AI assistants
An assistant answers the questions your team answers all day, and hands over the ones it should not touch. This page shows where that line is, and how we test it before it meets a customer.
What it can and cannot do
An assistant is a program that reads messages and replies in ordinary language, using your price list, your stock sheet and the rules you set. It is good at the questions that come in a hundred times a week and bad at anything that needs judgement. Being honest about that line is what makes it useful.
It can
- Answer price and stock questions from your own lists, in Tamil or English
- Take an order and read it back for the customer to confirm
- Send an invoice, a payment link or a catalogue when asked
- Tell a customer where their order is
- Remind a customer about an unpaid bill, in the tone you set
- Answer staff questions from your policies and manuals
- Work at eleven at night and on a Sunday
It cannot
- Decide a discount or give credit
- Know anything that is not in your records
- Handle an upset customer
- Be right every time; when unsure it says so and hands over
- Replace your sales person's judgement
- Promise, sign or commit on your behalf
- Learn your business on its own; we set up what it knows
It is not a person and does not pretend to be one. Customers are told they are talking to an assistant, and they can ask for a person at any point. Where the words it writes come from is explained on The AI itself.
When it hands over to a person
Handing over is not a failure. It is the design. These are the rules, in plain words. You can add your own, and you can change the limits yourself without asking us.
- It is not sure
- If the answer is not in your records or the question is unclear, it says so and hands over. It never guesses a price or a delivery date.
- The customer is upset
- Words like complaint, wrong, angry or refund, or a message in capitals: it apologises once, does not argue, and hands over.
- Money above a limit
- An order, refund or credit above a figure you set waits for a person. Below it, the assistant goes ahead and a person is told.
- Anything about a complaint or a return
- Straight to a person, with the order history attached.
- The customer asks for a person
- At once, with no questions, in whichever language they wrote in.
Handing over means your sales person gets the whole conversation so far on their phone, a one-line summary, and the reason. The customer is told a person will reply and by when. Nothing has to be repeated. The money limits and approval steps sit with the other safeguards on Keeping it safe.
How we test it before it talks to your customers
No assistant goes live on the day it is built. Three things happen first, and together they take about three weeks. This is the part of a Sprint that decides whether the assistant is ready, and it is where most of the fixing happens.
A list of test questions
We write down two hundred real questions from your WhatsApp history, in Tamil, English and the mix in between. That includes the awkward ones: a wrong price, a rude message, a request for credit. The assistant answers all of them, and we go through every answer with you.
A two-week shadow run
The assistant reads live messages but does not reply. It drafts, while your team sends their own replies as usual. Each evening we compare the two and fix what differs.
A weekly review
Once live, we read a sample of conversations with you every week for the first month, then every month. Each hand-over is checked both ways: was it right to hand over, and was it right not to?
Before and after
The same day, twice
Slide the divider. On the left, one person answers all eighty WhatsApp messages between everything else, and the slow ones wait until evening. On the right, the assistant answers sixty of them within a minute and hands twenty to a person, each with the conversation so far.
Before
The owner answers about 80 WhatsApp messages a day, between calls and dispatch. Replies slip to the evening, and some are missed.
After
The assistant answers about 60 of them within a minute. The other 20 reach a person with the chat so far, a summary and the reason.
A worked example
A price question that turns into a credit request
One conversation, start to finish. The assistant handles the part that is in your records and hands over the part that needs a decision, without the customer having to repeat anything.
In words
- A customer asks, in Tamil or English, what a 2 inch PVC pipe costs and whether 500 are in stock.
- The assistant looks up your price list and your stock sheet. It does not guess, and it does not quote a price that is not on the list.
- It answers in the language the customer wrote in, within a minute: the price, the stock, and the delivery time you set.
- The customer asks whether they can pay next month.
- Credit is a decision, not a lookup. The assistant tells the customer a person will reply, and your sales person gets the whole chat, a one-line summary and the reason on their phone.
For your technical teamTool use, frameworks, guardrails, evaluation sets, escalation, logging, red-teaming
- Tool use
- The model does not answer from memory. It calls functions we write, such as lookup_price, check_stock, get_order_status and create_draft_order. Each returns structured data from the client's own records, and every reply is grounded in that data.
- Frameworks
- A thin, hand-written loop rather than a heavy agent framework: one model call, a bounded set of tools, a fixed number of steps. We avoid LangChain-style abstraction in production because open-ended planning is harder to test and harder to explain to an owner.
- Guardrails
- Allow-lists for what the assistant may say and do: prices only from the price list, no discounts, no delivery promise without a date from the system. Money and credit above the threshold are blocked at the tool layer, not only in the prompt.
- Evaluation sets
- A held-out set of real conversations with expected outcomes, run before every change. A change ships only if answer accuracy and hand-over precision stay above the agreed bar.
- Escalation design
- Hand-over is itself a tool call. It opens a ticket with the transcript, a summary and a reason code, notifies the named person, and tells the customer the expected response time. The assistant stops replying on that thread until a person closes the ticket.
- Conversation logging
- Every message, tool call and model output is logged with timestamps and cost. Logs stay on the client's server in India for the period the client sets.
- Red-teaming
- Before launch we try to make the assistant misbehave: instructions hidden in customer messages, requests for other customers' data, pressure for discounts, abuse. Every finding becomes a case in the evaluation set.
The hand-over tool, in outline.
hand_over(reason, transcript, summary)
ticket = open_ticket(reason, transcript, summary)
notify(sales_person, ticket)
reply(customer, "A person will reply within an hour.")
pause_thread(until = ticket.closed)Tell us what your team answers all day
A week of typical WhatsApp questions is enough for us to say how many an assistant could answer, how many it would hand over, and what it would cost. We reply within one working day.
Monday to Saturday, 9:30 am to 6:30 pm IST