A general-purpose chatbot on a website answers questions nobody asked and fails at the one thing the business needed. An agent built for a single job — qualifying an enquiry, booking an appointment, recovering a cart — can be measured, improved and held to a number.
A niche AI agent is a conversational system built for one business and one outcome, grounded in that company's own data through retrieval, with defined escalation to a human. Unlike a generic chatbot, it is measured on a business metric rather than on containment rate.
Agents that do one thing well are measurable and improvable. Agents that do everything are demos.
Retrieval over your documents and records, with a refusal path when the answer is not there.
Low confidence, explicit request for a human, or a sensitive topic. Any one hands over immediately.
Including the enquiry that arrives at 11pm, which is a meaningful share of them in India.
Agent brief and success metric
Structured knowledge base
Working tool access
Agent in staging with an eval suite
Pilot report and fixes
Live agent and a monthly review
It greets every visitor, offers four buttons, cannot answer the actual question, has no access to any real data, and its only escape route is a form. It is measured on how many conversations it 'contained', which is a metric that rewards trapping people.
Each one has a single accountable outcome. That is what makes it possible to say whether it works.
Enquiries arriving at all hours about pricing, availability and location, answered the next working day if at all.
An agent that answers project and pricing questions from real inventory, qualifies budget and timeline, and books a site visit into the sales calendar.
Reception handling timings, doctor availability, preparation instructions and directions by phone, all day.
An agent handling scheduling and logistics from the real calendar, with anything clinical escalated to a human immediately and explicitly.
Order status, delivery and returns queries at volume, mostly answerable from data the business already holds.
An agent reading real order records, plus proactive cart recovery conversations on WhatsApp.
Applicants asking about fees, eligibility, syllabus and deadlines across a long admissions window.
An agent grounded in the actual prospectus, capturing applicant details and handing warm enquiries to counsellors.
Eligibility and document questions repeated hundreds of times a week before any application begins.
An agent explaining requirements and pre-checking eligibility, with every regulated statement drawn from approved copy.
Booking and availability enquiries across WhatsApp, Instagram and phone with no single queue.
One agent across channels, reading real availability and taking bookings.
Building something that responds is easy now. Building something a business can rely on is a design problem, and it comes down to four things.
The agent reads your actual inventory, calendar, prices and order records. An agent answering from a general model with no retrieval will eventually invent something expensive.
It must be able to say 'I do not know, let me get someone'. Agents without a refusal path fill the gap with plausible fiction, which is the failure mode that ends trust permanently.
A test set of real conversations with expected outcomes, run against every prompt or model change. Without it, an improvement is an assumption.
Appointments booked, leads qualified, carts recovered. Containment measures how well you trapped somebody, which is not the same as helping them.
The one job, the one metric, and what the agent is explicitly not for. Written down before anything is built.
Your documents, policies, prices and FAQs structured for retrieval — usually the largest and most valuable part of the work.
Live access to calendar, inventory, order data or CRM, because an agent that cannot see real data can only generalise.
Opening, qualification path, objection handling and the handover, written by someone who has read your real conversations.
Topic boundaries, approved copy for regulated statements, and a refusal behaviour that is graceful rather than blank.
Confidence, sentiment and explicit-request triggers, with the conversation handed over in full to a person.
Real conversations with expected outcomes, run against every change before it ships.
The business metric, plus a report of what the agent could not answer — which is the most useful output of the whole system.
An agent is not a single model call. It is a small system: a conversation layer, a set of tools it is allowed to use, retrieval over your own content, and a supervisor deciding when a human takes over. This is what sits behind a single deployed agent.
Handles the dialogue across WhatsApp, web chat or voice, keeping context across a conversation that may span days.
Fetches the relevant passages from your own documents, prices, policies and product data before any answer is composed.
Reads live availability, order status and inventory, and writes bookings back — the difference between describing and doing.
Asks the questions the business actually needs answered and records them as structured fields, not as chat transcript.
Blocks regulated claims, pricing it is not authorised to quote and topics outside its remit, using approved copy where a statement is required.
Watches confidence, sentiment and explicit requests, and hands to a human with the full conversation attached the moment any threshold trips.
Replays a growing set of real conversations against every change, so an improvement is demonstrated rather than assumed.
Reports the business outcome — bookings, qualified leads, recovered carts — alongside the questions it could not answer, which is next month's content plan.
A general-purpose chatbot on a website answers questions nobody asked and fails at the one thing the business needed. An agent built for a single job can be measured, improved and held to a number.
Containment rate measures how many people the system prevented from reaching a human. Optimised for, it rewards trapping people. Every agent we build is accountable for something the business actually wants more of, which usually means the agent should be *keen* to hand over.
Answers come from retrieval over your own documents and records, and the agent must be able to say it does not know. Agents without a refusal path fill the gap with plausible fiction, which is the failure mode that ends trust permanently and is entirely preventable at design time. How this is built is in RAG chatbot development.
A property agent answering pricing and availability from real inventory and booking site visits. A clinic agent handling timings and scheduling with anything clinical escalated immediately. A store agent reading real order records and recovering carts on WhatsApp. An admissions agent grounded in the actual prospectus. Each one accountable for a single number.
Hindi, Bengali and English with language detection, so the customer never has to choose. The knowledge base should be reviewed by a speaker in each language rather than machine-translated and left — the agent will faithfully reproduce whatever is in it.
The agent should say it is an automated assistant. What people object to is not talking to software; it is software pretending to be a person, and software with no way out to a human. Both are design choices, and both are avoidable.
Every stage ends in something you can hold — a document, a build, a live account. If a stage cannot name its output, it is a meeting, not a stage.
One outcome, one metric, and the boundaries. Most agent projects fail here by trying to do everything.
Gather, clean and structure the content the agent answers from.
Connect the systems it needs to read and write — calendar, inventory, CRM.
Conversation design, guardrails, escalation, and the evaluation set built alongside.
Run on a subset of traffic with a human reviewing every conversation before responses are trusted.
Live, monitored, with unanswered questions feeding the knowledge base monthly.
Everything here is part of the engagement at no extra cost. We do not itemise them on an invoice and we do not withhold them if you leave.
The one job, the one metric, and what the agent is explicitly not for. Agreed before building, which is where most agent projects go wrong.
Real conversations with expected outcomes, run against every prompt or model change, so an improvement is demonstrated rather than assumed.
Every month, what the agent could not answer. It is the most useful output of the whole system and it writes your content plan for you.
Running on a subset of traffic with a human reviewing every conversation before any response is trusted.
The cleaned, organised content the agent answers from — useful well beyond the agent itself.
Longer answers to the questions people ask before they hire anyone for ai agents by niche.
Built by us, free, no signup, nothing uploaded to a server. Take them whether or not you ever become a client.
It is scoped on the number of channels, how many systems it needs live access to, and the state of the content it will answer from — the knowledge base is usually the largest part of the work. Model usage is an ongoing cost that scales with conversation volume, and we report and optimise it monthly.
Tell us what you have now and what you are trying to reach. We will audit it and tell you what we would do, what it would cost and whether you need us at all. The audit is free and yours to keep.