Namici Chatbot: a WhatsApp assistant that sounds like your team

Before reading any of this, go and use it. Both doors are open and neither needs a form.

• Message the bot on WhatsApp: +62 878 2169 7205. It is a furniture shop today. Ask what it sells, argue with the price, say it is too expensive, ask for a photo.
• Open the dashboard: demo-chatbot.namicilab.com — sign in with demo@namicilab.com / Demo12345!@. It is a read-only account, so click anything you like.

The rest of this page explains what you are looking at, feature by feature.

What a conversation actually looks like

This is the shape of exchange the system is built to produce. Customer on the left of the colon, bot on the right.

• Daniel: an office chair. my back is killing me by 3pm every day
• Bot: That 3pm wall is usually lumbar, not the cushion
• Bot: How many hours a day are you sitting?
• Daniel: 9 or 10. work from home
• Bot: Proper full-time sitting then
• Bot: Are you tall? Seat depth starts mattering above 180
• Daniel: 186
• Bot: Arbor or Meridian. Both actually fit your height
• Bot: $420 and $680. The gap is armrests and 12 year warranty vs 5
• Daniel: what would you pick
• Bot: At 10 hours a day? Meridian
• Bot: The armrests are what save your shoulders. You’d feel it by week two

Notice what it does not do. It does not open with “Thank you for contacting us!”. It does not send a paragraph. It does not list all six chairs with bullet points and prices. It diagnoses before it recommends, asks one question at a time, and when asked for an opinion it gives one.

Four exchanges later it has a Saturday appointment and Daniel’s name in the leads table.

What it is

Namici Chatbot answers your WhatsApp Business number. It knows what you sell, quotes your prices from your own catalogue, sends your photos, recognises when somebody is ready to buy, and hands that conversation to a person at the right moment.

There are no flows to draw and no decision tree to maintain. You tell it who it is and what you sell; it holds the conversation.

Why it does not sound like a bot

Most WhatsApp bots sound like bots for a reason that has nothing to do with the model: nobody ever told them how the business actually writes. “Be friendly and professional” produces the same paragraph-shaped assistant every time.

This one is told something far more specific. You upload your best salesperson’s real WhatsApp exports — the ordinary .txt files WhatsApp gives you — and the system measures how that person types:

• Median words per reply
• How often they answer in fewer than five words
• How many messages they send in a burst before waiting
• The longest reply they ever send

Those exact figures go into the model’s instructions. If your best salesperson sends two short messages instead of one tidy paragraph, so does the bot, because it was measured against somebody who does.

The samples teach rhythm and technique. They never teach facts. That separation is what lets the same system run a furniture shop, a property agency and a dental clinic without touching the code.

The phrases that make somebody sound like themselves

The system also reads the samples for the phrases that person genuinely uses, counts how often each one appears, and puts the list on screen. You decide which the bot may copy.

This is a human decision, not an automatic one. Measuring finds everything somebody says, including things no company should repeat. A phrase you switch off is genuinely forbidden — it goes into the instructions as a banned list, so the bot will not use it even though it appears throughout the examples it is learning from.

Examples chosen by the question being asked

Rather than showing the model the same fifteen conversations every time, it retrieves the exchanges that answered questions like this one, ranked by similarity and narrowed to the right topic. Every exchange is classified first as business talk, rapport or off-topic, and off-topic never reaches a prompt.

A handful of exchanges can be pinned so they appear on every single reply — the greeting, the sign-off, the phrasing that best captures the voice. Everything else is left to relevance.

What it knows

Two sources, both edited from the dashboard by whoever runs the business.

The catalogue

What you sell. Shared columns for the things every business has — code, name, description, category, price, availability, location — plus a free-form attributes field for whatever your domain needs. Bedrooms and land size, or seat depth and warranty length, under your own labels.

Prices can be a single figure or a range, in which case the bot says “from $560” rather than quoting the top of the range as final. A price note carries wording you want repeated verbatim: “per month”, “per person”, “VAT included”.

Items are grouped into catalogues, and each bot is told which catalogues it may reach. Saving an item mirrors it into the knowledge base and indexes it in the same action, so there is no sync step anybody can forget.

The knowledge base

Everything that is not an item: opening hours, delivery areas, warranty terms, the returns policy, the questions three people ask every day. It is organised as folders so it can be filed the way a person thinks about it, and each bot is given the folders it is allowed to use.

