Long story short
Long story short
After the redesign, successful self-service agent creation increased by 72% and setup-related support requests dropped by one-third.
After the redesign, successful self-service agent creation increased by 72% and setup-related support requests dropped by one-third.
Redesigned the AI agent builder in Flametree so non-technical users can create and launch agents on their own. A code-heavy setup became a structured, natural-language experience organized into six sections: General, Knowledge, Capabilities, Channels, Workflow and Settings.
Context
Context
Flametree is an enterprise AI platform for customer operations and agent workflows.
Flametree is an enterprise AI platform for customer operations and agent workflows.
Businesses use it to create AI agents that talk to customers across channels like WhatsApp, Telegram, Instagram, email, phone calls and website widgets. Customers range from fintech companies to clinics, online stores and independent creators, and feedback from larger clients shapes how the platform evolves.
The core value of the product is that teams can configure and launch their own agents: define how they behave, what they know and how they interact with customers. But the setup experience had a strong technical bias, which made adoption hard for non-technical users.
Businesses use it to create AI agents that talk to customers across channels like WhatsApp, Telegram, Instagram, email, phone calls and website widgets. Customers range from fintech companies to clinics, online stores and independent creators, and feedback from larger clients shapes how the platform evolves.
The core value of the product is that teams can configure and launch their own agents: define how they behave, what they know and how they interact with customers. But the setup experience had a strong technical bias, which made adoption hard for non-technical users.
My Role
My Role
Contract Product Designer. I worked alongside another product designer, and we split ownership of different parts of the platform while aligning on shared components. The agent builder was one of several areas we worked on.
Team
Team
2 product designers, frontend and backend developers, analysts, an AI specialist who defined how agents are trained and behave, and a business representative who brought in real customer needs.
Timeline
Timeline
November 2025 – August 2026.
November 2025 – August 2026.
This was a period of iteration, not continuous work on a single screen: we returned to the builder several times while working on other parts of the platform.
This was a period of iteration, not continuous work on a single screen: we returned to the builder several times while working on other parts of the platform.
The original interface was designed with a developer-first mindset, exposing technical settings that created unnecessary complexity for non-technical users.
The original interface was designed with a developer-first mindset, exposing technical settings that created unnecessary complexity for non-technical users.
Problem
Problem
The setup flow exposed low-level concepts and settings, so new and non-technical users couldn’t create agents independently.
The setup flow exposed low-level concepts and settings, so new and non-technical users couldn’t create agents independently.
Key behavior and conversation logic were defined with Python code
Concepts like chunking and workflows were exposed directly
The interface had many unclear, low-level settings
As a result:
New users often failed to create even one agent during the trial
Customers relied heavily on support to finish setup
Non-technical users were effectively blocked from using the product
There was a big gap between the product’s potential and actual activation.
Key behavior and conversation logic were defined with Python code
Concepts like chunking and workflows were exposed directly
The interface had many unclear, low-level settings
As a result:
New users often failed to create even one agent during the trial
Customers relied heavily on support to finish setup
Non-technical users were effectively blocked from using the product
There was a big gap between the product’s potential and actual activation.
Process: Understanding the problem
Process: Understanding the problem
I started by reviewing the existing flow end to end and looking for friction points:
Walked through the setup experience step by step
Looked at where users dropped off, especially trial users who never created an agent
Used indirect signals: support requests, onboarding friction and feedback from the customer-facing team
Users weren’t failing because they lacked a need. They were blocked by complexity.
I started by reviewing the existing flow end to end and looking for friction points:
Walked through the setup experience step by step
Looked at where users dropped off, especially trial users who never created an agent
Used indirect signals: support requests, onboarding friction and feedback from the customer-facing team
Users weren’t failing because they lacked a need. They were blocked by complexity.
Process: Research & references
Process: Research & references
Users don’t want to “configure AI”. They want to describe what they need in plain language.
Users don’t want to “configure AI”. They want to describe what they need in plain language.
I explored:
How similar tools structure AI agent configuration
Patterns from no-code / low-code products
Interaction models in tools like ChatGPT, where behavior is defined in natural language
I explored:
How similar tools structure AI agent configuration
Patterns from no-code / low-code products
Interaction models in tools like ChatGPT, where behavior is defined in natural language
Process: Iteration & design decisions
Process: Iteration & design decisions
The final structure is the result of several iterations. Early concepts were simpler, and the builder grew in capability as we learned what customers actually needed.
The final structure is the result of several iterations. Early concepts were simpler, and the builder grew in capability as we learned what customers actually needed.
1. From code to natural language
1. From code to natural language
I replaced Python-based configuration with structured, human-readable settings for the things most agents need: who the agent is, how it speaks, what language it uses and how it greets users. These live in the General section.
I replaced Python-based configuration with structured, human-readable settings for the things most agents need: who the agent is, how it speaks, what language it uses and how it greets users. These live in the General section.
2. Six sections, one mental model each
2. Six sections, one mental model each
I reorganized the experience into six sections. Each answers one question a user is asking themselves.
I reorganized the experience into six sections. Each answers one question a user is asking themselves.
