Building an AI agent is becoming easier. Building a customer experience that works reliably across channels, connects with your business systems, and improves over time is a very different challenge.
That distinction became the heart of my conversation with Beerud Sheth on Dilse Omni Talks. Beerud is the Co-Founder and CEO of Gupshup, a global conversational AI platform for marketing, commerce, and customer support. An IIT Bombay and MIT alumnus, he previously co-founded Elance, which later became Upwork, and built Gupshup from a messaging platform into a global conversational AI company.
I put a practical question to him: if a CMO is planning a conversational AI strategy, should the company build, buy, or rent?
His answer was not a shopping list of tools. It was a masterclass in thinking beyond the first demonstration and designing for the entire customer lifecycle.
Table of Contents
The First Agent Is the Beginning, Not the Strategy
A Framework for Understanding the AI Customer Engagement Stack
Build vs Buy vs Rent: Own the Experience, Not Every Component
Many Channels, One AI Brain
Beware the Two-Hour Demo and the Point-Solution Dead End
A Safe Experimentation Loop for Marketing Teams
When AI Speaks for the Brand, the Bar Changes
The First Agent Is the Beginning, Not the Strategy
There is something genuinely exciting about how quickly we can now create. A blog post, landing page, email, or basic AI experience can come together remarkably fast. Beerud’s advice was to embrace that speed: experiment, play, and tinker.
But he drew an important line between creating something and operating it.
Once an AI experience enters production, somebody has to maintain it, support it, optimize it, and integrate it with everything else. Even if AI helps with that optimization, the underlying responsibility does not disappear.
My takeaway was simple: the ease of creating version one can hide the complexity of running every version after it.
An agent that answers a question is useful. A customer engagement system must do considerably more. It may need to initiate an outbound conversation, follow a structured journey, recognize when it cannot help, and hand the customer over to a human.
It also needs to distinguish between contexts. Someone arriving through an advertisement may need a different conversation from an existing customer being re-engaged. These are not minor additions to the agent. They are part of the experience itself.
A Framework for Understanding the AI Customer Engagement Stack
As Beerud unpacked the requirements, I began thinking in layers. The model is one layer, but it is not the whole platform. Messaging connectivity is another. Agents, journeys, analytics, and business integrations add further responsibilities.
Here is my way of organizing the architecture he described, moving from the customer experience down to the capabilities supporting it.
Customer experience and objectives: Define the journeys, segments, campaigns, and outcomes the brand wants to create.
Engagement orchestration: Connect outbound campaigns, structured journeys, advertising flows, and human escalation.
AI capabilities: Use appropriate models and agents for the tasks involved, rather than assuming one model does everything equally well.
Channel connectivity: Support the relevant messaging and communication channels, including WhatsApp, RCS, SMS, Instagram, voice, and video.
Business systems and learning: Integrate with core systems and use conversation analytics to refine the experience.
These are connected layers, not independent purchases. An agent cannot deliver a complete booking experience if it cannot connect with the booking system. A commerce conversation needs access to the relevant catalog. A banking experience must fit into the bank’s core systems.
The model is only one part of a much larger customer engagement stack.
Beerud explained that Gupshup works backwards from what brands need for customer engagement, while allowing businesses to use parts of the platform alongside other tools. That matters: a unified customer experience does not necessarily require every component to come from one vendor.
It does require those components to work together. This is also central to the broader shift explored in how AI is changing omnichannel marketing.
Build vs Buy vs Rent: Own the Experience, Not Every Component
Beerud did not offer a universal verdict that every company should build or every company should buy. Instead, he brought the decision back to core competence.
“Focus on the core competencies of what you need to do, which is design the customer experience.”
Beerud Sheth
For most marketing organizations, building frontier models is not the job. Nor is recreating all the technical machinery needed to make those models useful in production.
The marketer’s contribution sits elsewhere: understanding customers, choosing objectives, designing segments, shaping campaigns, developing conversational journeys, and learning from analytics.
