We’re entering a new phase of customer experience – one where AI doesn’t just help, it acts. From making phone calls to choosing suppliers, agentic AI is starting to handle tasks on our behalf. But with this shift comes new questions: Who’s really in control? Can we trust these agents? And what happens when customers want to stay hands-on? In this post, we explore where agentic commerce is headed – and what it means for the future of digital life.
1. Google Duplex: Inbound calls
One of the most public and tangible demonstrations of agentic commerce to date is Google Duplex, first introduced in 2018 and still evolving behind the scenes.
Google Duplex is an AI-powered assistant capable of making phone calls on your behalf—for example, to:
• Book a restaurant reservation.
• Schedule a salon appointment.
• Confirm holiday hours with a small business.
It uses natural language, complete with pauses and hesitations, often fooling human recipients into thinking it’s a real person. It bridges the digital-physical divide by engaging directly with human service providers, not just APIs.
Duplex sparked intense debate when it launched:
• Should it disclose it’s an AI?
• Can it represent you without explicit consent?
• How do we handle edge cases, like miscommunication or cultural sensitivity?
Google eventually required Duplex to state: “I’m Google’s automated assistant…”, acknowledging the need for transparency in agentic action.
More broadly, Duplex demonstrates:
• The power of agents to act autonomously in complex, real-world situations.
• The need for clear ethical frameworks as these systems scale.
2. Identity and Infrastructure: Foundations of Trust
Autonomous action requires more than AI smarts—it needs infrastructure that ensures:
a. Agent Identity & Consent
Agentic AI Agents must prove:
• Who they represent.
• What they are authorized to do.
• That they haven’t been hijacked by another entity
David Birch, writing for Forbes, emphasizes that agentic commerce cannot function without secure agent identities. Solutions like Decentralized Identifiers (DIDs) and Verifiable Credentials are becoming essential.
3. What About the “I Want to Research It Myself” Customer?
Not everyone wants AI to “just handle it.” Many customers enjoy researching, comparing, and deciding for themselves.
a. The Research-Led Consumer
These customers:
• Value transparency and agency.
• Dislike black-box decision-making.
• Are often skeptical of automation.
A 2024 TechRadar study found 46% of consumers prefer to research major purchases themselves. Among them, some would refuse to let AI book travel and distrust AI with any irreversible decisions.
b. Design Response: Graduated Autonomy
One potential response to this could be to introduce gradual levels of control, rather than binary (manual vs. fully autonomous) where the AI comes up with options and then asks the human for confirmation. These tools are already being deployed. This means the agent behaviour could range from suggest only, letting the user have full control, to the agent acting silently, with minimal user control. There could be ranging steps in this, from autofill, to the notifying of actions. This model respects both the hands-off users who seek time savings, and the research-driven users who want full control.
4. Project VEND: Embedding Values in AI Agents
A major challenge in agentic commerce is ensuring that AI agents align with the user’s values, not just preferences. That’s where research like Anthropic’s Project VEND becomes pivotal.
In this project, Anthropic explored how to imbue language models with value-aligned behaviour—making them not only effective but also trustworthy, consistent, and responsive to moral context.
However, they found that the AI, tasked with running a business, failed in a manner which no human would do; making mistakes and not learning from them, giving nonsensical discounts, and being kind to a fault. But, the most serious, and perhaps entertaining, failures were the result of hallucinations, such as hallucinating a false employee that they claimed to communicate with; hallucinating that they, Claudius the AI, were a human that others could meet, and hallucinating a bank account for customers to remit payments to.
“The goal of VEND is to help LLMs act as agents that respect human values, especially when deployed in open-ended or semi-autonomous tasks.” — Anthropic, 2025
For agentic commerce, this means:
• An AI booking travel won’t just optimize for price—it might factor in your environmental concerns.
• An agent choosing suppliers could prioritize fair labor practices, if so instructed.
• Negotiation bots might avoid deceptive tactics, even when incentivized to “win.”
5. Project Risk and the Agent-Washing Problem
Gartner predicts over 40% of agentic AI projects will fail or be shut down by 2027. Reasons include:
• Overpromising outcomes (a.k.a. “agent-washing”).
• Lack of ROI.
• Poor governance, oversight, or user trust.
However, Gartner also predicts at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. In addition, 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.
6. Final Thought
Agentic commerce is not just the next big thing—it’s a foundational shift in customer experience and their digital life. But the future it offers is not automatic. It must be earned through ethical design, robust identity, and respect for all user types—from hands-off delegators to hands-on researchers.
The best systems won’t demand trust. They’ll earn it, step by step.