Build vs Buy an AI Chatbot for SMBs

Most small businesses that call Agentic Academy Labs from India should buy a SaaS chatbot first. We say that even though we sell custom work. Buy when FAQs are public and integrations are standard. Build when wrong answers hit revenue or compliance, or when the assistant is the product. From Sikar we have watched shops pay for a custom RAG stack they never update, and we have watched others outgrow a website widget in a quarter. The checklist below is how we decide in discovery, not a pitch for either side.

Signals we use in the first call

What week 1 looks like when we advise 'buy'

1. List the twenty questions staff already answer on WhatsApp. If they are all on the website, a SaaS crawler is enough.
2. Pilot one vendor on a staging domain. Measure deflection and the questions it still escalates.
3. Assign a human owner for the FAQ. Unowned bots rot in two months.
4. Set a calendar reminder to revisit build vs buy after a quarter of real tickets.

What week 1 looks like when we advise 'build'

1. Discovery: corpus, tenants, channels, eval questions. See AI MVP cost and timeline.
2. Thin RAG slice on Next.js, Bun, and Postgres, not a multi-agent demo. RAG vs fine-tuning.
3. One channel that staff already live in. For many Indian SMBs that is WhatsApp or Slack, not a new URL.
4. Admin upload so a non-engineer can fix a wrong answer tomorrow morning.

Cost drivers without fake INR quotes

SaaS is a subscription you can cancel. Custom is people time plus model tokens plus hosting. We keep those lines separate. A built bot that nobody maintains becomes more expensive than the SaaS you rejected. Interns can staff eval scoring; they cannot be the only owner of a customer-facing bot. If you build, budget corpus cleanup. That was the expensive surprise in our Bangalore fintech case study, and SMBs hit the same wall with GST FAQs and rate cards.

Failure modes

• Buying a widget, then pasting private price lists into a vendor that trains on your data. Read the DPA.
• Building custom because a competitor announced 'AI,' then shipping with no eval set.
• Expecting a chatbot to update CRM. That is an agentic workflow, with approvals.
• Flutter app and website each talking to a different bot. Customers get two answers.

A simple decision rule we use

• If a wrong answer is embarrassing but not costly: buy, measure, revisit.
• If a wrong answer is a compliance or refund event: build RAG with citations and logs.
• If success is 'create the ticket' not 'explain the policy': you need tools and gates, not a smarter FAQ.

Agentic Academy Labs will tell you to buy when that is the honest answer. When it is not, we build under custom AI development and IT consulting. Schedule a call with the twenty real questions in a spreadsheet.