The Agent That Books Your Vacation
Imagine telling an AI agent: book a family trip to Italy, use my points, stay within budget, pick hotels we've liked before, handle all the details. Instead of returning a list of links to review, the agent simply handles it — comparing options, applying loyalty points, checking reviews against your preference history, booking hotels and flights, and presenting you with a confirmation. No research, no comparison tabs, no checkout flows.
This is the promise of agentic commerce: AI systems that don't just assist with decisions but make them. The technology to do this at a basic level already exists — large language models connected to booking APIs, calendar data, and purchase histories can execute multi-step transactions with increasing reliability. What determines whether the experience is delightful or disastrous, however, is not the model's intelligence but the quality of the information it operates on and the contextual understanding it brings to each decision.
Truth as Infrastructure
Agentic systems fail in ways that are different from traditional software. A booking engine with a bug will return an error. An AI agent operating on stale or inaccurate data will confidently complete a transaction that doesn't match what the user actually wanted — and may not flag the discrepancy at all. The agent's confidence can be inversely correlated with the user's awareness that something has gone wrong.
This dynamic makes data accuracy not just a technical requirement but a trust prerequisite. For agentic commerce to function at scale, every data source the agent interacts with — hotel availability, price feeds, product catalogs, loyalty program balances — must be accurate, current, and consistently structured. The supply-side infrastructure for agentic commerce is as important as the intelligence layer on top of it.
Enterprises building agent-ready data systems are increasingly talking about "truth and context" as the core design requirements. Truth means factual accuracy: real-time inventory, correct pricing, valid status. Context means the agent understands not just the data but its significance — that a hotel marked four stars means something different in Tokyo than in rural Bulgaria, or that a budget constraint means something different for a business trip than a honeymoon.
Context as Competitive Advantage
The contextual dimension is where agentic commerce diverges most sharply from traditional search and recommendation. A hotel comparison website shows the same results to everyone searching a given city on a given date. An agent that understands a specific user's travel history, preferred amenities, past complaints, loyalty tier status, and current trip purpose can make decisions that no general recommendation system could replicate.
This is why the companies investing most heavily in agentic commerce infrastructure are those with the deepest contextual data: airlines and hotel chains with decades of loyalty program history, banks with complete spending records, retailers with complete purchase histories. The agent's value proposition scales with the richness of the contextual data it can access.
For consumers, this creates a straightforward trust question: to delegate decisions to an AI agent, you must trust it with the data that makes those decisions good. The privacy surface of an agentic relationship is significantly larger than the privacy surface of a search session. This is not hypothetical — it is the immediate design challenge facing every company building consumer-facing agent products.
The Accountability Gap
When a human travel agent makes a booking error, accountability is clear. When an AI agent makes the same error, the accountability question is murkier. Did the model misinterpret the instruction? Was the underlying data incorrect? Did a connected API return stale availability? Was the user's stated preference inconsistent with their actual preference in ways the agent should have flagged?
The current generation of agentic products largely sidesteps this question by requiring human approval for consequential actions — the agent proposes, the human confirms. This is a sensible interim design, but it undermines much of the time savings that make agentic commerce attractive. Full autonomy requires not just technical reliability but a legal and accountability framework that hasn't yet been established.
Financial services regulators in several jurisdictions have begun engaging with the question of AI agent liability for transaction errors. The outcomes of those regulatory conversations will shape how aggressively enterprises can deploy autonomous commerce agents — and how the responsibility for agentic errors is distributed between technology providers, merchants, and consumers.
What Gets Built First
In practice, the first widely adopted agentic commerce applications are likely to be narrow rather than general: agents that handle a specific, well-defined class of transaction where the data environment is controlled and errors are reversible. Expense report filing, subscription management, recurring supply orders for businesses, travel booking within a corporate policy framework — these are all candidates for early autonomous agent deployment where the contextual complexity is manageable and the stakes of individual transactions are limited.
The general-purpose agentic assistant that can book a family vacation with full autonomy remains a more complex problem. It requires synthesizing preferences across multiple booking systems, handling edge cases, and making judgment calls that reflect personal priorities rather than policy rules. That capability is coming, but the infrastructure and trust frameworks required to deploy it at scale will take longer to build than the underlying AI technology.
This article is based on reporting by MIT Technology Review. Read the original article.
