Beyond Prompts: Exploring the Rise and Real Impact of Agentic AI

Featured Speakers:-

Links and Resources

• The Digital Handshake: Understanding Agent Communication Protocol
• From Demos to Deployment: Why MCP Is the Foundation of Agentic AI
• The Agent Architecture Decision: Workflow, Single Agent, or Multi-Agent

Transcript

[0:15] Host (Rajnish Ranjan)

Hello to our esteemed viewers. Welcome to another episode of AI Machines and Marketer. Today we’re diving into a topic that’s widely talked about in the world of AI: agentic AI. Our session title is “Beyond Prompts: Exploring the Rise and Real Impact of Agentic AI.”

Let me be honest — for the last couple of years, when we’ve talked about AI, it’s mostly been about chatbots, co-pilots, and better prompting. But with agentic AI entering the conversation, things have escalated. We’re now in a world where AI doesn’t just answer questions — it can reason, plan, take action, collaborate with other agents, and deliver real business outcomes.

Today we’ll focus on what makes agentic AI matter right now, where it’s creating real impact, who’s actually benefiting, and what businesses should get right before trusting AI agents with real work. Please welcome Mr. Arpit Srivastava, Vice President of Growth Marketing and Analytics at Growth Natives, accompanied by two rising leaders: Mr. Vaibhav Kumar, Team Lead – Data Scientist at Growth Natives, and Hitika Saini, Senior Data Analytics at Growth Natives.

Welcome, Arpit. Hello, Vaibhav. Hi, Hitika. Welcome to the show. Before I start, I always ask — how are you guys feeling today?

[2:09] Vaibhav

Really good.

[2:10] Hitika

Excited.

[2:12] Arpit

Great — I think both Vaibhav and Hitika joined our organization as freshers, and they’ve grown tremendously. It’s a pleasure to have these two together and share the stage. I think we’ll have a great session today.

[2:25] Host

Definitely, Arpit. So, are you guys ready? Shall we start with our first segment?

[2:36] Host

My first question is for you, Vaibhav — how would you define agentic AI beyond chatbots, co-pilots, and better prompting?

[2:49] Vaibhav

Let me keep it simple. What does an LLM do? It predicts the next word. An agent goes after the next outcome. Normally, an LLM does one thing — you give it a prompt, it gives you a response, that’s one round of inference, and it stops there. Agents are different. With an agent, you give it a goal, and it makes a plan, runs a loop, calls the right tools through function calling, calls the right APIs and databases, and based on the outcomes, decides the next step. If something looks wrong to the agent, it fixes itself and tries again — but it stays inside the limits we set.

[3:57] Host

Hitika, what’s your take on this?

[4:00] Hitika

I see it as a workflow-level intelligence system. A chatbot is conversation. A co-pilot assists you. But an agent can take real-world actions. For example, if I ask about analyzing a customer, an agent will actually go pull the real data, run the insights, give you the inferences, create a campaign brief, and even route it for approval.

[4:34] Host

That’s quite an answer. Arpit, what’s your take on this?

[4:35] Arpit

You’ve both explained it beautifully. The thing I’d add is that the big deal about “agentic” is that the system should be autonomous — that gets glorified a bit, but I think it’s really the baseline expectation. Think of an agent and draw a parallel to our own agency work: as an agency, we get access to a customer’s data sets, we get their brief, and based on our diagnostics and analysis, we know the gaps and what the future state — the goal — should look like, and the steps to get there. The same thing has to apply when we’re building agents. Now you’ve got the customer, the agency, and the agent all in the picture, and they all have to work together.

[5:38] Host

That’s quite an answer to start with. Hitika, my next question — people are still wondering why this conversation has so much momentum right now. What changed recently that makes agentic AI feel more real, and not just hype?

