Conversational AI design: A practical guide for brands

Most customers aren’t anti-AI. They’re anti AI that adds friction.

The difference matters because businesses keep misreading the signal. When customers abandon a chatbot, disengage from an automated flow, or ask for a human after two messages, it’s rarely because they object to AI in principle. It’s because the experience wasn’t seamless enough to see the interaction through, let alone earn their trust.

That trust gap is now the primary barrier to AI adoption. According to industry research, user distrust is the top reason brands are slow to implement conversational AI at scale. And nearly 90% of organizations still aren’t capturing measurable value from AI, despite widespread adoption.

The technology isn’t the problem. The design is.

This guide covers five principles for conversational AI design that closes that gap. We’ll be drawing on real experience from LAQO, Croatia’s first 100% digital insurer, who built a WhatsApp AI agent that now resolves 30% of all customer queries autonomously, empowering their 15-person team to focus on more complex cases without needing to grow headcount.

If you’re a marketer, CX lead, or product owner evaluating how to build AI experiences your customers will be satisfied with, this is where to start. Let’s begin by looking at where these experiences come up short. 

Why most AI experiences fall flat

The trust gap is the real adoption barrier

While AI investment is accelerating, customer willingness to engage with it lags. The gap between the two is where most conversational AI projects fail.

Industry research according to Lippincott identifies user distrust as the single biggest reason brands are slow to adopt AI at scale. Not budget. Not technical complexity. Trust. Customers have been burned by clunky chatbots that loop endlessly, give wrong answers, or pretend to be human when they clearly aren’t. That history shapes their skeptical approach to every new AI interaction.

Nearly 90% of organizations report they aren’t capturing measurable value from AI at scale, despite significant investment and adoption. The lack of results almost always points back to experience design.

The design problem isn’t technical, but experiential

Most conversational AI is built from the inside out. Teams map their internal processes, identify what the AI can technically do, and deploy it. The customer’s side of the conversation, including how it feels to use, whether it’s clear, whether it builds or erodes confidence, tends to get treated as secondary.

The result is AI that works on paper but frustrates in practice.

Your customers already know AI exists. The question they’re actually asking when they encounter your agent isn’t “is this AI?” It’s “is this worth my time?” Designing a chatbot with that question at the forefront is what creates experiences that build trust. Fixing this doesn’t require new technology. It requires better design decisions.

Fixing this doesn’t require new technology. It requires better design decisions.

Five principles for conversational AI design that works

Principle 1: Agents should act, not just answer

The most important distinction in conversational AI right now isn’t between good and bad chatbots. It’s between chatbots and agents, and understanding that difference is the foundation of everything else.

A chatbot responds, retrieves information, surfaces an answer, and waits. An AI agent reasons. It understands intent, makes decisions, triggers actions, and completes tasks without requiring a human to step in at every juncture. The distinction between a chatbot and a conversational AI agent goes deeper than capability. For the full breakdown, Conversational AI vs. Generative AI covers the divide in detail. For design purposes, the key difference comes down to one thing. Chatbots respond. Agents act.

Ervin Jagatic, Product Director at Infobip, made the point clearly in a recent webinar on conversational AI agents. “Chatbots can answer, but conversational agents can act.”

That shift has real implications for design. If you’re building an agent, you’re not just designing a Q&A flow. You’re designing a system that in a single conversation can schedule appointments, calculate premiums, validate inputs, generate payment links, and hand off to a human when something falls outside its scope.

The practical test is simple. Does your AI do something because of the conversation, or does it simply say something? If the answer is the latter, you have a chatbot, not an agent. That’s not necessarily wrong, but it sets the ceiling on what value you can deliver to customers.

Agents outperform traditional automation most clearly in complex, multi-step processes where the path isn’t entirely predictable. Where they need guardrails is in situations requiring fully deterministic outputs, where there’s no margin for error. In those cases, the agent should sit inside a structured workflow that validates its output before anything is executed.

Design for action, but design the boundaries of that action just as carefully. What your agent can do matters. How it presents itself matters just as much.

Principle 2: give your AI a character, not a human mask

One finding emerges consistently from brands that have deployed conversational AI at scale. Customers don’t mind talking to AI. They mind being deceived by it.

When an AI pretends to be human, using a human name with no disclosure or mimicking conversational patterns designed to obscure its nature, customers always notice. When they notice, the experience doesn’t just feel unhelpful. It feels dishonest. Now this UX problem has transformed into a brand problem.

Davor Bruketa, Chief Creative Officer at Bruketa & Žinić & Grey, addressed this directly at Infobip’s webinar Living up to the hype: Conversational AI agents on WhatsApp (February 2026), puts it directly:

The alternative isn’t to make your AI feel robotic. It’s to give it a genuine identity rooted in your brand, not in a simulation of humanity. Brand mascots and clearly defined AI personas give customers something real to engage with. They know they’re not talking to a person, but your brand’s voice made interactive. There’s much more honesty in this approach, and customers appreciate that.

