Chatbots that do the boring half of your inbox

What an AI chatbot actually costs to run

Build cost is the number every supplier quotes. Running cost is the number that decides whether the project makes sense, and it is quoted far less often. There are three components.

The WhatsApp charge is the one that surprises people. Meta bills per 24-hour conversation window rather than per message, which changes how the bot should be designed: a conversation that resolves the customer's question in twelve messages costs exactly the same as one that resolves it in three. Optimise for resolution, not brevity.

Model cost scales with conversation volume and with how much context you feed the model on each turn. A well-built retrieval layer that sends only the three most relevant paragraphs of your catalogue costs a fraction of one that sends the entire catalogue every time. This is invisible to you at build time and very visible on the bill.

Where AI chatbots actually fail

The failure mode people worry about is hallucination — the bot confidently inventing a price or a policy. That is real, and it is also the easiest problem to engineer away: build with retrieval so the bot answers only from your approved content, and instruct it to say it does not know rather than guess. We test this deliberately before go-live by trying to make the bot lie.

The failure modes that actually kill chatbot projects are less dramatic:

• No human handover. The single fastest way to make customers hate your business is to trap them in a loop with a bot that cannot help. Every bot we ship has explicit escalation triggers — pricing negotiation, complaints, repeated failure to answer — that route to a person with the full conversation attached.
• Stale knowledge. A bot trained on last year's price list will confidently quote last year's prices. Whoever owns the content needs a way to update it without calling us.
• Nobody reads the transcripts. The conversation log is the most valuable customer research your business will ever have — it is a list of exactly what people want to know, in their own words. Most businesses never look at it.
• Wrong tone. A bot that sounds like a US SaaS company talking to a customer in Vashi reads as fake. Tone is a build decision, not an afterthought.

Multilingual is not optional for most of India

English-only bots quietly lose a large share of enquiries. Customers who are not comfortable writing in English will either switch to a competitor who answers in their language or simply call — which defeats the point of the bot.

We build bots that detect the language of the incoming message and reply in it, covering Hindi, Marathi, Gujarati, Tamil, Telugu, Bengali and Kannada alongside English. In practice a large share of real Indian customer messages are code-mixed — Hinglish, or Marathi written in Roman script — and a bot that only handles clean Devanagari will fail on them. Handling the messy real version is the actual engineering problem.

How to tell whether your business is ready for one

A chatbot pays for itself when four things are true. If two or more are false, spend the money elsewhere first.

• You receive enough repeated enquiries that answering them is a real cost — as a rough rule, more than twenty a day.
• The answers are consistent and can be written down. If every answer depends on judgement, a bot will frustrate people.
• Someone in your business will own the bot after launch — updating content, reading transcripts, adjusting escalation rules.
• You are actually losing enquiries to slow response, and you can point to roughly how many.

What we hand over

A chatbot project ends with a trained bot connected to your real content, handover rules configured, lead capture wired into your CRM or a Google Sheet, and a dashboard showing what people actually asked. You get the source code and the deployment access. If you want to take it elsewhere afterwards, you can.

We also spend the first thirty days after launch tuning it, because the questions real customers ask are never quite the ones you predicted during the build.