Key takeaways
- A chatbot answers. A RAG system answers from your documents. An agent reads, decides and acts in other systems (CRM, ticketing, WhatsApp, databases) — usually with a human approval step.
- Costs scale with actions, not conversations: a RAG support assistant for a mid-size business runs ₹5–25k/month in model and hosting cost; an agent that touches production systems needs guardrails, logging and review that cost more than the model.
- The highest-ROI first deployment in most Indian businesses is boring: lead qualification on WhatsApp within 60 seconds, 24/7, with hand-off to a human. Speed-to-lead beats cleverness.
- Never give an agent write access to anything you cannot roll back. Start read-only; add actions one at a time with logs and an "undo".
- Measure in business units — minutes saved per ticket, leads answered under 1 minute, revenue per agent-month — or you will not know whether it worked.
1. Definitions that survive contact with reality
Chatbot. A conversational interface over a model (or over rules). It knows what the model was trained on plus whatever you put in the prompt. Good for FAQs, tone, drafting. Fails when asked about your price list from last Tuesday.
RAG (retrieval-augmented generation). Before answering, the system searches your content — policies, product sheets, past tickets, SOPs — and gives the relevant chunks to the model. The answer is grounded and citable. The hard work is not the model; it is chunking, indexing, permissions and keeping the index fresh.
Agent. A loop: observe (read a ticket, an email, a monitoring alert) → reason → call tools (look up the order, check the server, create a Jira, send a WhatsApp) → observe the result → continue or stop. Agents are useful precisely because they act, and dangerous for the same reason.
2. Which one for which job
| Job | Right tool | Why |
|---|---|---|
| Answer "what are your timings / prices / policies" | RAG | Needs your current documents, not the model's memory |
| Qualify inbound leads on WhatsApp, book a call | Agent (narrow) | Must read CRM, write a lead, send a message, escalate |
| Draft replies for support staff to approve | RAG + human | Grounded drafts, zero risk of wrong sends |
| First-line IT support (password reset, VPN how-to) | RAG → Agent | Start with answers; add actions (unlock account) once trusted |
| Monitoring triage (acknowledge, correlate, open ticket) | Agent (read-mostly) | Reads Zabbix/Prometheus, writes tickets, never restarts production |
| Weekly reports from GA4/GSC/ads | Agent (scheduled) | Read-only data, writes a document |
| Marketing copy, blog drafts | Chatbot / LLM with style guide | No retrieval needed beyond brand voice |
3. What they cost (India, 2026)
| System | Typical monthly run cost | Build effort |
|---|---|---|
| FAQ chatbot on website | ₹1–5k (model API) | 1–3 days |
| RAG assistant over 500–5,000 documents | ₹5–25k (model + vector DB/hosting) | 2–4 weeks incl. permissions and evaluation |
| WhatsApp lead-qualification agent | ₹8–30k (model + WhatsApp Business API + hosting) | 2–4 weeks |
| IT-ops triage agent (read-mostly) | ₹15–40k | 4–8 weeks incl. runbooks and approvals |
Model cost is rarely the problem. The hidden costs are evaluation (did it answer correctly 95% of the time? prove it), monitoring, and the humans who review the first month.
4. Guardrails that are not optional
- Read before write. First deployment is read-only. Add one action at a time.
- Approval gates for anything irreversible (refunds, deletes, customer-facing promises).
- Full logs of every prompt, retrieval and tool call — you will need them for the first bad answer.
- Scope by identity. The agent acts with its own credentials and least privilege, never a human's admin token.
- Fallback to a human in under one minute, with context handed over, when confidence is low or the user asks.
- Evaluation set. 100 real questions with known-good answers; run them on every change.
5. A 30-day plan for a first agent
- Week 1. Pick one workflow with volume and a clear metric (e.g., inbound WhatsApp leads; metric: % answered within 1 minute, % qualified correctly).
- Week 2. Build RAG over your product sheets and FAQs; test with the evaluation set.
- Week 3. Add actions: create lead in CRM, send WhatsApp reply, escalate to sales. Log everything. Human approves every outbound message for 7 days.
- Week 4. Remove approval for low-risk replies; keep it for quotes and promises. Report the metric weekly.
If the metric does not move in 30 days, stop and pick a different workflow — do not add cleverness.
We build and run exactly these systems — WhatsApp lead agents, support triage, reporting agents — with the guardrails above. See AI agent automation.