AI that runs inside the business, not in a demo
We automate one workflow at a time: the first reply to a lead, the invoice that gets typed twice, the report someone rebuilds every Monday. Each one ships with an evaluation set, a human approval step where money or records are involved, and a measured cost per task.
That is deliberately less exciting than a chatbot that can do everything, and it is the only version that survives contact with a real operation.
Lead qualification on WhatsApp
The first reply goes out in under a minute, at any hour. The assistant asks the qualifying questions your sales team asks, scores the lead and hands only the serious ones over with the conversation attached. Message fees are charged at cost, not marked up.
Document extraction and posting
Bills, invoices, purchase orders and bank statements arrive as photos and PDFs. We extract the fields, check GST and totals against your rules, flag mismatches and post the clean ones for approval instead of typing them twice.
Internal copilots over your data
A private assistant that answers questions from your own documents, tickets and database, with citations back to the source row or page. Access follows the permissions of the person asking, so a sales user cannot read payroll.
Forecasting and anomaly checks
Demand, collections and inventory forecasts built on your own history, plus alerts when a metric moves outside its normal range. Useful where the data is regular and the decision is recurring, not as a dashboard nobody opens.
Six rules every automation follows
Most AI projects fail after the demo, when nobody can say whether the output is right, what it costs, or what happens when it is wrong. These rules exist so those questions have answers on day one.
- A written evaluation set of real inputs before the first prompt is tuned, so quality is a number and not an opinion
- A human approval step on anything that spends money, sends a legal document or changes a record
- Hard limits on what the model may read, written as code, not as an instruction inside the prompt
- Cost per task measured and reported, so a workflow that saves a lakh a month does not quietly cost more than that in tokens
- Logs of every automated action with the input that caused it, so a wrong output can be traced and replayed
- A fallback path when the model is unavailable: queue and retry, or hand the task back to a person
One workflow, measured before it is widened
Pick one workflow
We start with a single recurring task that already has a cost you can measure: hours of typing, leads that go cold overnight, invoices keyed twice. If the saving cannot be estimated before the build, the project is not worth starting.
Build the evaluation set
Fifty to a few hundred real examples, including the ugly ones, with the correct answer written down. This set decides whether the system works, and it is the first thing we build, not the last.
Ship it narrow
One workflow, one interface your team already uses, whether that is WhatsApp, email or a screen in your admin panel. It runs beside the existing process for two weeks, and a person can override every result.
Measure and widen
We compare the automated output against the manual baseline weekly: accuracy, handling time, cost per task. Only when the numbers hold do we add the next workflow or remove the human approval step.
What you can check afterwards
- Accuracy measured against a written evaluation set, not demoed on cherry-picked examples
- Human approval on payments, legal documents and record changes
- Data stays in your infrastructure or a region you approve
- Token and API cost reported per task, monthly
- Full audit log of every automated action and its input
- Graceful fallback to a human when the model fails or is uncertain
Two workflows you can buy as they are
Both are live products, priced in public. A custom automation follows the same pattern, built around one process in your business.
WhatsApp Lead Desk
Answers every enquiry in under a minute, asks the qualifying questions, scores the lead and passes the serious ones to your team with the conversation attached.
₹35,000 setup + ₹10,000/mo
Accounts Desk
Bills sent on WhatsApp are extracted, checked for GST and totals, and posted to your accounting system for approval. Up to 300 invoices a month.
₹30,000 setup + ₹9,000/mo
Custom automation
A workflow specific to your operation: quoting, dispatch checks, support triage, report writing. Scoped after discovery, priced as a build plus a monthly run fee.
Build from ₹2,00,000, discovery from ₹25,000
Before you send a brief
What does an AI automation project cost?
The two packaged workflows start at ₹35,000 setup plus a monthly fee. A custom automation is quoted like software: discovery from ₹25,000, build from ₹2,00,000, plus a monthly run fee that covers hosting, monitoring and model cost.
How do you stop it making things up?
By narrowing what it is allowed to do. Extraction tasks return fields from a document, not prose. Answers over your data must cite the row or page they came from. Anything that spends money or changes a record waits for a person to approve it.
Where does our data go?
Wherever you require. Small models can run on your own infrastructure for extraction and classification. Larger models are called through an API in a region you approve, with a data-processing agreement and no training on your content.
What if the model gets it wrong?
Every automated action is logged with its input and output, so a wrong result can be replayed and corrected. The evaluation set is re-run after every prompt or model change, and a drop below the agreed accuracy pauses the workflow rather than shipping the regression.
Do we need clean data first?
For extraction and lead handling, no. For forecasting, usually yes, and the first phase is spent fixing the history rather than training a model on it. We say which case you are in during discovery.
Name the task nobody wants to do
Tell us the recurring job, how often it happens and who does it today. We will tell you whether it is worth automating, and what it would cost.