AI Chatbot for Your Product: Build vs. Buy vs. Skip
Learn the decision framework for when to build custom, buy off-the-shelf, or skip AI chat entirely for your startup.
November 22, 2024 5 min read
Every founder we talk to right now wants to add an AI chatbot. Investors love it. Users expect it. Competitors are doing it.
But here's what we've seen after integrating AI chat into a dozen products over the past year: about half of those chatbots actively hurt the user experience. They answer questions wrong, frustrate users with hallucinations, and create support tickets instead of resolving them.
The chatbot decision isn't "how do we add AI?" It's "should we add AI, and if so, what kind?" For a broader perspective on AI feature decisions, see our guide on when AI is overengineering your MVP.
The Real Question: What Problem Are You Solving?
Before touching any AI tooling, answer this: What specific user problem does a chatbot solve better than alternatives?
"Users can ask questions in natural language" isn't a problem statement. It's a feature description. The problem statement might be:
Users can't find documentation answers. Maybe. Or maybe your documentation structure is the real issue.
Support volume is overwhelming the team. Possibly. But chatbots that give wrong answers create more tickets, not fewer.
Users need complex data queries. Now we're talking. Natural language to structured queries is a legitimate AI use case.
Onboarding is too complicated. Perhaps. Or the product itself needs simplification.
Get specific. If you can't articulate the problem in terms of user pain and business impact, you're building a chatbot because it's trendy, not because it's useful.
The Build Option: When Custom Makes Sense
Building a custom AI chatbot means integrating directly with an LLM provider (OpenAI, Anthropic, or open-source models), handling conversation management, and creating your own retrieval and response systems.
Stop planning and start building. We turn your idea into a production-ready product in 6-8 weeks.
When to Build Custom
You have proprietary data that defines the value. If your chatbot's usefulness depends on understanding your specific product, codebase, or knowledge base, you need fine-tuned retrieval or specialized prompting that off-the-shelf solutions won't handle well.
Conversation flow is core to your product. If the chat interface is the product (an AI writing assistant, a customer service automation tool), you need full control over every aspect of the experience.
You're integrating deeply with existing systems. When the chatbot needs to read from and write to your database, execute transactions, or interact with internal APIs, custom integration is often cleaner than wrestling with third-party webhooks.
You need complete cost control. Third-party platforms charge per conversation or per message. At scale, building your own can be 3-10x cheaper on LLM costs alone. See our guide on LLM cost optimization for specific strategies.
What Building Custom Actually Takes
For a production-quality custom chatbot with Next.js and Convex:
2-3 weeks for basic implementation. Streaming responses, conversation history, basic error handling.
2-4 weeks for retrieval (RAG). If you need the chatbot to answer questions from your documentation or data, add time for chunking, embedding, and retrieval tuning.
Ongoing prompt engineering. The first version will give mediocre answers. Plan for 2-4 iterations of prompt refinement based on real usage.
Moderation and safety. Content filtering, rate limiting, abuse prevention. This isn't optional for production.
We use the Vercel AI SDK for streaming LLM responses in Next.js applications. It handles the streaming protocol cleanly and integrates well with both OpenAI and Anthropic. For conversation storage and retrieval, Convex provides real-time sync that keeps multi-device conversations coherent without extra infrastructure.
The Buy Option: When Off-the-Shelf Works
Plenty of platforms offer embedded chatbots: Intercom, Zendesk, Drift, and dozens of AI-specific tools. These make sense in specific situations.
When to Buy
You need a chatbot within days, not weeks. If there's genuine urgency (a launch deadline, a sales requirement), buying gets you something functional fast.
Your use case is well-defined and common. Customer support for documentation-based questions is a solved problem. If your needs match the platform's sweet spot, you'll get a polished experience without custom development.
You're testing the concept before committing. A $99/month chatbot subscription is cheaper than two weeks of engineering time. Use it to validate whether users actually engage with chat before building custom.
You don't have engineering capacity. If your team is at capacity on core product work, outsourcing a non-core chatbot makes sense.
The Tradeoffs of Buying
Limited customization. You're constrained to the platform's UI patterns, conversation flows, and integration options. Complex workflows often require awkward workarounds.
Recurring costs that scale with usage. Most platforms charge per conversation or per resolution. A chatbot handling 10,000 conversations monthly can cost $1,000-5,000/month. The math changes fast.
Vendor lock-in. Migrating conversation history, retraining on new platforms, and rebuilding integrations creates real switching costs.
Black box behavior. When the chatbot gives bad answers, debugging is limited to whatever logging the vendor provides.
For most startups testing chat-based support, Intercom Fin or similar tools provide reasonable quality without engineering investment. Just go in with realistic expectations about customization limits.
The Skip Option: When AI Chat Hurts More Than Helps
Sometimes the right answer is no chatbot at all. We've talked clients out of AI chat when:
Your Data Quality Won't Support It
Garbage in, garbage out. If your documentation is outdated, incomplete, or contradictory, the chatbot will confidently present that garbage to users. Fixing documentation is often higher ROI than adding a chat layer on top of broken content.
Users Need Certainty, Not Probability
Chatbots give probabilistic answers. They're usually right, but they're never guaranteed right. For products where incorrect information creates liability (medical, legal, financial), the risk of hallucination may outweigh the benefit of natural language interaction.
The Problem Is Discovery, Not Answers
If users struggle because they don't know what questions to ask, a chatbot doesn't help. They need better navigation, improved information architecture, or contextual help—not a text box.
Support Volume Is Low
If you're handling 50 support requests a week, a chatbot is premature optimization. A shared inbox and better documentation will serve you better until scale demands automation.
You Can't Commit to Maintenance
Chatbots aren't set-and-forget. They need prompt tuning, content updates, and ongoing monitoring. If you can't allocate 2-4 hours weekly to chatbot quality, it will degrade and create negative user experiences.
Document automation can cut drafting time from 3 hours to 15 minutes. But most MVPs fail by building too much too soon. Here are the 5 features that actually matter.
