When numbers come from a firm like Gartner, they carry the weight of a verdict. And the verdict for companies that have invested heavily in conversational AI is uncomfortable: customers overwhelmingly prefer ChatGPT over the in-house chatbot. Not by a small margin – according to a survey of 3,566 consumers, users are roughly three times more likely to turn to a third-party GenAI tool to solve a service problem than to the company’s own virtual assistant.

The finding upends the narrative of the past two years, in which enterprises raced to deploy proprietary AI assistants in the belief they would deliver a better, more personalized experience. Instead, the public appears to weigh perceived independence, response quality and, likely, trust on the same scale. ChatGPT has no ties to the brand, does not try to sell anything and – at least in the collective imagination – does not retain conversations to train internal models (though the technical reality is more nuanced). A company chatbot, however polished, is seen as a biased tool, and that perception erodes willingness to use it.

For those watching the space from a technical perspective, the alarm rings on the architecture and data-sovereignty front. Many businesses, eager to move fast, have adopted third-party LLM cloud APIs: the quickest route to a working chatbot. But with that choice, every customer interaction ends up on external servers, outside the company’s control perimeter. If distrust also stems from a sense of opacity, the cycle can become vicious: a chatbot built on shared cloud infrastructure risks appearing less trustworthy than a public ChatGPT, nullifying the investment.

This leads to a reflection that strikes at the core of AI-RADAR’s mission: on-premise deployment of LLMs, or in hybrid environments with granular data control, is no longer a niche exercise for extreme privacy requirements. It becomes a competitive lever to rebuild digital trust. When the entire stack – from inference to log management – stays under the company’s direct responsibility, it becomes possible to communicate data-handling practices transparently, comply with GDPR to the letter, and offer performance on par with public models, provided the hardware is correctly sized. This is no trivial choice: it demands in-house skills, investment in GPUs with enough VRAM to handle inference and fine-tuning, and a careful TCO analysis. Yet the Gartner figures suggest the ready-made cloud shortcut may not pay off in the long run.

There is another layer to read. The attachment to ChatGPT signals that users are developing a new literacy: they recognize LLM quality and demand it everywhere. A rule-based chatbot or a lightly quantized lightweight model no longer suffices. The public measures a corporate assistant against the same yardstick as the generalist model they use every day, and if the gap is too wide, they walk away. This puts pressure on companies to adopt up-to-date models, curate domain-specific fine-tuning datasets, and guarantee acceptable latency even with self-hosted setups – a technical balance that is far from trivial.

Gartner’s lesson, ultimately, is a lesson in humility for enterprise AI. Customer trust cannot be bought with a homepage widget: it requires an experience that measures up, built on solid technical foundations and governed transparently. For anyone reflecting on how to shape their AI strategy, the data is not a conclusion but an open question: is the architecture you have chosen really the best one to win preference, or are you merely driving traffic to your rival?