Building an AI agent · Guide 3 of 9

How to build an AI chatbot that isn't a chatbot wearing a hat

Building an AI chatbot in 2026 is no longer setting up a Q&A box. It's deciding whether you want a "marketing" chatbot (answers what you wrote), a "support" one (RAG over your knowledge base) or an "agent" (decides and executes actions). The three options have different architecture and price.

See it in production → 24/7 AI Support Back to the pillar

The 3 types of chatbot and which one you need

Type	What it does	When to use it
Marketing / FAQ	Replies with answers you wrote in advance	Small site, stable FAQs, low budget
Support (RAG)	Pulls info from your KB and answers with context	Support with a solid KB, >500 repetitive tickets/month
Agent	Decides and executes actions (not just replies)	Complex processes, system integration

Before you choose, rule out the obvious shortcut: a custom ChatGPT GPT works for none of the three if the destination is your own site, because it can't be embedded outside ChatGPT. How far it does go and where it breaks is laid out in building an agent with ChatGPT.

RAG architecture step by step

Structure your KB. Short blocks (200-400 words), self-contained, with metadata.
Generate embeddings. Turn each block into a vector with a model (OpenAI ada or equivalent).
Store in a vector database. Pinecone, Weaviate, pgvector depending on scale and budget.
Configure retrieval. For every query: pull the top-k most similar blocks, re-rank, filter by metadata if relevant.
Pass context to the LLM. The LLM gets the retrieved context + the user prompt + the system prompt.
Return the answer with a source citation (optional but recommended).

When the chatbot should hand off to a human (criteria)

Low model confidence (answer probability below threshold).
Cases flagged as sensitive (complaints, escalations, discount requests).
Explicit user request ("I want to talk to a person").
Number of turns without resolution — after 5-6 messages with no progress, escalate.
Frustration detection — keywords like "this isn't working", "I'm fed up".

Measuring CSAT and deflection without lying to yourself

Same thing as in support: inflated deflection is the trap. Measure:

Tickets resolved without escalation with confirmed resolution.
Post-interaction CSAT (>4/5 is healthy).
Clean escalation rate (every escalation must carry context).
CSAT before vs. after the chatbot.

Recommended stacks (self-serve, mid, enterprise)

Size	Stack	Monthly cost
Self-serve	Implementa Support AI, Intercom Fin, Chatbase	$79-300/mo
Mid-market	Zendesk + custom RAG + integration	$1,500-5,000/mo
Enterprise	Proprietary platform or Salesforce + custom agent	$5,000-25,000/mo

Frequently asked questions

If you get under 100 visits/day, it probably doesn't earn back the setup effort (it takes you more time to configure than it saves). From 500-1,000 visits/day with qualified traffic, yes — a lot. The honest metric isn't "traffic" but "repetitive queries/month": more than 50, it pays.

Self-serve: €0 extra beyond the subscription (€79-149/mo at Implementa). Self-hosted with your KB: €0 license, ~€20-50/mo in LLM tokens for reasonable volume. The real fight is content updates — that takes a human 2-4 hours/month so it doesn't go stale. That's a hidden cost.

Yes, modern LLMs handle multi-language without extra effort. Caveat: your KB has to be in at least one well-written language (English or neutral Spanish best). If your source content is low quality, the problems replicate across every language — amplifies them, doesn't fix them.

Free AI Impact Plan

The guide is generic. Your plan isn't.

Tell us about your company and we'll ship back a diagnosis with priorities, numbers and what to implement first. No sales call, no charge.

Keep going with these

Building an AI agent: the guide you should have read before the first prototype How to build an AI agent for sales (a real AI SDR)How to build an AI SDR: full architecture of automated outbound