The Importance of Model Context Protocol

In this episode of "Full Tech Ahead," host Amanda Razani interviews Amit Sharma, CEO and Founder of CData. They discuss the critical challenge of enterprise AI: securely connecting advanced AI models to proprietary enterprise data (like CRM and accounting systems). Sharma explains that while AI models have vastly improved, the real bottleneck is providing them with the right business context. 

He introduces the Model Context Protocol (MCP) as a key solution for this. The conversation also covers the shift toward Agentic AI—which demands near-perfect accuracy since there is no human in the loop—and data infrastructure, where Sharma advocates for data virtualization (leaving data where it resides, including on-premise) rather than moving everything into a massive central warehouse. 

Ultimately, he views AI as a massive enhancer of human capital that will radically accelerate business timelines.

Key Quotes

"The real power of AI is only captured when AI can actually connect to enterprise data."
"The models aren't the issue. The issue is, how do we make the data and context available to AI?"
"If you have a case for keeping data on prem, they should keep the data on prem. We in fact favor solutions like virtualization, where you can leave the data where it is..."

Takeaways

Context is King, Not Just the Model: Stop waiting for a "better model" to fix your AI problems. Recent models are already highly advanced; the actual challenge is securely feeding them your specific enterprise data and business context.

Embrace the Model Context Protocol (MCP): To effectively connect AI to business data without massive token waste, organizations should adopt MCP, which is becoming the standard for securely structuring and governing how context is brought into AI models.

Agentic AI Requires Extreme Accuracy: When moving from conversational AI to Agentic AI (where AI takes actions autonomously), the margin for error shrinks to zero. Without a human-in-the-loop to catch mistakes, data accuracy and strict agent governance become paramount.

Virtualize, Don't Centralize: You don't necessarily need to move all your data into a massive central data warehouse to use AI. Leaving data where it naturally resides (including on-premise) and using data virtualization is often more secure, compliant with data residency rules, and highly efficient.

Find Amanda Razani on LinkedIn.  https://www.linkedin.com/in/amanda-razani-990a7233/

Follow the FTA LinkedIn Page: https://www.linkedin.com/company/full-tech-ahead/

Visit the FTA website: https://fulltechahead.com/

Check out the Substack Channel: https://fulltechahead.substack.com/

Episode Transcript

SPEAKER_00 (00:20):
Hello and welcome to Full Tech Ahead.
I'm your host, Amanda Rizzani,and with me today, I'm excited
to have Amit Sharma.
He is the CEO and founder of CData.
How are you doing?

SPEAKER_01 (00:32):
Doing well.
Nice to be here with you,Amanda.

SPEAKER_00 (00:35):
Happy to have you on the show.
Can you share a little bit aboutC Data?
What services do you provide?

SPEAKER_01 (00:41):
Absolutely.
So C Data is all aboutconnecting data to AR.
The core capability we are, likeas you think about AI has
transformed all of our lives.
It's definitely in theworkspace.
But the real power of AI is onlycaptured when AI can actually
connect to enterprise data.

(01:01):
Think about all the systems thatmodern enterprises use today,
like Salesforce or accountingsystems or CRM systems.
And the ability to be able tohave AI to connect to this data
set in a secure, governedfashion is transformational for
most organizations.
And that's all we're about.

SPEAKER_00 (01:21):
Okay, great.
Well, we're on the show today totalk about enterprise AI and
infrastructure and the impactit's having on the world.
What are some of the AI issuesor problems that you're seeing
right now?
What are some of the toughestones?

SPEAKER_01 (01:37):
I think the toughest problem I'm seeing is like uh
most people have had uh used AIand see how transformational it
can be.
But for certain use cases thatthat involve enterprise data,
we're still at an early stage infiguring out how do we do that?
How do we make sure that we cando this in a secure manner?
How do we make sure that all ofour employees can connect AI to

(02:00):
the data sets that they need?
How do we model the data so AIis efficient?
There's a lot of uh questionsaround token efficiency, et
cetera.
Those are the problems I see inthe industry.
And people are approaching it indifferent ways, but um, we
believe that we have a solutionthat can be transformational.

SPEAKER_00 (02:18):
So a lot of the problems I'm hearing about is
the quality of the AI, um, theanswers, and um where do you
solve this problem?
It's a data problem, oftentimes.
So, what advice do you have?