Bruketa references System1 research presented at Cannes Lions 2019, covering three decades of brand effectiveness data. Brands with distinct characters consistently outperform those without in communication effectiveness. Now, in the age of AI, this correlation is even more relevant. Conversational AI is the most direct expression of brand character that exists. It’s the one touchpoint where your brand literally speaks, one-to-one, with millions of customers. Treating it as a purely functional tool is a waste of opportunity.

There’s one more design rule that separates premium experiences from ones that feel like cost-cutting. Always offer a human alternative. Not buried in a menu. Visible, accessible, early. Bruketa puts it directly: “If you force people to only use an AI chatbot, they feel that your service is not premium.” The option to speak to a person isn’t a concession. It’s a signal that you take the interaction seriously.

Design the character. Be honest about what it is. And always leave the door open to a human.

Once that’s clear, the next issue is the journey itself.

Principle 3: Collapse the journey, don’t add to it

Channel fragmentation is one of the most underdiagnosed causes of customer drop-off. It’s not that individual steps in a journey are too hard. It’s that customers are asked to switch between too many of them, and every switch adds cognitive load, hesitation, and another opportunity to abandon.

Marco Kosik, Director of Digital Business at LAQO, speaking at Infobip’s webinar Living up to the hype: Conversational AI agents on WhatsApp (February 2026), describes exactly this problem before they redesigned their customer journey around a conversational AI agent on WhatsApp:

Three channels. Three context switches. Not because the process was complex, but because it was fragmented. The purchase itself was simple, but the journey around it wasn’t.

The conversational AI agent collapsed most of that into a single guided thread on WhatsApp. The customer stays in one conversation as the agent gathers what it needs through dialogue rather than forms, validates inputs in real time, summarizes the selection before payment, and sends a single payment link at the end. The cognitive load drops. Clarity increases. Drop-off decreases.

The real question to ask when designing a conversational AI experience is straightforward. How many channels or steps does a customer currently touch to complete this task, and how many of those can be brought into one thread? The goal isn’t to add a chatbot to an existing journey. It’s to replace fragmentation with continuity.

When you design with the goal of channel consolidation rather than channel addition, the experience improves automatically, because you’re removing friction rather than just digitizing it.

However, a shorter journey doesn’t help if it’s still unclear.

Principle 4: Design for clarity, not speed

There’s a tempting assumption baked into most conversational AI design: that customers want things to move fast. Get to the answer. Complete the transaction. Minimize the messages.

Speed matters. But in many contexts, particularly anything involving money, health, insurance, or significant decisions, clarity matters much more. Designing for speed at the expense of clarity is a reliable way to lose customer trust.

LAQO operates in one of the most trust-sensitive industries. Designing around that reality was a deliberate choice, not a default.

That meant designing the agent to guide and explain rather than push. To answer questions in plain language and let the customer move at their own pace. It also included slowing down when needed and offering additional explanation before moving to the next step, rather than racing toward the close.

It also meant building a clear escalation path. When a conversation becomes complex or sensitive, the agent hands off to a human. LAQO frames this not as a failure of the AI, but as the system working correctly. Knowing its own limits is part of what makes it trustworthy.

The design implication is practical. Map the moments in your customer journey where confusion or hesitation is most likely. This is where the stakes are higher, the language is more technical, or the decision is more consequential. Design for those moments specifically and give the agent room to explain. Build in escalation that feels like a feature, not a fallback.

LAQO’s three core design guardrails are worth borrowing regardless of your industry. Accuracy, consistency, and knowing when to hand off.

Of course, even with the right design, execution can still go wrong. Let’s see how.

Principle 5: start focused, scale deliberately

The most common mistake brands make when building conversational AI agents isn’t technical, it’s scope. They start by mapping their most complex processes, trying to design every possible scenario. The result is something that’s too broad to work well and too ambitious to execute seamlessly.

Ervin Jagatic identifies this as one of the three most persistent misconceptions about agent deployment:

LAQO’s phased approach is a practical model. They started in 2023 with basic Q&A, questions that already had answers on their website. Payments, general definitions, and straightforward responses. That phase was explicitly about learning: how does the system perform, where does it struggle, what does a successful interaction look like?

From there, they moved into product-specific questions. Coverage explanations, policy conditions, anything that required the agent to be accurate, consistent, and compliant, not just helpful. Then came premium calculation for travel insurance, initially with manual logic supporting the conversation behind the scenes.