[5:59] Hitika

What I feel is that model capabilities have changed. We now have real tool connectivity, and we’re moving toward enterprise-ready systems. I remember when we first started — the first chatbot we built was a small open-source model, around 350 million parameters, doing basic question-answering: a user types a natural-language query, and it runs against a database without needing to write SQL. Today we’re talking about trillion-scale parameter models. We’ve shifted to the cloud, we’ve shifted to GPT-class models, we’re talking about memory architectures, feedback loops, and a lot of things that simply weren’t in the picture before. That’s a huge change in just the last few years — it feels like we’re growing every single day.

[6:48] Host

What’s your take on this, Vaibhav?

[6:50] Vaibhav

I think there’s also been a mindset change. The first wave of GenAI taught us to make one person a little faster — quicker emails, quicker code, faster drafts. Now the question itself has changed. It’s no longer “how do we make one person faster,” it’s “how do we redesign the whole process so the system can handle complex tasks.” That’s where the real impact is actually being created.

[7:40] Host

Arpit, what’s your final verdict on this?

[7:42] Arpit

Yeah, I agree it now looks real. The first impression I got was when I switched from ChatGPT to Claude — I could genuinely feel we were getting into a real agentic world, because my initial assumption was that this would stay more like a hybrid workflow. But Claude itself is actually built on an agentic architecture. While prototyping, building, and researching with it, you get a lot of clues about how it handles complex queries, and it gives you output you wouldn’t have imagined. That’s where I felt the shift had really happened, and it happened fast.

Take a step back, too — over the last 10 to 15 years, a lot has changed in our ability to connect to systems. Any commercially available platform — CRM, marketing automation, ERP, advertising platforms — has APIs now, so you have the ability to access data and take actions. With the rise of MCP and similar protocols, you can connect to these systems much faster. It’s not a six-month exercise to connect to an ERP system anymore, as long as MCP and the connection are in place. With that context layer, memory management, and tool-calling all in place, the agentic thing is now real, and it has clear value.

[9:34] Host

Vaibhav, what are the core building blocks of a reliable AI agent system?

[9:40] Vaibhav

Think of it like this — an agent isn’t a single model, it’s a full stack. At the base, there’s the model, the LLM layer, used for reasoning. On top of that, there’s a memory layer, which holds context — both short-term and long-term. Then there’s a retrieval layer, which pulls grounded data instead of just guessing. Then there’s a function-calling layer, for picking the right tools and databases. And on top of all of it, there are guardrails and an execution controller, which decide what the agent can and can’t touch. For sensitive tasks, we also keep a human in the loop.

[10:51] Host

Human is always there, as I always say. What’s your take on this, Hitika?

[10:59] Hitika

From a data science perspective, I’d add evaluation datasets, test cases, observability, and drift monitoring, so you can keep regular track of where the model is drifting as new data comes in. And most importantly, guardrails for security — especially when you’re handling personally identifiable information, which needs to be masked before it ever reaches the LLM. What I feel is that the most important thing isn’t just the LLM — it’s the whole system around the LLM. You can’t ask it to work in isolation; it needs support around it.

[11:38] Host

What’s your take on it, Arpit?

[11:40] Arpit

I think we’ve covered a great area here. What Vaibhav mentioned — that’s like the organs. What Hitika added is more like the immune system. On top of that, we need a structure, a skeleton, that defines and clarifies the actual use cases. When you’re moving from a deterministic system to a probabilistic one — which is what this whole agentic conversation is about — you have to deal with uncertainty and confidence levels. What’s really required is defining tiers of use cases.

If it’s high-stakes, meaning the cost of failure is very high, you design the system so there’s a human in the loop. If it’s low-stakes and the confidence level is high — say, using AI to enrich contact and company information, where you can undo and redo easily — that’s low-stakes, and it can run on its own. You don’t need a human in the loop for everything, because if you try to put a human in the loop for everything, you’re basically creating another job for that person, and the whole point of being agentic was to gain efficiency. So you design the system by looking at what’s high-stakes, what the cost of failure is, and whether something is reversible — and all those organs and immune systems are critical to making that whole thing run.

