Building Conversational Agents with Their Own Identity

What happens when you build on a platform that can change its rules without prior notice?

This question is not new, but it regained relevance on October 18, 2025, when Meta updated the terms of its WhatsApp Business API to specifically ban “AI Providers” (general-purpose artificial intelligence providers) starting January 15, 2026.The new terms are explicit:

“Providers and developers of artificial intelligence or machine learning technologies, including but not limited to large language models, generative AI platforms, general-purpose AI assistants, or similar technologies (...) are strictly prohibited from accessing or using the WhatsApp Business Solution”

when such technologies constitute the core functionality of the service.

As reported by TechCrunch, this change will directly affect companies such as OpenAI, Perplexity, and others that had built large-scale conversational assistants on WhatsApp.

However, beyond the immediate implications of the decision, this incident opens a broader reflection on what it means to build products with their own identity. The volatility of third-party platforms is a constant in the modern digital landscape. The question is not whether platforms will change their rules, but when. How can we navigate this reality so that we not only react to external changes but build with a wider vision from the beginning?

WhatsApp: Excellent for Validation, Limited for Scaling

Building on WhatsApp has an undeniable and immediate appeal: instant access to more than 3 billion users, a validated interface that requires no explanation, and a learning curve reduced to zero. In other words, you do not need to convince the user how to use your product; the conversational grammar of WhatsApp is already internalized.

These advantages make it the ideal environment to validate hypotheses quickly. Its immediate reach allows proof-of-concepts to be built with ease. Measuring engagement and retention becomes more accessible thanks to a massive and familiar user base.

However, here lies the paradox: what makes WhatsApp so effective as a channel is also its fundamental limitation as a platform for building unique products. Building a conversational agent on WhatsApp is a bit like trying to design a house on rented land: you can decorate, yes, but the structure, dimensions, and systems are predetermined.

This structural limitation generates large-scale commoditization. When all bots converse within the same framework, with the same interface restrictions and design limitations, the products inevitably start to resemble one another. The only remaining differentiation space is content, knowledge bases, and conversational tone. It is too narrow a margin to build sustainable competitive advantage.

At Promtior, where we work on developing integrations and multi-agent systems, we have observed this pattern repeatedly: clients with unique value propositions end up with experiences indistinguishable from their competitors simply because the platform does not allow them to express that difference. The question they ask us is not technical, but strategic: how can we build something memorable when everything looks the same?

In addition, there is a matter of context and brand perception: to what extent can we build long-term business relationships on the same platform where users receive family memes, group chats, and local pizza promotions? We are not only competing with other products but with everything that demands attention in that context.

In short, WhatsApp is excellent for validating ideas, testing assumptions, and measuring initial interest. However, it presents clear limitations for developing mature products with their own identity, distinctive personality, and the ability to evolve without external constraints.

AI Typologies in Products: Understanding Before Building

Before designing a conversational agent with its own identity, it is essential to understand how AI integrates with the core of the product. Not all products are the same, and design decisions depend directly on the role AI plays in the experience.

We can think of five main typologies that define the spectrum of AI integration in relation to the user and the product:

1. Invisibility: AI operates as a hidden layer that optimizes processes behind the scenes. This is the case with recommendation systems or algorithms that adjust interfaces without the user explicitly noticing.

2. Automation: AI replaces tasks that were previously performed manually and repetitively. Here, the challenge is to maintain quality and handle edge cases.

3. Augmentation: AI amplifies human capabilities without replacing them. It represents collaboration between humans and computers to solve complex problems. In this case, clearly defining the human-machine exchange is essential.

4. Agentic Experience: AI presents itself as an independent actor that the user interacts with directly. The conversational assistant has tools and a certain level of independence; communicating what it has understood and its progress at every step is crucial to keep the user in control.

5. Proactivity or Reactivity: AI can anticipate user needs based on observed behaviors or context over time (proactivity), or respond specifically to direct requests (reactivity). Each mode demands different interaction flows, complexities, and expectations.

These typologies are not mutually exclusive; a single product can combine several. What matters is identifying which one dominates in your use case, because that will determine key decisions: what information should the user see at each moment? When is it necessary to keep a human in the loop? What level of transparency is required regarding the agent’s reasoning process?

---

Anatomy of a Purposeful Agent

Once we understand the typologies, we can break down what makes a robust system with its own identity. An effective conversational agent that transcends third-party platform limitations requires deliberate architecture.

A robust conversational agent is not a chatbot. It is a complex system that requires four fundamental dimensions: solid technical components, a personality consistent with the brand, mechanisms that keep the user in control, and interfaces that express a unique identity. Let’s break them down.

