What an AI chatbot on your website can and cannot do in 2026
An assistant that answers visitors from your own content – service pages, brochures, a rate card – typically costs Rs. 90,000 to Rs. 2,50,000 to build in India in 2026, and after that it costs money per conversation rather than as a flat monthly fee. That cost moves with your traffic and with how long each answer is, so it is a number you design rather than one you are handed.
Looking for numbers first? Our AI chatbot cost breakdown for India lists what each type costs to build and to run every month. This guide covers how the technology works.
What these assistants do well is narrow and genuinely valuable: answering the same forty questions your team answers every week, at 11pm, in whatever mix of English and Hindi the visitor types. What they do badly is anything that needs judgement, negotiation, or a commitment made on your behalf.
The failure nobody forgets is not a bot that says it does not know. It is a bot that confidently quotes a price you withdrew in March, to a customer who takes a screenshot. Most of this guide is about engineering that away before you switch the thing on.
Rule-based flows versus an LLM assistant
These are two different products wearing the same chat bubble, and choosing the wrong one is the most expensive mistake on this list.
What a rule-based flow is
A rule-based bot walks the visitor down a decision tree you drew: buttons, branches, a fixed answer at the end. It cannot invent anything, because it was never asked to generate anything. It is cheap, instant and predictable, and it costs the same at any volume.
Its limit shows the moment a visitor types something you did not anticipate. Then it either loops, or dumps them into a contact form, and they leave.
What an LLM assistant is
An LLM assistant reads the question in plain language, pulls the most relevant passages out of your own content, and writes an answer from those passages. It copes with phrasing you never thought of and with follow-up questions. It also bills you per conversation, and it will happily produce a fluent wrong answer if you ground it badly.
Which one you actually need
If your top twenty questions have fixed answers – timings, location, documents required, how to track an order – a rule-based flow is the correct answer and anything cleverer is waste. If visitors arrive with messy, varied, comparison-shaped questions about a catalogue or a specification, a grounded assistant earns its cost.
In practice most sites want both: a flow for the four things everyone asks, with a free-text assistant behind it for everything else. Build the flow first. It is the cheaper half, and its logs tell you exactly what the assistant will need to know.
| What you are comparing | Rule-based flow | Grounded LLM assistant |
|---|---|---|
| Unexpected phrasing | Loops or exits to a form | Handles it; that is the point |
| Can it invent an answer | No | Yes, unless grounded and constrained |
| Cost per conversation | Effectively nil | Metered; varies with traffic and answer length |
| Maintenance job | Redraw the tree when a process changes | Refresh the index whenever content changes |
| Best for | Timings, location, documents needed, order tracking | Catalogue questions, comparisons, document lookups |
Grounding: the difference between useful and embarrassing
Grounding means the assistant is only allowed to answer from text you supplied. Your pages are chopped into passages and indexed; at question time the system retrieves the few passages that match, and instructs the model to answer from those and nothing else. If the passages do not contain the answer, it says so and offers a human.
Without grounding, the model answers from general knowledge about businesses like yours. That is precisely how a bot ends up promising a same-day service you do not offer.
What to feed it
Start with what already exists: services, pricing, FAQ, delivery and returns, contact. Then add the documents your team keeps re-sending – the rate card PDF, the specification sheet, the onboarding checklist. Then add the answers your team types into email every week that were never published anywhere. Those are usually the most valuable content you own.
Keeping the index honest
A grounded assistant is only as current as its index. Change a price on the website, let the index rebuild next month, and the bot spends a month being wrong with total confidence. Decide who owns the refresh and what triggers it: every publish, every week, whatever matches how often your content moves.
Two habits prevent most disasters. Never index a document whose version or effective date the assistant cannot see. And make every answer show a link to the page it came from, so the visitor can verify it and you can debug it.
Contradictory sources are the quiet killer. If two pages state different prices, retrieval will sometimes fetch one and sometimes the other, and you get intermittent wrong answers that are painful to reproduce. Fix the content, not the prompt.
What it actually costs to run
There are three separate costs, and businesses routinely budget for only the first.