SPEAKER_01 (02:35):
So, yeah, I mean, when you think about the quality
of uh the responses from AI,most people think that um that
there might be a better modelaround the corner that's gonna
fix the issue.
But I can tell you that theimprovements we have seen in the
recent models, starting from theend of last year with uh the
release of uh new models fromAnthropic and other LLM vendors,

(02:57):
it is truly a step functionabove what we had before.
The models aren't the issue.
The issue is like how do we makethe data and context available
to AI?
There's a lot of context that wecapture in our head that is not
available to AI.
And if you can actually bringthat to AI in a sensible way, if
you can bring the data to bearto AI in a sensible way, uh it

(03:21):
could truly be transformationalin how it works.

SPEAKER_00 (03:25):
When it comes to integrating AI tools, it there
are so many AI tools out there.
So, what is the first step forbusiness leaders as far as
determining what AI toincorporate?

SPEAKER_01 (03:40):
I'm sure many of your users would probably have
heard about MCP.
So, like you could start withany of the model providers.
Like you need a you need a modelthat you want to work with.
You could start with Cloud, youcould start with OpenAI, you
could start with uh Gemini.
In fact, what I'm seeing in mostorganizations, people are using
all three different set of usersare using one model, and I think

(04:01):
that's totally natural.
People have preferences, themodels are stronger in certain
areas.
So I would not try to fightthat.
And if people want to do that,that's that's okay.
The second step is how do webring more context and data into
these AI models?
And that is where, like, if yourusers haven't heard this, MCP is
the new wave and how to do that.
Lots of capabilities in the MCP.

(04:22):
MCP is the model contextprotocol.
It was designed to bring contextinto uh AI models.
There are various approaches todo that.
That is a second step thatpeople should think about how
what is my strategy to usesomething like an MCP to bring
uh data into the AI models to uhto answer whatever business
decisions are we're trying tomake or questions we're trying

(04:45):
to uh ask off of AI.
And once you have those twopieces, it's not enough to just
get any MCP.
You have to think carefullyabout how that MCP solution is
structured, how are you going togovern it?
How does it uh work so thatyou're not creating excessive
use of uh tokens?
Uh so there are some problems,but I think those two or three

(05:07):
steps would get uh your users along way into thinking about the
problem.
And then, of course, it needs tobe some discovery in finding the
best mix.

SPEAKER_00 (05:16):
So it seems like the next step, everyone's moving
towards agentic AI.
We're hearing agentic AI.
And so, what are some of thechanges and some of the things
to be aware of when using AItools that are starting to take
action on their own?
They're not just answeringquestions or uh responding.

SPEAKER_01 (05:37):
Yeah.
So I think uh the big stepchange and when you're thinking
about agentic use case versusconversational use case is the
there's no human in the loop,right?
The agents are taking actions onour behalf.
Uh so the biggest thing, thefirst thing that people need to
think about is the accuracy ofwhatever action they are taking.
Well, did they understand theproblem correctly?

(05:59):
Are they taking the rightaction?
Because the cost of taking anincorrect action is much more
larger than the cost of giving awrong answer to a human being
who can immediately detect itand ask for a clarification or
or ignore the answer and dotheir own research.
That cannot happen in an agenttech system.
So I think what happens is therequirements for accuracy are

(06:23):
much higher when you're thinkingabout agent tech solutions.
That is why this earlier thisyear, we published a paper on
accuracy and responses in AI.
Uh, I would encourage your userbase to go and check it out on
our website.
But I think like the first thingto think about is the accuracy
of what the agent is going toperform as an action is uh

(06:44):
dominant.
The next few steps is how do yousecure the agent so it can only
do what it wants to do?
How do you govern the agent arealso important uh following
questions.

SPEAKER_00 (06:55):
Absolutely.
And as they incorporate more andmore AI tools, as you mentioned,
many companies are using all thedifferent AI tools for various
purposes.
That's a lot of new technologyto keep track of.
What advice do you have forbusiness leaders as far as
making sure they understand allthe tools that they're using and

(07:16):
protecting their the company?

SPEAKER_01 (07:18):
Yeah, I mean, I think like uh if you think about
the AI stack, so to speak, asdifferent layers, uh, you could
probably you should have apreferred tool.
Like, like as I said, I'm veryopen for people switching
models.
People are very familiar doingthat.
You could switch uh models basedon use case, but you should
think about your AI stack indifferent pieces.