Technical Components

The architecture of an effective conversational agent rests on five core pillars. The complexity and sophistication of each will vary depending on the predominant typology: an agentic experience will demand more robust tools than a simple automation system, for example.

• Knowledge Base: This is not about a generic dataset, but about specific, curated, and contextualized knowledge. A financial assistant requires up-to-date market data; a technical support bot needs precise product documentation. The quality and relevance of the knowledge base directly determine the user’s perceived reliability.

• Tools: Tools are external capabilities the agent can invoke to perform actions beyond conversation. A conversational agent by itself can answer questions, but it cannot query databases, send emails, process transactions, or interact with external systems. Tools extend the agent’s functionality by connecting it with the real world: APIs, databases, legacy systems, web services. In essence, they transform the agent from an “informed conversationalist” into an “actor capable of modifying the environment.” The difference between a chatbot that only talks and an agent that can act lies in its tools.

• Models and Prompts: Model choice is not trivial, not all models have the same purpose, personality, reasoning ability, contextual understanding, or balance between creativity and rigor. A specialized model may be precise but lack conversational flexibility. The art lies in selecting (or orchestrating) the right set of models for the specific use case. Prompts are the instructions given to the model so it understands the context and performs the task; they are the mechanism through which the agent’s behavior is shaped, from its tone to its reasoning style. Different formulations, examples, and instruction structures completely change how the model processes and responds, transforming the same base model into different personalities and behaviors as needed. The relationship is bidirectional: the chosen model determines not only what the agent can do, but also how prompts must be structured, the level of detail required, the type of examples that work best, and the emphasis on explicit instructions versus the model’s emergent capabilities.

• Guardrails: Boundaries are not arbitrary restrictions, they are safety mechanisms that maintain coherence with the product’s purpose. Some are widely used to prevent bias, but there are gray areas where what works for one product does not work for another. For example, a legal advisory agent requires completely different guardrails than a customer support one. These limits not only prevent hallucinations but keep the agent within its domain, ensuring security and consistency, and protecting the integrity of the experience.

• Graceful Failures: Every system fails. The difference between a professional product and a prototype lies in how failure is handled. Graceful failures are those that do not break the experience, that communicate honestly about limitations, and that offer alternatives or escalation paths when necessary.

Brand-Consistent Personality

Beyond technical components, the agent’s voice is not an aesthetic detail; it is a fundamental design element that defines how the product relates to its users. Tone, chosen vocabulary, and communication style all express personality and, by extension, the brand’s identity.

Personality must align with the agent’s typology: a proactive agentic experience can allow for warmth and informal conversation, while an augmentation system should prioritize clarity and precision to avoid disrupting the user’s workflow.

Consider two contrasting cases: a financial assistant will prioritize clarity, precision, and rigor. Its responses will be direct, using technical terminology when appropriate, and avoiding ambiguity. In contrast, a creative assistant will aim for empathy, warmth, and rapid iteration. Its responses may be more elaborate, incorporate curiosity, and encourage exploration.

Adjusting personality through prompts is an art mastered with practice. It is not only about writing systematic instructions, though that matters, but about understanding how different formulations and examples shape the model’s emergent behavior. Gradually, you learn that adding “think step by step” changes reasoning patterns, that providing format examples affects response structure, and that adjusting temperature modulates between creativity and determinism.

The crucial point: the agent’s personality must be consistent with the product’s identity. The conversation is an extension of the product’s DNA; how the agent speaks communicates as much about purpose as it does about the intended experience.

Human in the Loop: Keeping the User in Control

Control is communicated in various ways: system states, progress indicators, visual feedback, etc. For the user to understand what is happening and why is fundamental to a satisfying experience.

In this context, “human in the loop” refers to keeping the human as supervisor and decision-maker in critical moments of interaction. The intensity of this control varies depending on typology: while in invisible automation the user may not even notice the AI working, in agentic experiences with action capabilities explicit control becomes imperative. But when should it be applied and how can it be integrated effectively?

• When to apply human-in-the-loop: Irreversible actions, such as deleting critical data or executing financial transactions, require supervision. Complex decisions involving significant trade-offs demand confirmation. High-uncertainty scenarios, where the model may be exploring unknown territory, benefit from human judgment.

• How to integrate it: Turn-based interaction is key. Instead of allowing the agent to act automatically, structure a flow where the user can review, confirm, or modify before proceeding. Progressive confirmations, like “You are about to send an email to 150 recipients, do you confirm?”, provide control without creating unnecessary friction.

Properly designed human-in-the-loop systems offer three benefits: first, they increase user trust by ensuring control over important processes. Second, they reduce operational risk by preventing unintended actions. Third, they foster agency, allowing users to influence outcomes, which is essential for building long-term relationships.