The build
One-time: content preparation, indexing, the prompt and refusal rules, the widget, handover wiring, testing. Content preparation is usually the biggest slice, not the code. A site with tidy, current pages is far cheaper to fit an assistant to than one carrying six years of drift.
The metered model cost
Every answer is billed by the model provider on the text going in and the text coming out. Text in is the retrieved passages plus your instructions plus the conversation so far; text out is the answer. Providers charge per unit of text and those rates change, so treat any figure anyone quotes you as a moving rate, not a fixed price.
The arithmetic is simple and worth doing before you commit: cost per answer, times answers per conversation, times conversations per month. What moves it:
- Answer length. A 400-word answer costs several times a 90-word one. Cap it. Short answers also convert better.
- How many passages you retrieve. Feeding ten passages to feel safe costs multiples of feeding three good ones.
- Conversation history. If every turn resends the whole chat, a long conversation gets steadily more expensive. Cap the history you carry forward.
- Model choice per step. Routing, classifying and in-scope checks can run on a small, cheap model; only the final answer needs the capable one.
- Traffic, and abuse. A public chat box is a public API. Without per-session and per-IP limits, one scraper can burn a month of budget over a weekend. This is the control people forget, and it is the one that turns a predictable bill into an unbounded one.
Maintenance
Reading the unanswered questions, refreshing the index, fixing contradictory content, tightening refusals. Budget it monthly, in the same conversation as the rest of your website maintenance budget. An assistant nobody maintains decays faster than a website does, because your business keeps moving and the index does not.
| What you are building | Typical India build cost, 2026 | Typical calendar time |
|---|---|---|
| Rule-based flow on an existing chat widget | Rs. 25,000 to Rs. 75,000 | 1 to 2 weeks |
| Grounded assistant over 20 to 60 pages and a few PDFs | Rs. 90,000 to Rs. 2,50,000 | 3 to 6 weeks |
| Assistant wired into live data: orders, bookings, CRM, WhatsApp handover | Rs. 3,00,000 to Rs. 8,00,000 | 6 to 12 weeks |
| Ongoing content refresh and tuning | Rs. 5,000 to Rs. 25,000 per month | Continuous |
These are what the Indian market charges in 2026, not a quote. Scope, integrations and the state of your content move them a long way, as on any custom software build. The metered model cost sits on top.
Handover: to a human, and to WhatsApp
Handover is not the failure path, it is the product. An assistant that answers twelve easy questions and passes the thirteenth to a person with the full transcript attached has done its job perfectly.
To a human
Three things to decide. What triggers it: the visitor asks, the assistant cannot ground an answer, or the question touches money, complaints or anything account-specific. What happens outside business hours, so nobody is left typing into a void. And what the agent sees when they pick it up – the full transcript, always. Making an irritated customer repeat themselves is worse than having no bot at all.
To WhatsApp
In India the conversation usually wants to go to WhatsApp anyway. A continue-on-WhatsApp button with a prefilled message moves the thread to a channel the customer actually checks, and hands you a phone number instead of an anonymous browser session.
Two mechanics are worth knowing before you promise this to anyone. First, when a customer messages your business first, a 24-hour service window opens in which your team can reply freely; once it closes you need a Meta-approved template message to reopen contact. Second, WhatsApp message charges are set by Meta and billed by Meta to your own WhatsApp Business Account, so your CRM cost and your messaging cost are two separate lines in the budget – a distinction that also applies to ordinary CRM software pricing in India.
Our own WhatsApp CRM, ZupiChat, runs on the official WhatsApp Business Platform. Starter is free forever; Growth is Rs. 999/month for 3 WhatsApp numbers and 3 team members; Enterprise is Rs. 2,499/month with unlimited agents. There is no free trial – the free plan is the trial – and API access is on every plan, which is what a chatbot handover needs.
What your chatbot should refuse to answer
A refusal list is a business document, not a technical one, and it is far cheaper to write than to regret. Put in writing, before launch, that the assistant does not:
- Quote a final price for anything that has to be scoped. Give the published range and hand over.
- Commit to a delivery date, a discount, or an exception to your policy.
- Give legal, tax, medical or financial advice, including questions about how GST applies to a transaction.
