Most businesses I speak to have already tried a chatbot. Usually it was a menu tree pretending to be a conversation — "Press 1 for orders, press 2 for returns" — and it annoyed customers enough that someone quietly switched it off.
That experience has made a lot of owners sceptical, and fairly so. But the technology underneath changed substantially in the last two years, and the gap between a bot that helps and a bot that irritates is now mostly about how you build it, not what you build it with.
We run an AI chatbot on our own site. It answers questions about pricing, checks availability and books consultation calls. Building it taught us more about where these things break than any vendor demo would have, so this is written from that side of the fence.
What actually changed
Older chatbots matched keywords against a script. If a customer phrased their question in a way nobody anticipated, the bot fell over. That is why the experience was so poor — real customers do not phrase things the way the script author imagined.
Modern systems understand the intent behind a message rather than matching words. Someone asking "my order hasn't turned up and I'm going away Friday" gets understood as a delivery query with urgency, without anyone scripting that sentence.
That is a genuine improvement. But it introduces a new failure mode that matters more than the old one, and it is the thing most businesses deploy without thinking about.
The problem you must solve first
These systems are built to be helpful. Ask one a question it does not know the answer to, and its instinct is to produce a plausible answer rather than admit ignorance.
For customer support, that is dangerous. A bot that invents a return policy, quotes a price you do not charge, or promises a delivery date you cannot meet has not saved you a support ticket. It has created a customer expectation you now have to break.
We hit this directly. Early on, our own assistant would answer pricing questions from general knowledge — plausible numbers, confidently delivered, and not our prices. A prospect anchored on a wrong figure is worse than one who got no answer at all, because now the conversation starts with a correction.
The fix is architectural, not a matter of instructions. We stopped putting prices in the bot's instructions and instead gave it a lookup into our actual pricing data, with a rule that it must check before stating any figure. It can only repeat numbers that genuinely exist in our system. Not "please don't guess" — structurally unable to.
Apply this to whatever matters in your business. Stock levels, delivery timelines, warranty terms, order status. If a wrong answer costs you money or trust, the bot should be reading it from your system at the moment it answers, not reciting something from its training.
Where chatbots genuinely help
Questions you answer constantly. Business hours, delivery areas, return windows, order status. In most businesses a handful of questions make up the bulk of support volume, and they are all answerable from data you already hold.
After hours. A customer with a question at 11pm either gets an answer or goes to a competitor. This is often where the value is largest and it costs nothing extra.
Qualifying enquiries. For service businesses, most incoming enquiries are not ready to buy. A bot that gathers what someone needs, their budget range and their timeline — and passes on a summary rather than just a name and number — changes what your sales team spends its day on.
Routing. Getting a customer to the right person with context already gathered removes the "let me transfer you" cycle that customers hate.
Where they do not
Angry customers. Someone whose order is three weeks late does not want a bot. They want a person who can fix it and apologise. Detect frustration and hand over immediately — with the conversation history attached so the customer does not repeat themselves.
Anything irreversible. Cancelling orders, issuing refunds, changing account details. You can let a bot prepare these actions, but a human should confirm.
Genuinely complex problems. If it needs judgement, it needs a person.
Regulated advice. Financial, medical, legal. Do not let a bot near it.
The businesses that get value here are clear-eyed about the split. The bot handles the routine 60–70%. Everything else goes to a human, fast, with context.
What it costs
Two components: building it, and running it.
Build. A useful chatbot connected to your live data is typically ₹1,50,000 to ₹6,00,000, depending on how many systems it plugs into. A simple FAQ bot pulling from a fixed document sits at the bottom. One that queries live order status, checks stock and books appointments sits at the top, because most of that work is integration rather than the bot itself.
Running. Cheaper than most people expect. The per-conversation cost of the underlying service is fractions of a rupee. A business handling 2,000 conversations a month typically spends under ₹1,000 on it. Hosting is the same server your site already runs on.
The economics are usually straightforward. If a bot handles 60% of 2,000 monthly queries, that is 1,200 conversations not reaching your team. At even three minutes each, that is 60 hours a month.
The mistake is treating it as a headcount replacement. It rarely is. What it does is stop your existing team spending their day on "what are your timings" so they can handle the queries that actually need them.
Six things worth getting right
Say it is a bot. Customers work it out in two messages, and pretending otherwise costs you trust for no gain.
Make the human handover obvious. A visible route to a person at every point. Ours sits in the header of the chat window and never disappears — including when the assistant itself is unavailable. That last part matters: if your bot goes down, your contact channel should not go down with it.
Read the transcripts. This is the most under-used benefit. Every conversation is a record of what customers actually ask, in their own words. Ours has produced product page fixes and FAQ entries we would never have thought of. Read them weekly for the first month.
Start narrow. Pick your top five questions. Get those genuinely right. Expand from evidence rather than imagination.
Give it a scope and enforce it. Ours will not discuss anything outside our services. It says so and steers back. A bot that will chat about anything will eventually be screenshotted saying something you would rather it had not.
Test the failure path harder than the success path. Anyone can demo a bot answering a question it knows. What matters is what it does when it does not know, when a customer is rude, and when someone tries to talk it into something. Spend your testing time there.
What good looks like six months in
The businesses getting real value share a pattern.
Their bot answers a specific set of questions from live data and hands off everything else quickly. Their support team handles fewer routine queries and more real problems. They read transcripts and improve their website based on what they find. Their customers get answers at midnight.
Nobody has replaced their support team. They have just stopped that team spending half their day repeating themselves.
The honest summary
An AI chatbot for customer support is worth having if you have repetitive questions and data it can read to answer them. It is not worth having if your queries are mostly complex, emotional or high-stakes — and forcing one into that situation makes things worse.
The single most important decision is not which platform you use. It is making sure the bot answers from your real systems rather than from guesswork, and that a customer can always reach a person. Get those two right and the rest is refinement.
If you are considering one and want a straight assessment of whether it fits your business — including a "probably not worth it" if that is the answer — Seven Digital Services runs free consultation calls.
Want to talk this through?
Free 30-minute call — no obligation, and we'll tell you if we think it isn't worth doing.
Book a free strategy call →