[13:48] Host

That’s great information on how agents work in theory. Now let’s get into real-world scenarios. Hitika, where is agentic AI delivering real production value today?

[14:02] Hitika

The strongest use cases in production are where tasks are repetitive but still judgment-heavy — customer support, IT and ticketing, software engineering, some procurement and compliance reviews. These aren’t fully autonomous, science-fiction-style systems. They’re real agents delivering workflow-level, specific tasks.

[14:32] Host

What’s your take on this, Vaibhav?

[14:37] Vaibhav

Let me bring it to my world — data. Agents are already earning their place there. They can run data quality checks every morning before anyone even logs in. They can do root-cause analysis when a pipeline fails overnight. My test is simple: does it shorten cycle time? If support tickets get resolved faster, if data incidents get fixed faster — that’s real. If it only looks good in a demo, it isn’t.

[15:26] Arpit

The tangible results we see across clients — the area where it’s made the most impact is at the support level, because there you have defined tasks and defined preferences that a regular client support rep would normally refer to, and agents can easily replace or augment that kind of work. There’s tangible value in efficiency, accuracy, and time saved.

The other area I’d point to is anywhere you’re trying to bring alignment across teams. As a leader, part of your job is bringing alignment across different teams, and your time is precious. If an agent is actually solving a problem that forces alignment across teams, that’s a top-notch production use case. For example, in our case, we have an SEO team focused on organic traffic and a paid media team focused on performance marketing and spend. One thing common to both is the end audience — the research, the keywords driving intent. If we can build some level of agentic support that surfaces information valuable to both teams, and they both use it and align around it, that’s a great use case, because I don’t personally need to spend my time just getting teams to align — there’s a process and an agent in place forcing that alignment.

There are similar use cases elsewhere too. Say you get a campaign brief just 24 hours out and need to launch creative across 6 or 7 platforms — can we have an agent that understands everything about the client, their brand, their voice, the feedback they’ve historically given, and produces a list of creative that’s compatible across all those platforms? The workflow helps you get feedback and keeps self-improving. Again, that’s about bringing alignment between what the client has said in the past and what you know about the brand.

[18:35] Host

Vaibhav, how do you prevent hallucinations when an agent is taking real-world actions?

[18:44] Vaibhav

Let me be honest, because I think this matters most — you can’t fully remove hallucination. Confidence and correctness are two different things. A model can be fully wrong while still sounding fully correct. So we never try to build toward a perfect model. Instead, we build the system assuming the model can be wrong. Before taking any real action, we set rules for the agent — for sensitive decisions, it has to ask a human, what we call human in the loop. And once an agent produces an output, we validate it against a schema and against policy. It’s also about grounding — we ground responses with retrieval.

[19:58] Hitika

Like Vaibhav said, hallucination can’t be eliminated completely — what we can do is reduce it as much as possible. What’s in our hands is trusting the workflow, not just the raw model output. We use retrieval-augmented generation to ground responses, we use schema checks, and we use audit logging and feedback mechanisms so we can continuously monitor how the model is performing and improve its execution over time. I think Arpit has much more to say on this.

[20:35] Arpit

I did mention this earlier — once you’re pivoting from a rule-based, deterministic model to a probabilistic one, you have to deal with uncertainty. There’s an old saying: all models are wrong, some are useful. You need to align stakeholders — the people actually using these agents — around the fact that there’s uncertainty, that failure is possible, and that there’s a defined mechanism to deal with it so it fails gracefully and brings the right people into the loop when needed.

The goal is to be a trustworthy, accurate agent, and for that, it’s fine to take a blended approach — a hybrid of rule-based elements and model-driven autonomy — which collectively raises the accuracy level a few notches. From there, you can even have an agent decide whether a deterministic approach or an agentic approach is the better fit for a given moment, bringing the best of both together to handle accuracy, confidence, and uncertainty all at once.

