Building an AI Agent: Part 1 — Ideation and Planning

I have recently embarked on a journey of building an AI agent for a design agency’s website and thought I’d document the whole process from the learning to the agent's final (hopeful) deployment. It might take time since I will be learning and doing it on the fly while also working on projects and other commitments, so do excuse me if there is a long gap between articles! 🫣

Photo by Ales Nesetril on Unsplash

Let’s dive in!

AI agents are revolutionizing how businesses interact with their customers, and building one for your website can be a game-changer. This is the first part of a series of articles where I’ll document my journey in creating an AI agent for the agency’s website. In this article, we’ll focus on the beginning of this project, and every project that we approach — ideation and planning.

Why Build an AI Agent?

Building an AI agent stems from the need to enhance user experiences and streamline workflows. As an agency that focuses on product, UX and AI, it will also be sort of a showcase of the agency’s skills, while we learn the requirements of our users, and help them understand what we do and how we can help. For the agency, the goal is to:

• Improve Efficiency: Automate repetitive tasks like answering FAQs.
• Boost Engagement: Offer real-time assistance to website visitors.
• Showcase Innovation: Demonstrate our expertise in AI-driven solutions.

AI agents can handle tasks 24/7, ensuring users get the support they need without delay. Whether directing them to the right service, sharing a quote for a particular service, or answering common questions, the potential to save time and increase satisfaction is immense. Fingers crossed, this AI agent will be able to accomplish its goal. I am not expecting it to be perfect at the first shot, but I look forward to learning that will come with this.

Identifying Use Cases

Before jumping into development, defining what your AI agent will do is crucial. For our website, I identified these core use cases:

1. Answering FAQs: Automating responses to common queries like pricing, services, timelines, about the agency and work.
2. Navigation Support: Helping users find the right pages or resources quickly.
3. Lead Qualification: Engaging with potential clients to gather basic information before handing it to the team.

How did I arrive at these use cases? A little bit of research is still ongoing. This is what I am currently doing:

• Reviewing website analytics to see where users struggled.
• Talking to the team and the founder about frequent customer, client and business questions.
• A quick analysis of any other agencies and businesses with an agent/chatbot on their website.

A pro tip I learned from a few weeks of research is to start with simple use cases and expand over time. An AI agent which tries to do everything might end up doing nothing well.

Choosing the Right Tools and Technologies

I am currently exploring the tools and platforms that would best suit this project. This stage is all about identifying potential solutions that align with the project’s goals. Here are some options I’m considering, which you might find useful to explore as well:

Frameworks:

• GPT-4 via OpenAI API: For generating natural language responses with high accuracy and fluency.
• LangChain: For chaining tasks together and improving contextual understanding in conversations.
• Rasa: A customizable open-source framework for building conversational AI workflows.
• Hugging Face Transformers: Provides pre-trained models for various natural language processing tasks.
• Auto-GPT: Enables autonomous AI agents that can execute multi-step tasks independently.

Platforms:

• Dialogflow: A user-friendly platform for integrating conversational AI into websites and apps.
• IBM Watson Assistant: Offers robust enterprise-grade capabilities for creating AI-powered assistants.
• Microsoft Azure AI: Provides tools for building and deploying AI agents with cloud integration.
• Amazon Lex: Helps develop conversational interfaces using AWS’s machine learning services.
• Kore.ai: A platform designed for creating intelligent virtual assistants with advanced integrations.
• Google Cloud Conversational AI: Offers end-to-end solutions for building and deploying advanced conversational agents.

Other Tools to Explore:

• Runway ML: Simplifies creative workflows with AI-driven tools for video, image editing, and more.
• Teachable Machine by Google: A browser-based tool that allows anyone to train machine learning models without coding.
• Tidio: Combines live chat and chatbot functionalities, making it ideal for small businesses.
• Landbot: A no-code platform for building conversational chatbots with rich customization options.
• GaliChat: A user-friendly tool for deploying conversational agents quickly.
• YourGPT.ai: Focused on deploying custom AI agents to websites effortlessly.

These are some of the tools I came across while doing my research. There could be something I missed or upcoming. In that case, do let me know, so I can add it to the list of resources I can share once I finish this series.

Once I finalize the tools for our project, I’ll share my choices and the reasoning behind them in the next article. For now, these options provide a starting point for anyone interested in building an AI agent.

Have you built an AI agent or explored similar tools? I’d love to hear what worked for you. Are there any tools you’d recommend exploring? Drop your suggestions in the comments.

Setting Goals and Metrics for Success

Clear goals help measure your AI agent’s effectiveness. For our project, I’ve set the following:

1. Response Accuracy: The agent should answer 80% of user queries correctly.
2. Engagement Rates: Aim to increase time spent on the site by 15%.
3. Task Completion: Ensure the agent can guide users to the correct resource at least 85% of the time.

Tracking these metrics ensures we stay focused on delivering value. Tools like Google Analytics and built-in AI platform dashboards will help monitor performance.

Challenges and Considerations

Planning isn’t without its challenges. While I sat upon understanding how to build the agent, here are some hurdles I expected I would hit and how I am going to solve them.

• Identifying Relevant Data: Ensuring the agent has access to the right information for accurate responses.

Solution: I’m compiling FAQs and key resources into a centralized knowledge base. I am also scrapping all the information I can from the website which can be a great asset to ensuring the agent has answers to most of the users’ questions.

• Balancing Simplicity and Functionality: Avoiding feature creep while ensuring the agent meets user needs.

Solution: Starting with the most requested tasks and iterating based on feedback. I am keeping it simple for starters and as I make the agent, I will keep working on the ideas I might gain when I share the progress of this agent with the team.

User Expectations: Managing what users expect versus what the agent can realistically deliver. Let’s face it, it can’t do everything, right? We need to set a line somewhere.

Solution: Setting clear boundaries for the agent’s capabilities (e.g., “I can help with FAQs but not detailed consultations. For more please contact ABC.”).

Let’s get this agent started!

A well-planned foundation sets the stage for a successful AI agent. By defining clear use cases, exploring the right tools, and setting measurable goals, we’re ready to move to the next phase: design and architecture.

In the next part of this series, I’ll share how we’re structuring the agent’s workflows, the tools I plan to use to build the agent, and designing its conversational flow. I’ll also add the resources for all the research I am working on at the moment. Stay tuned!

I am always open to learning from this wonderful community. If you have built an agent before I would love to hear your thoughts on the way you approached the research, the actual building, conversational flows, the tools you used and the challenges you faced. Let’s learn from each other! :)