AI agents are smart – they just have no idea what you did before you opened the chat 

A human at the airline check-in counter will confirm your destination even if your boarding pass is in their hand. A hotel associate will ask for your name and number of nights you’re staying at the property even when they have your reservation in front of them. These are rituals that travellers accept without much thought. 

But when an AI agent asks the same questions, people leave. 

New research from Conviva, which sells technology designed to fix the problem it describes, found that consumer-facing AI agents spend an average of two minutes and 32 seconds gathering basic context before they can address a customer’s actual request. This includes pages visited, confirmation numbers and more. The setup loop drives 8.5 per cent of shoppers to abandon the session entirely, before the agent has provided any help at all. 

The gap between those two reactions – patience with a human, abandonment for a bot – points to something the AI industry has been slow to reckon with. Customers were not promised an electronic version of a legacy experience. They were promised an AI super-brain tailored to their unique wants and needs. 

According to the Zendesk CXtrends26 report, 83 per cent of consumers believe customer experience should be far better than it is today, and 67 per cent believe brands should offer more personalised service because of AI. 

The frustration is not, industry analysts are quick to note, a failure of the underlying models. The large language models powering consumer AI agents have become capable at reasoning, nuance and generating relevant responses. The problem is what those models are given to work with. 

“Most consumer-facing AI agents can only see information from the current conversation,” says Keith Zubchevich, president and CEO of Conviva. “They can’t see what the customer did before the chat opened, their unique patterns, or their buying behaviour.”  

That information exists. It simply lives in separate systems, such as booking databases and loyalty platforms, that most AI agents were never built to access. The result is that every chat opens with the agent greeting a stranger. The shopper who just spent 15 minutes failing to process a return, browsed the refund policy twice and clicked through three dead-end menus before opening the chat is, to the agent, indistinguishable from someone who arrived with no prior history. 

The harder question is whether the technology was ever going to clear the bar that its own proponents set. Brands accelerated AI deployments across customer service through 2023 and 2024, often positioning the technology as a major improvement in personalisation. The pitch was not “as good as a human”. It was faster, always available and drawing on everything the brand already knows about you. 

Gartner, in a March 2026 report on agentic AI architecture, identified the missing piece as a dedicated “context layer” – a system that organises behavioural signals, real-time data and prior history into something the agent can use before a conversation begins. Without it, agents default to asking. The report concluded the layer “cannot be bought off the shelf; it must be engineered to fit your organisation’s unique needs” – a significant complication for companies that have framed AI customer service primarily as a cost-reduction exercise. 

Conviva's research, which the company says is based on full-census client-side telemetry across e-commerce and travel platforms, argues the two minutes and 32 seconds of setup time is not an inherent feature of AI agents but a symptom of that architectural gap. It found 30 per cent of sessions failed outright because the agent misread the customer’s intent from the opening message – a misread that behavioural context would in theory correct. 

Whether or not Conviva’s specific solution emerges as the industry standard, the underlying dynamic it identifies is widely corroborated. The agents are not unintelligent. They simply have no institutional memory – no way of knowing that the person typing “I need help with my order” has already been trying to solve this problem for the better part of an hour. 

A human agent asking for your booking number is following a script. An AI doing the same thing is, in the customer's mind, failing a test it was supposed to have already passed. 

Conviva’s full research report is available at conviva.ai. 

Gartner®, “The 3 Core Components of the Context Layer for AI Agents”, Andrés García-Rodeja, Michael Gonzales et al, 10 March 2026. GARTNER is a registered trademark and service mark of Gartner, Inc and/or its affiliates in the US and internationally and is used herein with permission. All rights reserved. Gartner does not endorse any vendor, product or service depicted in its research publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organisation and should not be construed as statements of fact.  

Header image credit: Conviva