Hitika, my next question is for you — when should a company use multiple collaborating agents instead of one powerful agent?

[22:28] Hitika

In my opinion, you use multiple agents when you have tasks that are naturally different but need to run in parallel — similar to microservices, where you split responsibilities for more control and easier maintenance. That said, it doesn’t mean you build multiple agents just because it sounds fancy when it isn’t actually required.

[23:02] Vaibhav

Let me add a warning here — I think most teams over-engineer this. More agents means more variance and more places to fail, since these are predictive models. Like Hitika said, don’t go multi-agent just because it sounds advanced. If one agent can do the job, use one. Go multi-agent only where it’s genuinely needed — parallel tasks running at the same time, or independent verification, where one agent validates another agent’s output. Our goal is to build a reliable system, not a bigger architecture. The goal, ultimately, is efficiency.

[24:08] Host

What would you say, Arpit?

[24:09] Arpit

If I put on my product-management hat — at the end of the day, the agentic thing is really a kind of product we’re building. So first, you have to keep the end user in mind — which persona are you actually solving for? A CMO? A CRO? A CTO? A performance marketer? And what does their job actually involve? If you’re solving for a specific persona, there’s a list of things they do, and we try to consolidate everything into one workflow, because we don’t want the person using it to feel like they have to travel across multiple agents and systems every day just to get their work reviewed. That consolidation matters.

The other thought is that not everything needs to be “agentic.” Build solid solutions and utilities that still leverage AI, and then put an agentic layer on top of that — one that has an open conversation with your target persona and routes them to the right tool that solves their real-world problem. That’s the way we’re thinking about it and building it.

[25:42] Host

Vaibhav, what does a weak or failed agentic AI implementation usually look like?

[25:50] Vaibhav

The easiest way to spot a weak one is a chatbot with API access wearing an “agent” label. If you look under the hood, nobody owns it, there are no logs, so whenever it fails, you never know why. There’s no business case for it, so you never know if it’s even worth running. There’s no permission model — no clear definition of what the agent can and can’t touch. And there’s no evaluation framework, so you never know whether it’s getting better or worse over time.

[26:45] Hitika

To add to that — another failure pattern I’d call “agent washing”: labeling a normal process automation, or a plain chatbot, as an “agent” when it isn’t. It won’t optimize for workflow impact or governance the way a real agent should, and a badly failed implementation will usually be missing exactly that.

[27:13] Arpit

One thing to keep in mind is actual cost. As you move from a demo, or a staging environment, into real production with multiple users, there’s real AI inference cost involved. If that cost ends up overriding the actual value and tangible benefit it’s delivering, it turns into more of an engineering cost center than a business win — so you need a clear understanding of the ROI on whatever AI solution you’re offering.

The second thing is alignment from leadership. Does the agent have real backing from someone at the top who’s actually driving adoption? If that’s not happening, it’s a recipe for failure — people will try to avoid it, try to derail it, there’s politics, there are access issues, all kinds of scenarios. That initial alignment — “this is happening, this is a mandate from the CEO, everyone has to show up and adopt it” — needs to come from the top. Otherwise it just becomes a technical project that doesn’t make business sense.

[28:53] Host

How will agentic AI change business models, workflows, or team structures over the next couple of years?

[29:09] Vaibhav

The biggest change I see coming is workflow redesign. Teams will stop asking “how do we add AI to this process” and start asking: if AI can handle routing, do first-level analysis, and take some actions, how should our processes change? How should our teams work more closely together? That’s where the real change is going to happen.

[29:35] Hitika

According to me, the team chart changes too. Right now, engineers, product managers, and each team’s domain experts work in fairly separate, fixed lanes. I think AI puts all of them into a single room. After a couple of years, I think there will be new job titles — Agent Product Owner, Agent Reliability Engineer, evolving ML and data scientist roles, and Agent Risk specialists. Honestly, today’s answer is really for the HR recruiters — be ready for new openings, because an agent can touch data, touch the product, and touch risk, so all the teams need to work together under one roof.

