Why AI Agents Hallucinate: The Data Problem Nobody Is Fixing | Mario Moscatiello, Airbyte
Watch on YouTubeVideo summary
The current landscape of artificial intelligence is dominated by the rapid emergence of new models, ranging from general-purpose systems to highly specialized tools, leading many experts to believe that the challenge regarding model capabilities has been largely resolved. However, despite this abundance of advanced AI technology, enterprises are facing a significant hurdle in effectively utilizing their own internal data and context. The core issue is not a lack of sophisticated algorithms but rather an inability for agentic models to access reliable enterprise information due to poor data connectivity. This disconnect forces agents to hallucinate because they cannot verify facts against the actual business reality, often resulting in inconsistent answers where different users receive conflicting responses to the same query simply because there is no unified semantic layer connecting their disparate systems.
The root of this problem lies in a mismatch between AI readiness and data maturity within organizations. While companies with high levels of data trust are naturally better positioned to adopt AI successfully, many large enterprises still struggle because the specific data agents need to function effectively is trapped in silos or unavailable across various integrations and warehouses. When an agent attempts to solve a problem without access to this fragmented information, it fills the gaps with fabricated details rather than accurate insights. This lack of coherent context creates a major bottleneck that prevents organizations from realizing the full potential of their AI investments, as the agents are essentially operating in a vacuum where they cannot distinguish between truth and invention based on incomplete or inaccessible data sources.
To address these critical challenges, Airbyte is developing a fully agentic platform built upon six years of experience in data movement and integration. The goal of this system is to organize vast amounts of data from diverse sources into a coherent structure that makes practical sense for the business environment. By assembling context effectively before it reaches the AI agents, the platform enables these models to work smarter rather than harder by providing them with accurate, accessible information directly within their daily operational systems. This approach ensures that when an agent interacts with users or solves complex problems, it does so based on a unified and trustworthy foundation of data, thereby eliminating the need for hallucinations and ensuring consistent, reliable outcomes across different departments and use cases.
Read the full video transcript
AI or GenAI or agentic AI, it has kind
of become the only topic that we talk
here at TFA. No matter who I talk to, AI
is the only topic. And what we have
realized so far is that almost every
enterprise, they have no shortage of AI
models. Where they struggle is to
actually their own enterprise data,
enterprise context. If I ask you, based
on your interaction,
do you also feel that data connectivity
is becoming a real bottleneck for
agentic models as compared to the model
versus what are you seeing when it comes
to model versus data?
>> Um I think every week and and every
month we have new models that are coming
out and they're better and better and
there is you go from, you know, general
models to very specialized models,
whether they're closed source or open
source.
Um so I would say the model problem has
been largely solved. Um I think what
you're seeing is that if you look at
uh companies, whether they're small
companies or large companies and
enterprise and organizations, um they
sort of have, you know, two things. That
you look at
you look at data maturity and you look
at AI readiness. And yes, a lot of in a
lot of the cases, companies that are
highly data mature are usually the first
to want to be embracing AI because they
don't get they they can trust their
data. But in a lot of companies and you
know, at Airbyte we work from startups
to Fortune 500s, uh we see that even in
the case of very large enterprises, like
a lot of the data that these agents
would need to work on is actually not
available in their in the data
integrations or in their warehouses or
in their systems. And so I think data
connectivity is is what's causing a lot
of agents also to hallucinate because
they think that uh you know, they're
giving the right answers to a problem,
but there is no notion of semantic
layer. Um if 10 people in the business
were to ask an agent the same question,
they would probably get five or six
different answers. Um and so we're
really seeing that becoming um a huge
bottleneck for for companies and that's
definitely something that we're here,
you know, to help with. So.
>> Excellent. And what is Airbyte doing to
address this problem?
>> So, in a way where what we're seeing is
that because organizations have, you
know, different levels of data maturity,
what we're building is really a system
that can take a lot of data from
different sources for companies and sort
of like organize it in a way that makes
sense.
Um and then helping them companies
helping those companies put that bit
put that data like into into action.
Um put that data where where it needs to
work. So, like back into into the
systems um that that they use every day.
So, we're building a fully agentic
platform in that sense where
um we're building on top of, you know,
six years of data movement and data
integration to say, "Hey, like how do we
help companies assemble context?" Um
so that their agents can can work in a
smart way.