How it finds things

Both are retrieved by meaning rather than keyword, so “my back hurts by the afternoon” finds the ergonomic chair without anybody having tagged it that way.

The bot has three tools it can decide to use mid-conversation: search the catalogue, search the knowledge base, and check whether a specific item code is still available. It chooses; you do not script when.

One column on the knowledge screen matters more than it looks: whether each row is actually indexed. Whether the bot can retrieve something is normally invisible until a customer asks and gets nothing. Here you can see it, and one button fixes the backlog.

Getting your information in

Nobody wants to retype a price list, so you upload the file you already have.

Catalogue import

Drop in a PDF, a Word document, an Excel workbook or a CSV. A model reads it and proposes rows — because real price lists put the header on row four, merge a category across a band, hide an internal cost sheet on another tab, and arrive as a PDF at least as often as a spreadsheet. Mapping columns by position gives up on all three.

You then review every row before anything is written. Each is labelled new, update or error, with a count of each and the existing price shown beside the new one, so a price change is something you approve rather than something you discover. Name, description and price are editable right there.

Prices written as “$1,120.00”, “Rp 4.200.000” or “4,2 jt” all become plain numbers. Anything that is not clearly a number — “call us”, “price on request” — is flagged as an error rather than guessed at, because a hallucinated price is quoted to a customer as fact.

Knowledge upload

Upload a policy PDF or a Word FAQ and it is split into entries you can read, retitle, merge or delete before saving. You can make the split finer or coarser and see the result immediately.

The split happens before the confirm step because splitting is free and indexing is not — and because a scanned PDF that produced no text at all is obvious on that screen, rather than a week later when the bot cannot answer from it.

Nothing reaches a customer that somebody has not looked at first.

Contacts

Your customer list, with bulk import: name, phone, email, gender, age, birthday, city, occupation, tags, source and free notes. Searchable and filterable, and the audience you pick a broadcast from.

Every word the bot says is a setting

The prompt being editable is table stakes. The harder part is everything nobody thinks of as a prompt — and in this system all of it is on a screen:

• Identity — who the bot is, and what it must never claim to be.
• Approach — how it sells: diagnose first, one question at a time, what to do when somebody hesitates.
• Safety rules — the things it must never do, which are read before everything else.
• Using information and tools — answer from retrieved data, never invent a name or a price, what to do when a lookup comes back empty.
• Photos and video — how media is handled, since the system attaches it and the model cannot.
• WhatsApp message format — one line per message. This is a contract, not a preference: the channel splits replies on line breaks.
• Recording customer details — what is worth noting about somebody, phrased as recording rather than interrogating.
• When the customer commits — what counts as agreeing to the next step.
• Using the retrieved reference — the instructions attached to the retrieved information, which sit closest to the answer and outrank everything above them.
• Never do this — voice and behaviour prohibitions, kept separate from safety policy because they exist for different reasons.
• Reply length and rhythm — normally generated from the measurements, and overridable when you disagree with them.

Then the small text nobody thinks of as dialogue, all of it editable:

• What a search says when the knowledge base has nothing, when the catalogue has nothing, when a code does not exist, and when an item is or is not available. The model reads these and says them out loud in its own words — they are the bot’s voice, not internal plumbing.
• What the model is told arrived when a customer sends a photo, a video, a document, a voice note or a sticker with no caption.
• Greetings answered in kind, as plain phrase pairs: when they write this, reply with that. No pattern syntax, because one stray bracket typed into a settings box would throw an error on every incoming message.
• The opening greeting, sent verbatim so the first impression is authored rather than generated.
• The words that signal a customer has committed — “fitting”, “test drive”, “site visit” — in your language and your industry’s vocabulary.
• The extra lead fields worth capturing for your business, beyond the standard ones.

There are no defaults hiding behind an empty field, because a default is a rule the person running the business cannot read, cannot edit, and does not know exists.

Empty fields are not left as a blank page either. Every one shows a worked example as faint placeholder text, and one button fills them all with suggested wording — written through the normal save path, so the text lands in the settings where you can read and change it like anything else.

Photos, files and voice notes

Customers send pictures. A photo with no caption is turned into words the model can act on, in wording you control, so “here’s my room” is understood rather than dropped.

Going the other way, items carry their own photos and videos, and the system sends them alongside the reply. It remembers what it has already sent in that conversation, so a customer who asks twice does not get the gallery twice.