2.1 General: Who is this agent and what should it deliver?
2.1 General: Who is this agent and what should it deliver?
General brings together three blocks that used to be spread across the product:
Agent Profile: an internal name for finding the agent on the platform, plus the name it uses when talking to customers, a description and tags
Behavior: tone of voice (a ready-made style or a custom one defined by instructions), default language, additional instructions and an optional welcome message with quick-reply buttons
Conversation Result: what data the agent should extract from a conversation and send to internal systems such as a CRM, using standard fields like name, topic, phone and email, or custom fields defined by instructions
Conversation Result used to be a separate tab. Moving it into General lets users define identity, behavior and business outcome in one place.
General brings together three blocks that used to be spread across the product:
Agent Profile: an internal name for finding the agent on the platform, plus the name it uses when talking to customers, a description and tags
Behavior: tone of voice (a ready-made style or a custom one defined by instructions), default language, additional instructions and an optional welcome message with quick-reply buttons
Conversation Result: what data the agent should extract from a conversation and send to internal systems such as a CRM, using standard fields like name, topic, phone and email, or custom fields defined by instructions
Conversation Result used to be a separate tab. Moving it into General lets users define identity, behavior and business outcome in one place.
General answers “Who is this agent and what should it deliver?” It combines profile, behavior and conversation results.
General answers “Who is this agent and what should it deliver?” It combines profile, behavior and conversation results.
2.2 Knowledge: What does the agent know?
2.2 Knowledge: What does the agent know?
Q&A pairs, text, web links, files and YouTube videos. A central Knowledge section lets teams manage materials independently of any single agent and connect the same content to several agents, so company information doesn’t need to be set up again each time.
Q&A pairs, text, web links, files and YouTube videos. A central Knowledge section lets teams manage materials independently of any single agent and connect the same content to several agents, so company information doesn’t need to be set up again each time.
Knowledge answers “What does the agent know?” Q&A, text, links, files and videos work the same way, so text is shown in full and the rest in miniature.
Knowledge answers “What does the agent know?” Q&A, text, links, files and videos work the same way, so text is shown in full and the rest in miniature.
2.3 Capabilities: What can the agent do?
2.3 Capabilities: What can the agent do?
The most important new concept, with three levels of depth:
Building Capabilities: ready-made abilities that can be switched on and configured without code, such as image understanding, knowledge search, quick replies, transfer to a human agent with custom conditions, sending lead data by email and confirming a received request to the customer
Tools: actions the agent can perform, either connected from existing ones or created new
Skills: more complex abilities that combine several tools with instructions on how to use them
Tools and Skills are managed centrally and can be reused across agents. Creating custom ones may still require technical knowledge, so they are aimed at advanced users. Most users can cover core scenarios with Building Capabilities alone.
The most important new concept, with three levels of depth:
Building Capabilities: ready-made abilities that can be switched on and configured without code, such as image understanding, knowledge search, quick replies, transfer to a human agent with custom conditions, sending lead data by email and confirming a received request to the customer
Tools: actions the agent can perform, either connected from existing ones or created new
Skills: more complex abilities that combine several tools with instructions on how to use them
Tools and Skills are managed centrally and can be reused across agents. Creating custom ones may still require technical knowledge, so they are aimed at advanced users. Most users can cover core scenarios with Building Capabilities alone.
Capabilities answers “What can the agent do?” Ready-made abilities work out of the box, while Tools and Skills extend the agent for advanced scenarios. Empty state above, configured agent below.
Capabilities answers “What can the agent do?” Ready-made abilities work out of the box, while Tools and Skills extend the agent for advanced scenarios. Empty state above, configured agent below.
2.4 Channels: Where does the agent talk to customers?
2.4 Channels: Where does the agent talk to customers?
WhatsApp, Instagram, Facebook, Telegram, email, phone calls and the web widget, all in one place. The functionality is the same, but the settings are now organized more clearly.
WhatsApp, Instagram, Facebook, Telegram, email, phone calls and the web widget, all in one place. The functionality is the same, but the settings are now organized more clearly.
Channels answers “Where does the agent talk to customers?” WhatsApp, Instagram, Facebook, Telegram, email, phone calls and the web widget are managed in one place.
Channels answers “Where does the agent talk to customers?” WhatsApp, Instagram, Facebook, Telegram, email, phone calls and the web widget are managed in one place.
2.5 Workflow: How does the conversation unfold?
2.5 Workflow: How does the conversation unfold?
A visual editor for conversation scenarios. Users create states as cards, connect them with transitions and define which tools and skills the agent can use in each state. Logic that used to be written in Python can now be built visually.
Workflow is optional. Simple agents don’t need it, but it opens up more complex scenarios for those who do.
A visual editor for conversation scenarios. Users create states as cards, connect them with transitions and define which tools and skills the agent can use in each state. Logic that used to be written in Python can now be built visually.
Workflow is optional. Simple agents don’t need it, but it opens up more complex scenarios for those who do.
Workflow answers “How does the conversation unfold?” Dialogue logic that used to live in Python code is now built visually, as connected states.