(07:42):
So there is the data layer.
Uh, how do you you should thinkabout the data layer?
How do you want to bring data toAI?
I'm obviously biased, and Ithink C data is a fantastic
solution for that.
But regardless of solution youuse, uh, you should think about
how I'm gonna manage my data,how I'm gonna model it, how am I
going to bring it to where toAI, and there has to be a

(08:03):
strategy towards that.
On top of that, is there is theLLM layer where you're thinking
about which LLMs will takeactions on it.
And then you need to think aboutan agent orchestration platform
that's going to orchestrate allthe agents that you're going to
be building.
Multiple solutions for all ofthose uh pieces, but I think
like uh it is beneficial tothink about the broad

(08:25):
architecture first and thenidentifying the best solution
for each segment of thatarchitecture.

SPEAKER_00 (08:32):
And let's talk about the infrastructure here for a
second, too.
As more and more companies areusing AI, and in fact, everyone
is using AI, there is a concernabout where all that data is
housed, and we're needing, youknow, more and more data centers
as the solution.
So I've been hearing aboutactually um, you know,

(08:53):
on-premise data centers comingback to companies.
What are your thoughts on this?

SPEAKER_01 (09:00):
Absolutely.
I think uh data privacy and uhdata residency are important
issues for many organizations.
Uh, we actually believe thatthere are reasons, separate from
how you're building the AIstack, decide where your data is
uh resides.
And people shouldn't just giveaway up on those residency

(09:21):
requirements very easily.
If they have a case for keepingdata on-prem, they should keep
the data on-prem.
We, in fact, uh favor solutionslike virtualization, where you
can leave the data where it isuh and be able to act on it at
real time.
The alternative approach is, ofcourse, to bring warehouse all
the data in one place so AI canwork on it.

(09:42):
So that involves like bringingall the data from various places
into one central warehouse andthen operate on it.
There are use cases where thatis useful, but in our uh in our
experience, oftentimes that'snot necessary and it creates
more problems than it solves.
Uh, so we we actually recommendleaving data in place where it

(10:02):
is uh needs to reside.
As you mentioned, like a lot oforganizations choose to keep
that data on-prem.
So we're in favor of actuallyleaving it there and then
building infrastructure on topthat you can model it,
virtualize it, and still connectit to AI.

SPEAKER_00 (10:15):
Well, AI is advancing rapidly, and as fast
as it's evolving, what do youenvision for the future?
What is the next great wave thatyou're gonna see hit companies?

SPEAKER_01 (10:27):
I think the next wave is going to be of uh of uh
efficiency and more value out ofAI.
So, like uh, I think most peopleare familiar about AI, but we
have only scratched the surfacein terms of uh how much uh value
we can get out of it.
As people start deploying agentsand start working with these

(10:47):
technologies, they will findthat they can do a lot more.
And I'm also a firm believer,like some people worry about the
AI disrupting human capital.
I'm a firm believer that AI isgoing to enhance that.
Like it's not a question of uh II if you go to any company, like
uh they will tell you there isnever a dearth of ideas that

(11:09):
people want to implement.
So it's not a question of whocan do this work instead.
It's a question of how much canof the future we can build, how
quickly.
So I'm a firm believer in that.
So I think we will see immensechange in organizations that are
adopting AI and we will see uhthen move much faster.
Uh things that they were mightbe thinking of doing in the next

(11:30):
10 years, they probably aregoing to be able to do in the
next three to four years.
And even faster.
Yeah.

SPEAKER_00 (11:36):
Well, if there was one key takeaway you could leave
our audience with today, whatwould that be?

SPEAKER_01 (11:41):
Uh the key takeaway is like uh if you are not
already on this journey, uh, Idon't I don't like to uh create
a FUD factor and and scarepeople into doing this.
Uh there is not much to thistechnology.
You can learn it quickly.
Your organizations can left seenpeople not knowing it at all and
being able to adopt it very,very quickly from conversational

(12:02):
AI to agentic AI.
Take the first steps.
Uh, you will be amazed what itcan do.
We have uh I've seen demos wherepeople are brought into a room
that have not seen thecapabilities of AI and they are
wowed.
Create that wow experience uhinside your organization so
people are like what AI can doand that it can truly transform
their work lives.

SPEAKER_00 (12:23):
Wonderful.
Well, thank you so much forcoming on the show and sharing
your insights today.

SPEAKER_01 (12:28):
Thanks, Amanda.
This was great.

SPEAKER_00 (12:30):
And thank you to our audience.
If you have any questions orcomments or concerns, put those
in uh below in the comments, andI'll try to respond back as soon
as possible.
Have a wonderful day.