- Discuss a specific customer’s order, invoice or account unless that person has actually been authenticated. A chat box is not proof of identity.
- Compare you against a named competitor.
- Answer anything outside your business at all, however politely it is asked. Public assistants get used as free general-purpose chatbots, and you pay for every one of those answers.
Write the refusal wording yourself, because the default wording loses leads. Something like “I cannot quote that here, but I can get you an exact figure – shall I pass this to the team?” keeps the conversation alive. A flat “I cannot help with that” ends it.
Capturing leads without blocking the answer
The most common self-inflicted wound is demanding a name and phone number before the assistant says anything useful. Visitors read that as a contact form in a costume, and close it.
Answer first. Ask afterwards, once, when there is a reason to ask: a quote, a callback, a document to send. Ask for one field, not four – in India a phone or WhatsApp number is worth more than an email address. Then continue either way. A visitor who declines and still reads three answers is a warmed lead you can retarget.
When a lead is captured, send the transcript with it. A sales call that opens with a reference to bulk pricing for 200 units is a completely different call from a cold one. Wire this into the same pipeline as your existing forms rather than a separate inbox nobody opens; that small piece of custom software work is often what decides whether the project pays for itself.
Measuring whether it deflects work or just annoys people
Resist the vendor dashboard’s favourite metric, conversations handled, which happily counts a visitor asking the same thing three times and giving up as three successes.
Track these instead:
- Clean containment. Conversations that ended without a handover and without that same person contacting you by another channel within a few days. The second half is what makes the number honest.
- The unanswered log. Every question the assistant could not ground. This is your content roadmap, and it is the single most valuable output of the whole system.
- Handover rate, and what happened next. A rising handover rate is not automatically bad; it can mean the easy questions have stopped reaching your team.
- Drop-off point. Where in a conversation people close the window. That is usually a wording problem, not a model problem.
- Leads by source. Tagged, so you can eventually argue about revenue rather than chat volume.
Then read transcripts. Thirty to fifty real conversations a month, by someone who knows the business. No dashboard tells you the assistant is technically correct and reads like a cold form letter. Feed what you find back into your pages: the questions people ask a chatbot are the questions they type into Google, so an unanswered log is free keyword research for your website.
Data and privacy to settle before you switch it on
India’s Digital Personal Data Protection Act, 2023 governs personal data you collect, and a chat transcript carrying a name and a phone number is personal data. Settle these before launch rather than after a complaint:
- Notice and consent. Say in the chat window itself that the conversation is stored and why, with a link to your privacy policy.
- Retention. Decide how long transcripts live, and delete them on schedule. Keeping them forever is also a decision, and a poor one.
- Deletion requests. Someone will ask you to erase their data. Know in advance how you locate and remove their transcripts, including backup and analytics copies.
- Who your processors are. The model provider and the chat platform both see the text. Read their terms on whether your inputs are used to train models, and pick a tier where they are not.
- What must never enter the box. Instruct the assistant to refuse card numbers, Aadhaar, PAN, passwords and OTPs, and mask phone numbers in your logs. If a customer pastes a card number anyway, you are suddenly holding card data you never wanted.
- Read-only connections. If the assistant queries your database to answer order questions, give it a scoped, read-only view. Never a connection that can write or delete, and never one that returns another customer’s record just because an ID was changed in the request.
- Staff access. Transcripts contain customers’ problems. Decide who inside the company can read them.
Publish a contact for data complaints on your privacy page while you are there; you want a named human for that regardless.
Where to start this month
Export the last three months of enquiries – email, WhatsApp, the contact form – and count the questions. If twenty questions cover most of the volume and each has one fixed answer, build a rule-based flow and put the savings elsewhere. If the questions are varied and specific to your catalogue or documents, a grounded assistant will earn the larger build.
Either way, do the content before the chatbot. Publish those answers as real pages first: pages that answer real questions earn search traffic on their own, and the assistant then becomes a second use of work that was already paying for itself. Codelith Lab is based in Pune and works with businesses across India. If you want this scoped properly – flows, grounding, WhatsApp handover, refusal rules and a privacy review in one pass – see how we approach AI and automation projects or send your enquiry pile through the contact page and we will tell you which half you actually need.