Workflow answers “How does the conversation unfold?” Dialogue logic that used to live in Python code is now built visually, as connected states.
2.6 Settings: Advanced configuration
2.6 Settings: Advanced configuration
AI modules (language model and other components), environment variables, session limits and operator interaction rules, additional parameters and agent credentials. This is deliberately a separate, advanced layer, so most users never need it to create a working agent.
AI modules (language model and other components), environment variables, session limits and operator interaction rules, additional parameters and agent credentials. This is deliberately a separate, advanced layer, so most users never need it to create a working agent.
Settings answers “How is the agent configured under the hood?” AI modules, limits, credentials and other advanced parameters live here, away from the basic setup.
Settings answers “How is the agent configured under the hood?” AI modules, limits, credentials and other advanced parameters live here, away from the basic setup.
3. Workflow: moving complex logic out of code
3. Workflow: moving complex logic out of code
Basic settings and ready-made capabilities turned out not to cover every customer scenario. Complex dialogue logic still required programming. Workflow came from this gap rather than from a wish to add another screen: it gives users control over agent behavior without asking them to write code. Advanced tools and skills can still require technical knowledge, but the barrier for everyday scenarios is much lower.
Basic settings and ready-made capabilities turned out not to cover every customer scenario. Complex dialogue logic still required programming. Workflow came from this gap rather than from a wish to add another screen: it gives users control over agent behavior without asking them to write code. Advanced tools and skills can still require technical knowledge, but the barrier for everyday scenarios is much lower.
4. Playground: available, not distracting
4. Playground: available, not distracting
Users can test their agent in real time in the Playground. In the first version it was always visible and competed for attention even when nobody was testing. In the new version it lives in a collapsible side panel on the right. It stays one click away, and the main settings get the space they need.
To keep the first steps frictionless, an agent is created as soon as it is named, with no manual save, and users can start chatting with it immediately.
Users can test their agent in real time in the Playground. In the first version it was always visible and competed for attention even when nobody was testing. In the new version it lives in a collapsible side panel on the right. It stays one click away, and the main settings get the space they need.
To keep the first steps frictionless, an agent is created as soon as it is named, with no manual save, and users can start chatting with it immediately.
Playground answers “How does the agent actually behave?” It sits in a collapsible right-hand panel: always one click away, out of the way when not needed. Expanded here, with conversation results visible.
Playground answers “How does the agent actually behave?” It sits in a collapsible right-hand panel: always one click away, out of the way when not needed. Expanded here, with conversation results visible.
5. Naming
5. Naming
We went through many rounds of naming for sections and tabs, for example Knowledge vs. Knowledge Base vs. Knowledge Sources. Names were checked internally, with the sales team and with customers, because labels often decide whether people understand the structure at all.
We went through many rounds of naming for sections and tabs, for example Knowledge vs. Knowledge Base vs. Knowledge Sources. Names were checked internally, with the sales team and with customers, because labels often decide whether people understand the structure at all.
Process: Validation
Process: Validation
We checked decisions in several ways.
We checked decisions in several ways.
Walkthroughs with the sales team, who know the product and hear customers’ questions daily
Interviews with existing customers, who explored the interface and asked questions
Observing how quickly people complete key tasks and where they get stuck
Findings fed back into names, element placement and the organization of features.
Walkthroughs with the sales team, who know the product and hear customers’ questions daily
Interviews with existing customers, who explored the interface and asked questions
Observing how quickly people complete key tasks and where they get stuck
Findings fed back into names, element placement and the organization of features.
Solution
Solution
The final solution turns agent setup from a technical configuration tool into a structured, human-readable workflow.
The final solution turns agent setup from a technical configuration tool into a structured, human-readable workflow.
Users can now:
Describe agent identity, behavior and expected results in natural language
Add knowledge without technical setup and reuse it across agents
Switch on ready-made capabilities, and reach for tools, skills and visual workflows only when they need them
Connect channels in the same flow
Test the agent instantly
The product shifted from “configuring AI” to “describing intent”, while keeping depth for advanced users.
Users can now:
Describe agent identity, behavior and expected results in natural language
Add knowledge without technical setup and reuse it across agents
Switch on ready-made capabilities, and reach for tools, skills and visual workflows only when they need them
Connect channels in the same flow
Test the agent instantly
The product shifted from “configuring AI” to “describing intent”, while keeping depth for advanced users.
Impact
Impact
The redesign lowered the entry barrier and made the product accessible to non-technical teams.
+72% successful self-service agent creation
−33% setup-related support requests
Non-technical users create agents independently
Existing clients build agents without assistance
The redesign lowered the entry barrier and made the product accessible to non-technical teams.
+72% successful self-service agent creation
−33% setup-related support requests
Non-technical users create agents independently
Existing clients build agents without assistance
What’s next
What’s next
A new onboarding for the updated builder is in design to help users learn the six-section structure. It is still at the design stage and not yet in the product.
A new onboarding for the updated builder is in design to help users learn the six-section structure. It is still at the design stage and not yet in the product.
