Most business chatbots are bad. They are decision trees wearing an AI badge, and customers learn within two exchanges that the fastest route to an answer is to escape to a phone number.
The ones that work share a specific property: they answer from your actual documentation, admit when they do not know, and hand over to a human cleanly. Here is where that is worth building and where it is not.
Where chatbots genuinely earn their cost
Works well
- Repetitive factual questions — hours, pricing, policies
- Order and delivery status lookups
- Qualifying leads before a salesperson calls
- Answering from a large product catalogue
- After-hours cover when staff are offline
- Internal helpdesk for your own staff
Works badly
- Complaints and emotionally charged issues
- Complex negotiation or custom quoting
- Anything needing genuine judgement
- High-value B2B relationship conversations
- Situations where being wrong is expensive
- Replacing support entirely to cut headcount
The most common expensive mistake: deploying a chatbot to reduce support headcount, then discovering it deflects 30% of queries while making the other 70% angrier. Deploy it to handle volume your team cannot reach, not to replace the team.
The three tiers, honestly compared
| Type | Cost | What it does | Honest verdict |
|---|---|---|---|
| Rule-based widget | ₹0 – ₹20,000 | Fixed buttons and scripts | Fine as an FAQ menu; not AI |
| Generic AI widget | ₹1,500 – ₹8,000/mo | LLM with a scraped site index | Adequate; sometimes invents answers |
| Grounded custom agent | ₹80,000 – ₹3,00,000 | RAG over your verified content, connected to your systems | The only tier that reliably works |
What "grounded" actually means
The difference between a chatbot that helps and one that embarrasses you is architectural:
- Your content is indexed, not the model's memory Your policies, pricing, and product data are stored in a searchable index. The bot retrieves relevant passages before answering.
- The model answers only from what it retrieved It is instructed to use the retrieved passages and nothing else. If nothing relevant is found, it says so.
- Answers carry citations The customer can see which page the answer came from, which builds trust and makes errors traceable.
- A confidence threshold triggers handoff Below a set confidence, the conversation routes to a human with the full context attached.
- Everything is logged Every question and answer is reviewable, so you can find gaps in your documentation and improve it.
A chatbot that says "I'm not sure — let me get someone who is" is more valuable than one that confidently invents your refund policy.
Where to deploy it in India
Website chat is the default, but frequently not the highest-value channel here:
- WhatsApp — usually the highest-engagement channel for Indian consumers. Requires Business API approval and has per-conversation costs.
- Website widget — best for pre-sales questions from people already evaluating you.
- Internal Slack or Teams — often the fastest ROI of all, answering staff questions about your own policies and processes.
- Email triage — drafting replies for human approval rather than sending autonomously.
Running costs, realistically
Costs scale with usage, and caching common questions reduces them substantially. The recurring work people forget is content maintenance — a bot grounded in documentation is only as accurate as that documentation.
How to start sensibly
- Read your last 200 support messages Count how many are the same handful of questions. That number is your realistic deflection ceiling.
- Fix your documentation first A bot grounded in vague content gives vague answers. This step alone often reduces support volume.
- Launch narrow One category of question, on one channel. Prove it works before widening.
- Measure deflection and satisfaction together Deflection alone is a vanity metric if the deflected customers left frustrated.
Considering an AI agent for your business? Tell us what your team answers repeatedly — we will tell you honestly whether it is worth automating. See our AI agent development service for how we build these.