Capability 06 · Code
Most businesses do not need more artificial intelligence. They need the same twenty questions answered at midnight, enquiries that reach the CRM without anyone copying them, and a person ready to step in when it matters. From Mumbai, we design and build AI assistants, WhatsApp chatbots and automations that do exactly that, with the guardrails written before the first prompt.
41%
of customer questions answered end to end by a Dubai skincare brand's AI concierge
38%
higher conversion for the same brand once its store, app and concierge worked as one
95%+
email delivery for a B2B trading house, with its WhatsApp moved to official broadcasts
Where it usually starts
Nobody writes to us wanting a model. They write because something repetitive is eating a team's day, or because a first attempt went wrong.
Your founder or support team answers the same questions on WhatsApp, Instagram and email. The replies are good. They also arrive hours late, and after hours they do not arrive at all.
A quick pilot on a general model answered confidently and wrongly: an invented return policy, a discount that never existed. Now nobody trusts it and someone wants it switched off.
Messages from personal numbers, broadcasts through unofficial tools, numbers getting blocked and no record of who agreed to hear from you. The channel works until the platform notices.
Enquiries, orders and invoices arrive as emails, PDFs and forms, and someone retypes each one into the CRM or ERP. It is the work everybody agrees should not need a person.
What we do
Designers shape the conversation, engineers build the plumbing and both sit in the room when we decide what the assistant must never do.
Retrieval augmented generation, or RAG, over your product data, policies and documents. The assistant searches only what you have approved, answers from it, shows its source and says so when it does not know.
The WhatsApp Business Platform set up through an official solution partner, with a verified profile, approved templates and a chatbot that handles orders, bookings and common questions before a person steps in.
Clear rules for when the assistant stops and a person takes over: by topic, by order value, by tone or simply on request. The conversation arrives with a summary, so your customer never repeats themselves.
A written list of what the assistant must never do, a sprint spent trying to make it misbehave, and a scored set of your real questions that every release has to pass before it reaches a customer.
Automations for the copy and paste work between your inbox, CRM, ERP and spreadsheets, and agents that act through your systems, with a person approving anything that spends money or changes a customer's record.
Search that understands a question, summaries, recommendations and drafting tools inside your website, app or portal, designed by the same team that designs the rest of the product.
The Ampersand Method, applied
The same six phases as every D&C project. On AI work, the real conversations come first and the testing takes longer than people expect.
How the method worksWe read a sample of your real conversations and tickets, with names removed, and sit with the people who answer them. By the end we know the questions, the tone and the topics a machine should never touch.
A knowledge audit of which documents are current and who owns them, automation candidates ranked by hours saved and risk, three measures agreed with you, and a running cost estimate per thousand conversations.
Conversation design for greetings, fallbacks and hand off moments, drawn alongside the architecture: which model, what it can read, where data is stored and what it is never allowed to see.
Two week sprints with a demo every Friday. Every build is scored against a few hundred of your real questions, and one whole sprint goes on trying to make the assistant say something it should not.
A staged start, often with people reviewing replies before customers see them, then one channel at a time: website, app, then WhatsApp. Your team is trained on the inbox and the hand off before day one.
A monthly read of conversation samples with your team, new answers where it hesitated, tighter rules where it was too sure, and every new model tested on your question set before we switch.
Proof
A bilingual concierge that answers 41% of a Dubai brand's questions and knows exactly when to hand over, and a trading house whose WhatsApp moved from blocked blasts to official broadcasts its own team runs.
All workWe went from messages landing in spam to a channel our buyers actually read. The team trained us properly and then stayed on call.
Tools we trust
How we engage
Most AI work starts small on purpose: a two to three week pilot on one channel and one set of questions, measured against your own conversations. If the numbers hold, we widen it one channel at a time. Programmes with agents that act inside your systems begin with a full three week Discovery Sprint.
The most useful AI we build is a little boring. It answers the question a customer asked at eleven at night, correctly and in their language, and passes the conversation to a person when it should. It moves an enquiry into the CRM without anyone copying it across. Nobody calls it a breakthrough. They simply stop dreading Monday’s inbox.