---

Interfaces That Express Identity: Generative UI

Here lies one of the key differences between building on restricted platforms like WhatsApp and developing products with their own identity: the ability to design interfaces that adapt to the interaction context.

Natural language is powerful for many tasks, but it poses significant challenges when designing precise and efficient experiences. Understanding these limitations leads us to explore what we call generative UI: interfaces that combine the best of conversation with traditional interactive components.

The Limitations of Natural Language

Imagine a user needs to configure complex filters for a search: 15 different categories, specific numerical ranges, and boolean relationships between criteria. Trying to do this purely through natural language creates several problems: first, the user may not know how to express exactly what they want (“I want between 10 and 20, but not 15”). Second, broad or mutually exclusive categories like “product type” with 30 options are hard to navigate conversationally. Third, the inherent ambiguity of language can lead to misinterpretation.

It is not that natural language is deficient; the point is that it is not the optimal tool for every task. Decades of validated UX research support the advantages of dropdowns for multiple selections, radio buttons for mutually exclusive options, sliders for continuous ranges, cards for visual comparisons, and numeric fields for precise quantitative data. These components exist for a reason: they mitigate human error and reduce cognitive load in specific tasks.

Every interface is born with a purpose. WhatsApp, for instance, is excellent for conversational immersion when the goal is to maintain a natural dialogue. But for complex option deliberation, visible trade-off calculation, or navigation across multi-dimensional parameters, pure conversation is not the most effective solution.

The Smart Hybrid

Generative UI proposes a smart hybrid: combining conversation with interactive components that are generated contextually. Imagine a cost-tracking app where a user writes in natural language, “I had lunch today for $15 at the lobby restaurant.” The agent interprets the intent, identifies the category, amount, and details, but instead of automatically logging the data, it displays a structured modal where the user can confirm, correct, and complete any misinterpreted fields.

This is not conceptually new; adaptive interfaces have existed for a while. The difference now is that we can dynamically generate components based on the agent’s contextual understanding. Internal validation structures ensure the loaded information is reliable, while the user keeps control to adjust anything the model misunderstood.

The result: we preserve the power of natural language for fluid intent capture while leveraging the precision and clarity of visual interfaces when truthfulness and control are required. The user gains comfort in input; the system gains data integrity. It is a win-win.

Transparency and Control: Timelines and Intermediate States

One of the most significant challenges for conversational agents is managing asynchronous states and processing times. When the agent needs to query a database, analyze long documents, or perform time-consuming tasks, the user is left waiting without knowing what is happening.

Timelines and visible intermediate states that communicate real-time progress solve this issue. Instead of waiting in silence, we show understandable progress messages: “Querying knowledge base...”, “Analyzing documents...”, “Generating personalized response...”. These should be written in language the user understands, avoiding technical jargon.

The benefits of this transparency are multiple. First, it gives meaning to waiting: it is not empty time but a process underway with a clear direction. Second, it helps users infer reasonable delays (“if it is analyzing 1000 pages, it will take a few minutes”). Third, it reduces anxiety because users perceive progress. Fourth, it builds trust through honesty about timing and processes.

Transparency also applies to turn-taking in critical actions. It is not enough to show what is happening; sometimes we must show how it is happening, offering control points where the user can intervene, stop, or modify the course of action before completion.

---

From POC to Product with Identity

The change in WhatsApp’s terms is not a crisis, it is a clarification. It forces us to answer a question we should have asked from the start: are we building a product or renting an experience?

WhatsApp will remain excellent for validating hypotheses and quickly building proof-of-concepts, especially for reactive agents with limited functionality. But developing sophisticated agentic experiences, adaptive interfaces, complex feedback loops, and systems that evolve without external constraints requires sovereignty over the platform.

Recent changes only accelerate an inevitable transition: from a reactive attitude to a proactive one. The strategy is not to abandon established channels, but to understand how to articulate different platforms according to their purpose. WhatsApp is an excellent channel for acquisition or human support, while owned platforms are better suited to build long-term relationships where every element (from the knowledge base to the visual interface, from the agent’s tools to its personality) can be expressed without compromise.

There is no universal formula. Each product has its purpose, tone, and specific way of integrating AI that provides genuine value. What all share is the need for clarity in decisions, transparency in processes, and design that keeps the user in control. These three pillars enable the shift from utilitarian conversation to meaningful, lasting connections.

In a world filled with tools and platforms, true differentiation lies in using available technology thoughtfully, but with intentional purpose. The fundamental questions remain: what kind of relationship do you want to build with your users, and how can you add value in their everyday life?

The difference between just another chatbot and a memorable product lies in the answers to these questions.

—

Article written by the Promtior team, where we build AI-powered products and solutions.