That gives us a practical decision filter:
Own the customer decisions: Your objectives, audience understanding, brand experience, and journey design.
Use tools where they reduce complexity: Especially when they are cheaper, easier, and faster than recreating the capability internally.
Evaluate beyond the immediate use case: Ask whether the platform can support the next channels, integrations, and experiences you are likely to need.
This is not an argument against building. It is an argument against building indiscriminately. The question is not whether your team can create something. It is whether creating and maintaining it is the best use of that team’s attention.
Many Channels, One AI Brain
The channel discussion made the complexity particularly real for me. We have SMS, email, WhatsApp, RCS, Instagram messaging, and other social platforms, alongside voice and video.
I mentioned my son’s preference for Instagram DMs. For him, WhatsApp is not automatically the default. That small example captures the larger challenge: customers choose their channels, and brands have to make the experience work across them.
Beerud’s warning was to avoid building a separate intelligence for every channel.
“Ultimately you want the same brain.”
Beerud Sheth
His analogy was wonderfully clear. Imagine messaging someone, but a different clone of that person handles voice and yet another handles video. That is effectively what happens when disconnected agents represent the same brand.
As policies and experiences change, the organization must then update multiple places. Consistency becomes harder to maintain.
The principle I took away was: different channels, shared intelligence, consistent brand experience. It is a useful test for any platform decision. Does the next channel extend your existing experience, or force you to rebuild it?
Beware the Two-Hour Demo and the Point-Solution Dead End
I admitted something during our conversation: short tool demonstrations can make everything seem so simple that you wonder why you have not built the same thing in two hours.
But the demonstration rarely carries the full burden of production. Channel integrations, analytics, human handoffs, business systems, and ongoing maintenance all sit behind the polished interaction.
Beerud highlighted two traps. The first is a point solution that handles today’s need but leaves you stranded when the next requirement arrives. The second is a vendor still delivering older services without investing in newer capabilities.
His advice was not to stop experimenting. It was to think a couple of steps ahead while doing so.
Conversation analytics is a good example. A click path tells you where somebody went. Conversational logs require a different kind of analysis of the exchange itself. If your ambition is conversational customer engagement, the ability to learn from those conversations belongs in the platform discussion from the beginning.
A Safe Experimentation Loop for Marketing Teams
The strongest practical advice came when we discussed how to move without putting the brand at unnecessary risk.
Beerud suggested sandboxes, internal testing, and phased rollouts. A large organization could begin by running campaigns with employees. It could then expand to 1,000 users, followed by 10,000, and grow from there.
I would summarize his approach as a five-step learning loop:
Recognize the change: Accept that new AI capabilities deserve serious exploration.
Create a sandbox: Give the team room to test within a controlled environment.
Set boundaries: Define the agent’s purpose, guardrails, and human escalation path.
Expand gradually: Use phased exposure rather than an immediate, broad launch.
Learn and optimize: Refine the experience and build organizational confidence before expanding further.
The important phrase here is freedom to fail within boundaries. Without that freedom, experimentation stalls. Without boundaries, experimentation can create avoidable risk.
When AI Speaks for the Brand, the Bar Changes
Beerud made a distinction between AI as a copilot and AI as a pilot.
When I use an AI tool to draft something, I remain in control. I can reject an answer or correct it before using it. When an agent speaks directly to a customer on behalf of the brand, that review may not happen before the response reaches them.
The standard therefore rises: accuracy, staying on message, and respecting the agent’s scope become essential. Beerud gave examples of guardrails around competitors, politics, and unrelated topics. The agent should serve its defined purpose, not become an unrestricted spokesperson.
But his conclusion was not paralysis. It was to find solutions, create guardrails, test safely, and optimize.
That is the balance I would encourage every CMO to pursue: protect the brand while creating a practical path to learning. Building an AI agent is a milestone. Building the connected system around it is what turns that milestone into a customer experience.
I am Saurabh Agrawal and we come with a new episode on Dilse omni talks every fortnight and cover different aspect of omnichannel with amazing speakers.
