04/06/2026
Most teams overpay for AI chatbots. We didn't.
Here's the 6-week playbook we used to ship our own Gen AI chatbot for ~$25/mo:
Week 1. Define goals and use cases
Start with the problem:
- Customer support?
- Lead qualification?
- Onboarding?
Lock the channels (web, WhatsApp, Slack), the audience, and the 3 metrics that actually matter: resolution rate, CSAT, and cost per conversation.
Week 1–2. Conversation design
This is the step everyone skips and it's the one that decides whether users like your bot.
Map the user journeys, set the tone, plan fallbacks, and define exactly when the bot should escalate to a human (low confidence, frustration signals, sensitive topics, explicit "talk to a human" requests). Build for v1, not for the perfect vision.
Week 2. Choose your stack
Four real decisions:
- LLM provider (GPT, Claude, Gemini, or open-source),
- framework (Vercel AI SDK, LangChain, Rasa, or custom),
- data layer (RAG vs tool-augmented / MCP),
- hosting (cloud vs on-prem if HIPAA/SOC 2 applies).
Week 2–4. Build the core bot
Intent recognition, response generation, integrations with your CRM/CMS/calendar, lead qualification flows, and multi-turn context.
If it's an MVP, ship 1–2 use cases first - don't try to solve everything.
Week 3–4. Security & compliance
Auth, rate limiting, input validation, and encryption from the start, not bolted on later.
For regulated industries add GDPR consent, HIPAA-compliant data handling, and audit trails. This step alone can add 30–50% to the timeline.
Week 4–5. Testing
Chatbot testing is not standard QA.
Test conversation flow regression, adversarial prompt injection, hallucination spot-checks, and role adherence. Automate the adversarial tests - manual testing will miss the edge cases.
Week 5–6. Deploy and monitor
Roll out in stages: internal team → beta users → production.
Wire up observability with Langfuse, LangSmith, or Arize so you see every conversation, every token, every cost.
Ongoing. Optimise and iterate
Weekly: review resolution rate, fallback rate, CSAT, cost per chat.
Monthly: deep-dive logs, A/B test response strategies.
Quarterly: audit knowledge base, review security logs, decide whether to expand scope.
Want the full build doc with our exact stack and security checklist?
Drop a "+" in the comments.