The billable work in an agency chatbot service is the content, not the install. Package it as a setup fee plus a line on the maintenance contract, fix the client's published answers first, and write down where your support ends before the first client asks you to answer their customers.
Clients started asking for AI on their websites about eighteen months ago, and most agencies answered it one of two ways. They built something with a vector database and a model key, and now one developer owns a service thirty clients depend on. Or they said no, and watched a client buy a foreign tool that put someone else's logo on the site they maintain.
There is a third way to run this, and it looks like every other recurring service an agency already sells.
What you are actually selling
Not the widget. Pasting a script tag takes a minute, and a client who watches you do it will price it accordingly.
What you are selling is the reason the bot gives good answers: someone went through the client's website, found the twenty questions their customers ask, checked which ones had a published answer, wrote the missing ones, and deleted the page from 2023 that still lists old rates. That is a day of work per client, it is visible, and it improves the website whether or not the chatbot exists.
Sell that. The chatbot is what makes the work legible to the client, because for the first time they can watch their own website answer a customer.
Packaging it so it sells itself
Two lines on the invoice.
A one-time setup fee covering the content audit, the writing, the install and the first two weeks of tuning. Price it as you price a day or two of work.
A monthly line added to the annual maintenance contract you already invoice, alongside hosting and backups. This is the part that compounds, and it survives because the client already pays you monthly for the site to keep working.
Do not sell the chatbot as a standalone product with its own contract. That turns it into a thing the client evaluates every year against whatever they saw on Instagram, rather than a component of the service they already renew.
An onboarding sequence you can repeat
The agencies that make this work run the same five steps on every client, which is what makes a second client cheaper than the first.
Collect. Ask for their last fifty customer enquiries, from the inbox and their WhatsApp, before you look at their website. That is the question list, and it comes from the client rather than from your imagination.
Audit. Check each question against the live site and mark it answered, unanswered or contradicted. The content checklist is the pass, and the output is a short document the client can see.
Fill. Write the missing answers as two or three plain sentences each, get them approved, publish them. Enter the ones that do not belong on a public page as question-and-answer pairs.
Install and test. Paste the tag, then test the client's bot from your own dashboard before the client sees it. Ask it their own pricing question. If it refuses, the content work is not done.
Hand over. Show them where the leads land, show them the conversations, and tell them what to do when a customer reports a bad answer.
Getting the content out of a client
This is where the service dies if you let it. You ask for the rate card, and three weeks later you are still asking.
Two things that work. Take the questions rather than the documents: a fifteen-minute call where you ask the owner "what does a customer ask you every single day" produces more usable material than a month of chasing a PDF. And write the answers yourself, then send them for approval. Clients are slow to write and fast to correct.
Put the boundary in the proposal in plain words: the bot answers from what the client publishes, so anything they choose not to publish will not be answered. That sentence saves an argument later, and it is also the honest description of how grounded chatbots work.
Where your support ends
Write this down before the first client, because the failure mode is predictable. A client forwards you a customer's question and expects you to answer it.
A workable boundary: you own the widget working, the crawl being current and the leads arriving. The client owns replying to their own customers and telling you when their prices change. Bad answers get reported to you, and you fix the content that produced them rather than defending a model.
Set a review cadence, quarterly is enough for most, where you read the month's refusals and turn them into a short list of content to publish. That meeting is also where next year's renewal gets decided, because you are arriving with evidence that the thing works.
Pricing it
Cost-plus is the easy start: your platform cost per client, plus your margin, plus the setup fee. It is defensible and it undersells the work.
Value-based is harder to quote and closer to the truth. A client who runs a ₹40,000 service and captures two extra enquiries a month from customers who would have left is not comparing your monthly line to a software subscription.
Either way, the platform economics decide the floor. On Chatterbox that floor is the ₹3,000 monthly agency fee spread across your book, minus 15% off each client tenant's plan, which improves with every client you add. The white-label page has the details, including what the client sees on the widget.
Start with one client. Preferably one whose website is already decent, so the first run is fast and you learn the sequence rather than the exceptions. The second client is where you find out whether you have a service or a favor.