[31:00] Arpit

From an overall model and pricing perspective, I do see change coming too. If it’s a SaaS business, pricing has traditionally been based on number of seats. If it’s an agency, it’s traditionally based on selling time. I think those models will shift toward being more outcome-based.

And like Vaibhav mentioned, there will be new roles. I think every job will eventually have a relevant agent attached to it — the same way everyone uses something like Excel in their job today, we’ll have agents for our particular jobs too, and we’ll need people who can maintain those agents and get real value out of them. There’s also a shift coming in how we work with clients — new roles similar to what OpenAI calls “forward-deployed engineers.” You’ll have forward-deployed marketers, forward-deployed campaign managers — people embedded directly into the client’s ecosystem. That matters because these agents need a lot of business context, and that context doesn’t fully live in documents, forms, or questionnaires — a lot of it just sits in people’s heads. It becomes the job of that forward-deployed person to capture that context from leadership and feed it into the system, so that what worked in a clean demo also works in the messy real world of a client, where data is scattered and information is siloed across people and systems.

Rapid Fire Round

[33:34] Host

Let’s start the rapid-fire round with you, Arpit.

One word that defines agentic AI. Not a magic wand — you need the infrastructure and the execution to back it up.

One capability that surprises people today. Doing an entire process end to end, and actually achieving outcomes through agentic solutions.

What fails faster — strategy or execution? Execution.

What’s more dangerous — over-automation or under-understanding? Under-understanding.

What’s harder — making an agent act, or making it act correctly? Making it act correctly.

One signal that a company faking “AI-first” gives away? No data foundation, and no real business use case.

One-liner — what kills most agentic AI projects? (This answer was cut off in the source recording — Arpit begins “I think the right AI system is…” and the sentence isn’t completed.)

[34:44] Host

Hitika — one thing every production agent needs?

One underrated technical layer. Memory architecture.

One red flag in an agentic AI demo. That it only works under perfect conditions.

One sentence — what separates a chatbot from an agent? A chatbot responds. An agent takes action.

[35:18] Host

As we head toward the end of the segment, I’d like one final thought from each of you on what you want to leave the audience with. Let’s start with Hitika.

[35:31] Hitika

I believe the winners won’t be the companies with the flashiest agents — they’ll be the ones that build reliable, measurable agentic workflows.

[35:43] Vaibhav

Agentic AI doesn’t mean handing someone a key to a black box and walking away. We have to build the system, redesign the whole architecture, so multi-step or repetitive tasks get handled efficiently — while humans take care of judgment on the agent’s responses, the context, and the edge cases that genuinely need one. AI, our agents, won’t replace us. They’ll just work next to us. It’s collective work.

[36:37] Arpit

I believe it’s a genuinely exciting time for all of us, because it’s one of those moments where we get to do things that have never been done, that were never even discussed before. For me, the whole definition of top performance has changed — it’s no longer about who can write code, or produce good content, or run a good campaign. It’s about who can learn and relearn, who has the curiosity to actually embed themselves into the agentic system, whatever their field is. That curiosity, and the hunger to keep relearning, is what I think will define people going forward.

[37:42] Host

As we move toward the conclusion, one thing is clear — agentic AI has taken us well beyond the simple prompt. It’s brought us to systems that can plan, act, collaborate, and deliver real outcomes. The opportunities are significant, but success will come down to better governance, clear data, human insight, and, most importantly, better execution.

A big thank you to Arpit Srivastava, Hitika Saini, and Vaibhav Kumar for joining us today, and to our viewers — thank you for being with us. Until next time, keep learning, keep innovating, and stay ahead in the world of AI. Thank you.

See more webinar