screening

Anyone who's struggled with a six-digit verification code knows that security can make even simple digital experiences frustrating.

Anurag compared visiting a website to passing through a TSA checkpoint. Customers repeatedly prove their identity before they're allowed to accomplish anything.

The irony is that these security measures often create friction without necessarily improving protection.

A better approach, he argues, is continuous authentication: evaluating trust throughout an interaction rather than concentrating security checks at the beginning.

For routine, low-risk activities, authentication could happen quietly in the background using contextual signals such as device information, location, and behavioral patterns.

When something unusual happens, the system could request additional verification.

Think of it as a car's safety systems. Drivers don't need to repeatedly demonstrate their ability to operate the vehicle. But when conditions change, the technology responds.

For product managers, the opportunity is to make security more responsive to customer intent.

An AI agent reviewing flight options shouldn't face the same authorization requirements as one spending $2,000 on tickets.

And customers shouldn't have to navigate complicated settings to stop an agent from doing something unexpected.

"Security needs to shift from being this annoying gatekeeper at the front door to more of a safety net that protects them in real time without getting in their way," Anurag said.

The fine line between helpful and creepy

Product teams love frictionless experiences.

Customers, however, don't always share that enthusiasm.

Anurag described an experience at Twilio involving Silent Network Authentication, a technology that verifies users through mobile carriers without requiring traditional SMS codes.

Technically, it offered a more secure, seamless experience.

But early user research revealed something surprising.

Customers perceived the invisible authentication process as less secure because they couldn't see anything happening.

Twilio addressed the problem by adding a brief message explaining that secure authentication was taking place in the background.

A small design change helped customers understand and trust the experience.

"The transparency, you know, builds trust when the technology is working," Anurag explained.

This matters enormously for AI-powered customer experiences.

Product managers often measure success through task completion, efficiency, and reduced friction. But those metrics don't necessarily reveal whether customers feel comfortable with what the technology is doing.

An AI agent might complete a task flawlessly while leaving its user unsettled.

The experience succeeded technically.

It failed emotionally.

That's why customer trust in AI needs to be evaluated alongside functional performance.

Product managers need to test the boundaries of delegation

Here's where UX research becomes especially important.

Product teams cannot assume customers want maximum automation simply because the technology makes it possible.

Anurag identified three areas product teams should investigate: which tasks customers are comfortable delegating, how comfort varies among different users, and what controls people need when something goes wrong.

These questions require more than usability testing.

They require understanding how people make decisions, perceive risk, and respond when technology acts independently.

For example, a customer might happily let an AI agent reschedule meetings but insist on approving financial transactions.

Another might trust an agent to book travel within a spending limit but want confirmation before canceling an existing reservation.

Neither preference is irrational.

Trust depends on context.

Product managers and UX researchers should test these boundaries early, ideally before investing heavily in development.

A practical research agenda might include:

• Delegation: Which tasks will customers hand over completely, and which require approval?
• Transparency: What information helps customers understand an agent's actions without overwhelming them?
• Control: How quickly can customers interrupt, reverse, or revoke an agent's permissions?
• Confidence: Does the experience make customers feel safer, or merely reduce the number of clicks?

Prototype testing and customer interviews can reveal concerns that product analytics alone won't capture.

Teams should also observe how preferences change across customer segments and risk levels.

The goal isn't to eliminate every moment of friction. It's to understand which moments build confidence.

Trust is becoming a product strategy

There's a broader business implication here.

AI is accelerating how quickly companies can build products. But building faster doesn't guarantee customers will adopt what gets shipped.

An agentic experience that saves time but creates uncertainty may struggle to earn repeat usage.

For product leaders, that means treating security, UX, and customer research as connected responsibilities rather than separate workstreams.

It also changes how success should be measured.

Beyond task completion and conversion rates, teams should examine whether customers understand agent permissions, feel confident delegating tasks, and continue using the experience after encountering unexpected behavior.

Those signals can help explain adoption, retention, and ultimately the return on AI investments.

Anurag believes digital identity will increasingly become a continuous, contextual trust layer, supported by verifiable credentials that establish exactly what an AI agent is authorized to do.

It's an ambitious vision. But the immediate product challenge is much more practical.

Talk to customers. Test their assumptions about AI. Observe where automation creates confidence and where it undermines trust.

And build controls that ordinary people can understand.

Because the next generation of successful AI products won't necessarily be the ones that automate the most tasks.

They'll be the ones customers trust enough to use.

As Anurag put it, "Security is no longer just about lock on the front door. It's the stating wheel and the brakes that give customers the confidence that to let AI drive in the first place."

Additional resources

These resources complement the conversation with Twilio's Anurag Dodeja, particularly around AI agent security, customer trust, human oversight, and the role of UX research in designing AI-powered experiences.

The selections prioritize practical insights for product managers and UX researchers who want to understand how customers interact with AI and how to build experiences people trust.

• Webinar: Beyond technical evals: How to evaluate AI for real-world success — Learn how to evaluate AI experiences beyond technical accuracy by measuring customer trust, understanding, and the quality of human-AI interactions

• Guide: How to be sure your AI experience is excellent: Introducing ARQ™ — Explore a framework for evaluating AI-powered experiences across five dimensions: trust, control, understanding, outcomes, and affinity.

• Podcast: Insights Unlocked: UX research for AI: building trust in experiences — Microsoft senior UX researcher Priyanka Kuvalekar explains why traditional usability metrics aren't enough to understand how customers experience and trust AI.

• Blog post: Beyond accuracy: how to evaluate the full AI experience — Discover why an AI agent can complete a task successfully yet still leave customers frustrated or uncomfortable, and how product teams can measure what conventional performance metrics miss.