The AI itself

The thinking part of what we build is a service we rent by the month, like electricity or a phone plan. This page explains what that means for your bill and for your data.

A service you rent by the month

When an assistant reads a WhatsApp order, drafts a reply or pulls the total off a scanned bill, the reading and writing is done by an AI model. A model is a very large program trained to read and write ordinary language, Tamil and English among many others. Nobody builds one for a single business. A handful of large companies run them as a service, and we pay for what your systems use, the way you pay for electricity or a phone plan.

Nothing is installed on your machines for this part. Your system sends a question over the internet, the service answers within a second or two, and the answer comes back to be checked and used. Everything else, the parts that hold your data and move the work along, sits on a server you control. A server is a computer rented by the month in a data centre. Where yours lives is explained on Your data.

Why we use more than one

Different models are good at different jobs. One is careful with long documents and rarely misses a line. Another is quick and cheap at short questions, which is most of what a WhatsApp assistant does. A third turns Tamil speech into text well. None is the best at everything, and their prices move every few months.

So each task in your system is matched to the model that suits it, and each can be swapped on its own. If a supplier raises prices or a better model appears, we move that one task and mention it in the monthly note. You are never tied to one company. This is the third of the three rules on How it's built: nothing you will be locked into.

Reading a scanned bill: A careful model that reads slowly and checks that the totals add up. Used a few hundred times a day at most, so its higher price hardly shows.
Answering a price question: A quick, cheap model. Used thousands of times a month, so its price is the one that matters, and the one we watch.
Writing out a voice note: A model built for speech, in Tamil and English, that turns the note into text before one of the others reads it.

What it costs each month

The models' share of the bill is small next to the hours they give back, but it is real, and it grows with use. These are the ranges we see for the models alone, in businesses of five to two hundred people. Our own fees are separate and are set out on Pricing.

Scroll sideways to see the rest of the table.

Typical monthly cost of the models alone
What it is doing	Typical range, a month
A staff assistant answering questions from your documents	₹1,500 to ₹6,000
Reading and entering a few hundred orders or invoices a day	₹3,000 to ₹12,000
A customer-facing WhatsApp assistant for a busy business	₹8,000 to ₹40,000

Billed by the model's supplier for what is used, to an account in your name where possible. A quiet month costs less than a busy one, and the figure appears on your monthly note.

Two things move the number: how many messages or documents there are, and how long each one is. A price question is a few lines. A scanned bill is a page. A voice note is longer still, because it is written out first. We tell you where you sit in the range before we start, and we show you the real figure every month.

What happens to your data when a question is sent

Every time the assistant needs the model, it sends only what that one task needs. That is the customer's message, the few lines of your price list that match, and the rules we wrote for how to answer. Not your whole set of records, not your customer list, not last month's sales.

Sent: the text of that one task, with names and phone numbers taken out where the task does not need them.
Answered: within a second or two, back to your server, where the answer is checked before anything is sent on.
Not kept: the services we use do not store the exchange beyond a short window for abuse checks, and do not use it to train their model or anyone else's.
In writing: we only use services that make that promise in their business terms, and a copy of those terms sits with your agreement.

Some services run their computers in India and some abroad. For everyday work such as reading an order, either is fine, and we tell you which is in use. For anything you consider sensitive, we can choose a service with computers in India, or run a smaller model on your own server so that nothing leaves it. Keeping it safe covers the rest: who can open what, backups, and what we never do with your data.

How we keep the bill predictable

A bill that depends on use can run away if nobody is watching it. So we settle four things before anything goes live, and they stay in place for as long as the system runs.

We agree a ceiling: A monthly limit for each assistant or workflow, in rupees, written into the plan. Use cannot go past it without someone on your side saying yes.
The cheap model goes first: Routine questions go to the quick, cheap model. Only the hard ones, a scanned bill or a long document, go to the careful, dearer one.
We watch the use: We check the count every day. At 80 percent of the ceiling we tell you, and we look at why. Usually it is good news: more orders.
You see the figure every month: The monthly note shows what was used, what it cost, and whether the ceiling still fits the business.

One question, start to finish

What happens when the assistant needs to think

A customer asks whether an item is in stock. This is the route the question takes, and where the checks sit on either side of the service. Nothing goes out unchecked and nothing comes back unchecked.

In words

You ask a question on WhatsApp, in Tamil or English.
Your server works out what the task needs: the question, the matching lines from your price list or stock sheet, and the rules for answering. It takes out any names and numbers the task does not need.
Your server sends that to the AI service, which reads it and writes an answer in a second or two. The service keeps nothing afterwards.
Your server checks the answer against your records and your rules. A price that is not on your list, or an answer the model is unsure of, does not go out; it asks a person instead.
The checked answer goes back to you on WhatsApp, in the language you wrote in.

For your technical teamModels, selection by task, routing, caching, data terms, regions

Models in use: Claude (Anthropic), OpenAI models, Google Gemini, and open-weight models from the Llama and Qwen families served on a rented GPU or on the client's own server where data must not leave it.
Selection by task: Reasoning and long-document extraction: the strongest tier we can justify per document. High-volume classification and short replies: a small, fast model. Speech: a dedicated transcription model for Tamil and English, then a text model. Cost is a first-class criterion; each task has a rupee budget per call.
Routing and fallbacks: A thin routing layer maps each task to a primary model and one or two fallbacks. Timeouts, rate limits and provider outages fail over automatically. Every call is logged with model, tokens and cost.
Prompt caching: Stable prefixes (system instructions, price lists, policy text) are cached where the provider supports it, which cuts cost and latency on repeated calls.
Data-processing terms: Business API terms only: no training on inputs or outputs, zero or short retention. The data-processing addendum for each provider is kept with the client agreement.
Region choices: Indian regions where the provider offers them; open-weight models on servers in India for sensitive work; otherwise the nearest region, and the client is told which.

An example routing table for one client. Budgets are per call and are enforced, not advisory.

task              primary      fallback   budget per call
read-order        small, fast  mid        under 0.20 rupees
read-invoice      strong       mid        under 2.00 rupees
transcribe-note   speech       none       under 1.00 rupee a minute
monthly-summary   strong       mid        under 20 rupees

Ask which service your work would use

Tell us the task. We will say which kind of service suits it, what it would cost a month, and where its computers are. We reply within one working day.

Talk to us

hello@pakshitechnologies.com

Monday to Saturday, 9:30 am to 6:30 pm IST