Every project starts in Listen & Learn with the conversations you already have. We read a sample of WhatsApp chats, emails and support tickets, with names removed, and sort them into three piles: questions with one right answer, questions that need judgement, and work that is really copying data from one screen to another. For a D2C skincare brand in Dubai, three months of messages became the concierge’s first syllabus. The piles decide the product, not the other way round.
Our guide to chatbot costs and how RAG works covers what each one costs to build and to run.
On every AI project, the first document we write is not a prompt. It is a short list headed things it must never do. Never give medical advice. Never promise a refund. Never guess a delivery date. Then we spend a whole sprint trying to make the assistant break those rules, with awkward questions, instructions hidden inside messages and topics it should politely decline. It fails a few times. We tighten the rules each time, and it goes live only when it holds.
A good assistant knows when to stop. Complaints, sensitive topics, high value orders and anyone who simply types “person” reach a human with a two line summary, so nobody asks the customer to repeat themselves. For the Dubai skincare brand, the concierge now answers 41% of questions end to end. The rest reach the team faster and already summarised, which is the half of the result founders tend to value most.
In India and the Gulf, most customers would rather message than call. We set up the official WhatsApp Business Platform with opt-in, approved templates and a shared inbox, then connect the chatbot to it, in English and Arabic where needed, which matters to teams in Dubai and across the Gulf. A B2B trading house moved from blocked numbers to official broadcasts, with replies routed to the right person on its sales desk. Automations also feed your CRM and ERP, so a conversation becomes a lead or an order without retyping, and our growth team plans the broadcasts customers actually want.
Models change, products change and customers find new ways to ask. Each month we read a sample of conversations with your team, add answers where the assistant hesitated and tighten it where it was too sure of itself. New models are tested against your own question set before we switch. It is the habit we learned keeping websites online: look after what you ship. When you are ready, tell us what your team answers every day.
FAQ
A focused FAQ assistant for your website usually costs ₹2 to 5 lakh to build. A RAG assistant that answers from your documents and hands over to people typically runs ₹6 to 15 lakh, and agents that act inside your systems cost more. Running costs, mostly model usage, hosting and WhatsApp messages, often sit between ₹5,000 and ₹50,000 a month. Our AI chatbot cost guide breaks it down.
RAG stands for retrieval augmented generation. Before the assistant replies, it searches your own documents, product data and policies, then writes an answer using only what it found, and can show the source. It matters because a general model on its own answers from whatever it learned on the internet. With RAG, your return policy comes from your return policy.
Yes, and for most businesses it should. We restrict the assistant to your approved knowledge, instruct it to say when an answer is not there, and test that it actually does. It still writes in natural, friendly language, but the facts come from you. When a question falls outside that knowledge, it offers a person instead of improvising.
The free Business app suits a small team replying by hand from one or two phones. You need the API, now called the WhatsApp Business Platform, for a chatbot, several people on one number, links to your store or CRM, or broadcasts at volume. It needs Meta's approval and a verified business, and our WhatsApp readiness checklist covers what your website needs.
Meta charges for template messages by category and by the recipient's country. Marketing messages cost the most, utility and authentication messages less, and replying to a customer who wrote to you is free inside the 24 hour customer service window. Solution partners add a platform fee, and GST applies. Meta revises its rates, so every estimate we send uses the current rate card.
Mostly the patient, repetitive work between systems: turning an enquiry into a lead in your CRM, reading order details from an email into your ERP, checking delivery status, booking a slot or drafting a reply for someone to approve. We start with tasks that are frequent, well defined and easy to check, and a person approves anything that spends money or changes a customer's record.
Whenever the stakes rise or the assistant is unsure: complaints, refunds above a set limit, health or legal questions, a customer asking for a person, or two failed attempts to understand. The hand off should carry a short summary so nobody repeats themselves. For a skincare brand in Dubai, any mention of a reaction, pregnancy or prescription goes straight to a trained adviser.
It can be, if it is designed that way. The Digital Personal Data Protection Act expects a clear purpose, consent where it is needed, as little data as possible and deletion once the purpose ends. We mask personal details before anything reaches a model, choose providers that do not train on your data, keep logs for a set period and document where every piece of data goes.
Send a rough list, or a week of messages with the names removed. A partner replies within one working day with what we would automate first, and what we would leave to people.