When a customer is ready to buy

The moment somebody agrees to the next step, the conversation switches to a human and the people you nominate get a push notification on their phones. What counts as agreement is configurable, because the next step is a site visit to one business, a fitting to another and a consultation to a third.

Every conversation is in one of three modes — answered by the bot, answered by a person, or a mix — and you can move between them at any time:

• Take over a conversation from the dashboard, and hand it back when you are done. Both are written into the conversation itself, so the timeline shows who was answering and when it changed.
• Reply directly into a live WhatsApp chat from the dashboard, bypassing the bot entirely. Sending a message takes the conversation over automatically, because somebody typing by hand almost never wants the bot to answer the next one.

Each bot reply also carries a confidence score, so the conversations worth reading are easy to find.

Leads, without anybody filling in a form

Every conversation quietly builds a lead record: who they are, how to reach them, what they asked about, and what they need before they will buy. Leads move through stages — new, contacted, qualified, negotiating, won or lost — and the list is searchable and filterable by all of it.

The fields that matter are your decision. A property business wants bedrooms and budget; a clinic wants which treatment and when. You name them, the bot collects them in conversation rather than by interrogation.

Reaching out first

WhatsApp only allows businesses to start a conversation using a template that Meta has approved, which is normally a tedious back-and-forth between two systems. Here it is one screen:

• Write the template, with placeholders for the values that change.
• Submit it to WhatsApp for approval without leaving the dashboard.
• Pull its review status back, so you know whether it was approved or rejected rather than guessing.
• Reconcile every template against your WhatsApp Business Account in one action, when the two have drifted apart.
• Send an approved template to a chosen set of contacts as a broadcast.

Test it before a customer sees it

Three things let you check the bot before it is live, and after every change:

• A simulator in the dashboard that runs the exact same pipeline as WhatsApp — same retrieval, same tools, same persona. Not an approximation of the real thing, the real thing without the phone.
• A prompt preview showing the exact instructions the bot will be given for a given question. Not a summary — the real text. A strange answer can be traced to the words that caused it, and anything you left blank is visibly blank.
• An example tester where you type a real customer question and see which past exchanges the bot would draw on, with a similarity score on each.

More than one number, more than one business

Each WhatsApp number points at its own persona, and the persona decides everything: the voice, the rules, and which catalogues and knowledge folders it may reach. A sales line and a support line can run from one installation and know nothing about each other. The Graph API version is pinned per number, so one can be upgraded without moving the others.

A number with no persona assigned does not borrow somebody else’s. It says nothing, records the customer’s message so a person can pick it up, and the log names the screen that fixes it.

A wrong voice is worse than no voice. A bot that stays quiet costs one message; a bot answering as the wrong business costs trust, and you hear about it from the customer.

Knowing it is working

• Reports — conversation volume, the daily trend, and average response time.
• Every conversation is stored and searchable, with the full message history, the media that went out, and a timeline of who was answering.
• Notifications go to the people you nominate, with a test button so you find out it works before you need it.
• Startup checks print what is not ready to serve customers before the first request, and the deploy refuses to run at all when something is configured in a way that would fail silently.

That last one earns its place. The failures worth engineering against are the ones that never produce an error: a bot that looks healthy and answers nobody because a key is missing, or a dashboard built pointing at the wrong address so every request fails on the customer’s machine and never on yours.

Built to be handed to a team

• Roles, from full administrator down to a read-only account that can open every screen and change nothing. The demo login on this page is exactly that.
• Credentials are encrypted with AES-256-GCM before they are stored, and never returned to the browser. A database dump, a backup on somebody’s laptop or an open database viewer yields nothing usable.
• Incoming webhooks are signature-verified, so nobody can post messages into your bot.
• Sessions are per device, so signing in on a phone does not sign anybody out on a laptop.
• Uploaded conversation exports are stored outside the public folder, because they contain your customers’ names and numbers.

Try it

The fastest way to judge any of this is to use it:

• WhatsApp: message +62 878 2169 7205. Ask for a recommendation, push back on the price, ask what the warranty covers, ask it to send a photo, tell it you want to come and see one.
• Dashboard: demo-chatbot.namicilab.com — demo@namicilab.com / Demo12345!@. Open the persona and read the instructions the bot is actually given. Open the catalogue and see where its answers come from. Open a conversation and read the timeline.

The demo number allows a set number of messages per person per day, which is plenty to form an opinion. The dashboard account cannot change anything, so nothing you click will break the demo for the next person.

If you want one of these pointed at your own catalogue and your own number, get in touch and we will set it up.