How to Build a Customer Support Function for Startups Using AI Tools with Alex Velinov

What Is an AI Agent?

AI agents are tools that use language models to perform tasks in real-time. Think of them as virtual assistants that can:

• Answer customer questions
• Search internal documentation
• Handle voice or video conversations
• Help sales teams

Depending on how you deploy them, AI agents can be used for internal knowledge sharing, customer-facing support, or outbound sales.

Why AI Agents Matter for Startups

For a startup, time and bandwidth are everything. AI agents let you extend your team without hiring, allowing your startup to punch above its weight when it comes to:

• Handling inbound support requests
• Providing instant answers 24/7
• Reducing repetitive tasks for your team
• Improving customer experience and speed of resolution

They can also support in multiple languages, route requests based on sentiment—such as escalating if the agent detects confusion or frustration in the user’s language, and scale up as your customer base grows.

If you want to know how to also automate sales outreach, check out our guide on how to build an automated sales outreach engine with ColdIQ.

Types of AI Agents: 3 x 2 Framework

There are two primary ways to classify AI agents: by who they serve and how they operate.

By Function:

1. Internal Agents — Used by your team.
   
   
   • These can act as Virtual Assistants for your employees, trained to navigate internal knowledge bases and company policies.
     
     
   • Key functions include:
     • Accessing the internal knowledge base
     • Searching company policies
     • Answering FAQs from employees
     • Searching and indexing internal documents
       
       
   • These agents reduce internal friction and help new hires onboard faster while decreasing dependency on managers for basic information.
     
     
2. External Agents — Customer-facing AI agents.
   
   
   • These are your front-line customer support assistants, operating 24/7 with consistency.
     
     
   • Capabilities include:
     • Automated query resolution
     • Handling detailed FAQs
     • Providing rich knowledge on your business, products, and services
     • Supporting multiple languages
     • Escalating conversations based on sentiment or confusion.
       
       
3. Custom Agents — Agents built for a specific role or goal.
   
   
   • A great example is a Sales Agent, tailored to:
     
     
     • Pre-qualify leads through dynamic questioning
     • Perform deep research on companies and people using ChatGPT. Use it in deep research mode to extract detailed, structured information about a target company—its products, services, customers, pricing models, and more. The included prompt template walks you through exactly what to ask: begin with the company name (and website, if available), then prompt ChatGPT to deliver information across multiple categories including features, use cases, pricing, support docs, and case studies. This can power sales agents with high-context personalization and pre-call intelligence.
     • Personalize emails and outbound sequences
     • Provide real-time coaching during sales calls
     • Analyze calls and deliver actionable recommendations

By Interface:

1. Text-based Agents — Chatbots or embedded widgets on your website.
2. Voice-based Agents — Phone-based support powered by tools like Eleven Labs.
3. Video Agents — Synthetic avatars capable of visually interacting with prospects or customers.

To create a voice agent with a specific accent, you need at least 2 minutes of recorded audio. But for best accuracy, 60 minutes of recorded speech is recommended.

Real-World Examples

• Internal Agent: Apollo, a digital employee trained on 300+ internal documents at Tag Digital, answers HR, brand, and operations questions 24/7. It was built using Chatbase to ingest and structure the company’s internal knowledge, allowing the team to interact with it through a custom chatbot interface. This made it possible for employees to access detailed internal info instantly without pinging HR or senior staff.
• External Voice Agent: A voice bot that answers FAQs and escalates based on frustration, trained with Eleven Labs and embedded on a website.
• Custom Video Agent: A cloned digital twin of a CEO that answers common sales questions, powered by generative video using Synthesia.io. This way, a busy CEO can now talk with hundreds of clients that couldn’t before.

Building Blocks of a Great AI Agent

Live Demo Recap: Tools Used (shown at minute 27:00 in the webinar)

• ChatGPT for company/product data scraping and FAQ generation
• Markdown formatting for structured knowledge ingestion
• Chatbase for building and testing the chatbot interface
• Zapier to automate lead capture and route to CRMs like HubSpot
• Midjourney for creating a profile avatar image

Before jumping into implementation, it’s important to set realistic expectations. AI agents won’t replace your entire support team on day one, but if they can resolve 80–90% of incoming queries accurately, you’ve already achieved a massive efficiency win.

1. Knowledge Base
   • Start with internal docs, SOPs (Standard Operating Procedures—which are step-by-step instructions for how your company handles repeatable processes), product info, and especially FAQs.
   • If you don’t have an FAQ, don’t worry, create one using transcripts from your support calls. Tools like NotebookLM from Google are perfect for this, especially when dealing with proprietary or sensitive data. Unlike traditional chatbots, NotebookLM keeps your content private and isn’t used for model training, making it a secure option for handling internal documents or unannounced product material.
   • Bonus: Publish your FAQ on your website. It not only helps your customers, but also improves your site’s SEO and answer engine optimization.
     
     
2. Platform Choice
   • Use tools like Chatbase—it lets you launch a working agent in under an hour and is ideal for testing out your first support workflows quickly.

Don’t wait to be perfect. Build a simple FAQ-driven agent, test, learn, and improve.

Additional Resources

We’ve included a PDF of prompt templates to help you apply everything covered in this guide. Inside, you’ll find:

• A company deep research prompt for ChatGPT or Perplexity
• A format for generating FAQs from your team or support transcripts
• A Midjourney prompt to create a custom AI avatar for your support agent
• Main prompt instructions to define your agent’s personality, tone, and behavior

Use these templates to jumpstart your implementation.

Download our bonus document with sample ChatGPT prompts you can use to:

• Extract FAQs from call transcripts
• Generate sales research on prospects
• Build custom conversation flows

These templates help ensure your agent is built on accurate, structured, and actionable data—whether you’re supporting customers, training your sales team, or improving internal workflows.

Bonus: How to Evaluate Your Agent’s Responses

Once your agent is live, you can test it using a set of sample questions. Start by collecting 50–100 common customer or internal questions, then feed them one by one to your AI agent and record the responses. Copy these outputs into a spreadsheet or document.

Next, use a large language model (LLM)—like ChatGPT—as a neutral evaluator. Give the LLM each question, the agent’s answer, and your knowledge base or FAQ as reference. Ask it to score the agent’s response using a rubric:

• Accuracy: Is the information correct?
• Clarity: Is the language easy to follow?
• Completeness: Does it fully answer the question?
• Tone: Is the tone friendly, on-brand, and appropriate?
• Adherence to Source: Does it stick strictly to your knowledge base?

Have the LLM assign a score (1–5) for each criterion, and optionally add a comment. Use the results to identify weak spots in your dataset, clarify vague instructions, or add new training examples.

Running these evaluations regularly ensures your agent is learning and improving without drifting from your core information.

Key Takeaways

By this point, you’ve seen how to design, build, and deploy AI agents across support and internal ops. Whether you’re doing this to free up your time, reduce team strain, or just scale smarter—here are the essential principles to carry forward.

• AI agents help startups deliver faster, better customer service without hiring a support team.
• Start with what you already have: internal docs and support transcripts. Use that material to build an FAQ, which serves as both your agent’s foundation and a helpful SEO asset.
• Use tools like Chatbase (for chatbot creation), Eleven Labs (for voice capabilities), and NotebookLM (for private document analysis) to build and deploy your first agent.
• Focus on accuracy and relevance before scaling complexity.
• A strong FAQ doubles as a support tool and a traffic driver.