Today, the agent is connected to LAQO’s web shop and can complete an entire purchase journey autonomously. From understanding the customer’s needs to calculating the premium to handing off to payment. That’s a fundamentally different capability than where they started, but it was built deliberately, phase by phase, with a clear view of what success looked like at each stage.

The lesson isn’t to start small because ambition leads to failure. It’s to start focused because an overly complex scope produces diffuse results. Pick one problem. Pick one channel. Define the KPI you’re trying to move. Design the agent around that, and expand once it’s working.

That discipline is also what makes it easier to get internal buy-in. A focused pilot with measurable results is a far more persuasive case for the next phase than a broad deployment with ambiguous outcomes.

LAQO is a clear example of how this plays out.

What good design looks like in practice: LAQO on WhatsApp

The before: A fragmented journey

Before LAQO deployed their conversational AI agent, their core purchase journey was already digital and efficient on their website and mobile app. The fragmentation appeared when customers started that journey in WhatsApp instead, which many of them naturally did.

A customer asking about travel insurance on WhatsApp triggered a multi-step, multi-channel process. A human support agent would collect the necessary information, calculate the offering in a back-end system, and send a proposal by email. The customer would then open that email, find the payment link, and complete the purchase on the web.

The three modes of interaction are also three context switches, making the process fragmented regardless of its simplicity.

The change wasn’t in the product. It was in the structure of the journey.

The after: One guided thread

The conversational AI agent, built with Infobip, collapsed that journey into a single WhatsApp thread.

The agent extracts the information it needs directly from the conversation and validates inputs in real time. Based on where the customer is travelling and other relevant details, it surfaces additional coverage options in context, not as a sales push. Before moving to payment, it summarizes the customer’s selection clearly, so everything is transparent and easy to review. Only then does it send a single payment link.

The experience is fluid by design. Customers don’t need instructions because they’re already on WhatsApp and already having a conversation. The agent meets them there.

This is where design decisions turn into measurable outcomes.

The results

• 30% of all customer queries on WhatsApp are now resolved by the AI agent, many within just a few messages
• The same 15-person team, covering both sales and customer support, handles significantly higher volume without additional headcount
• Human agents are freed to focus on complex, judgment-heavy cases where their involvement has greater impact

For LAQO, the agent isn’t a cost-cutting tool. It’s a capability multiplier. It lets a small, skilled team deliver a level of responsiveness that would otherwise require a much larger operation.

LAQO’s agent runs on AI chatbot builder. The agent is trained exclusively on LAQO’s approved content, including product information, policy conditions, and internal knowledge base, giving LAQO full control over what the agent can say and how it explains insurance products. Continuous monitoring tracks resolution effectiveness and conversation quality, with the team iterating prompts, tone, and product knowledge on an ongoing basis.

If your goal is an AI agent that creates seamless interactions, start here.

Where to start: A practical design checklist

You don’t need a finished AI strategy to take the first step. You need a clear enough answer to six questions.

1. What is the one problem you’re solving? Not “improve customer experience”, but something much more specific. Pick a process that’s fragmented, a query type that’s overwhelming for your team, or a journey that’s losing customers at a predictable point.

2. Which channel does your customer already use? Don’t design for the channel that’s convenient for you. Design for where your customers already are. LAQO chose WhatsApp because their customers were already there, and the familiarity made adoption frictionless.

3. What does your agent’s character look like? Define tone, personality, and the brand values it should reflect before you write a single prompt. This is a brand touchpoint. Don’t miss the opportunity.

4. Where does the human handoff happen? Map the escalation path from day one. What triggers a handoff? Is it complexity, sensitivity, uncertainty? How does the agent communicate that transition to the customer? The handoff should feel like the system is working, not breaking.

5. What goes into the agent? Quality of input determines quality of output. What approved, accurate content will the agent draw from? Who owns that content and keeps it current?

6. What does success look like at 30 days, 90 days, six months? Define the KPI before you build it. This includes resolution rate, customer satisfaction score, and volume handled without human intervention. Pick the metric that maps to the problem you’re solving and track it from the start.

Frequently asked questions about conversational AI design

Ready to design an AI agent your customers will actually use?

Good conversational AI design isn’t about deploying the most sophisticated technology. It’s about earning trust through honest personas, clear journeys, and experiences that do something useful for the customer.

The brands getting this right aren’t starting big. They’re starting focused. One problem, one channel, one clear measure of success. Then they scale.

Infobip’s AI chatbot builder lets you build, deploy, and manage conversational AI agents across the channels your customers already use, including WhatsApp, SMS, and more. It’s how LAQO went from basic Q&A to full transaction execution, with one platform and a 15-person team.

Talk to an expert about our Conversational AI Platform that’s already turning fragmented customer journeys into seamless, single-thread experiences at scale.