CustomerBot indexes your docs, changelog and help centre, then answers setup and how-do-I questions with the passage it used — during a trial, when it decides whether someone activates, as much as after they have paid. The tickets that remain are the ones actually worth a human.
Support deflection is the one that gets measured. The onboarding ones are worth more and nobody counts them, because a user who quietly gives up during a trial never files a ticket.
The user who cannot get the webhook working on day one rarely opens a ticket. They just stop logging in, and you record it as a trial that did not convert.
Every product has vocabulary new users do not share yet. A patient explanation at 11pm is cheaper than a call, and it happens whether or not anyone is on shift.
Reference documentation is precise and badly suited to browsing. Retrieval turns it into something you can ask rather than search.
If your docs explain an error code, the assistant can find it from the visitor’s paraphrase of it — which is how people actually describe errors.
Train it on the changelog and “did you break X?” gets an accurate answer instead of a ticket that interrupts an engineer.
Pricing, SSO, data residency, seat limits. These are on your site already and they are the questions that quietly decide evaluations.
Every unanswered question is a gap in your docs, ranked by how many people hit it. This is the output SaaS teams undervalue most.
Same assistant, same knowledge base, three very different jobs — configured per placement rather than duplicated.
Search boxes return ten links. An assistant returns the answer.
Point CustomerBot at your documentation domain and it pulls every page it can reach, chunks each one and indexes it. A user asking “how do I rotate an API key” gets the two sentences that matter plus the page they came from, instead of a results list they have to triage themselves.
Bulk-train from a sitemap in one pass
Context-aware help without building a help centre in-app.
The same one-line script tag works inside your authenticated app. Set the welcome message and the suggested questions per environment, so a user stuck on the billing screen is offered billing questions rather than a generic greeting.
Custom questions, welcome copy and proactive bubbles per assistant
Pre-sale questions are the ones worth answering fastest.
Lead capture runs inside the conversation: the assistant collects a name and email at the right moment, tags the transcript, and the lead lands in a list you can filter by status and export as CSV. Sales gets the question, not just the address.
Built-in lead capture, statuses and CSV export
A fine-tuned model goes stale the moment you deploy. Retrieval does not — the answer is read fresh from your content every time, so updating the page is the whole update.
You ship
A release goes out and the docs page changes with it.
You re-crawl
Re-train the URL — or the whole sitemap — from the training screen.
It re-embeds
The page is re-chunked and its vectors replaced in the index.
Answers move
The next question about that feature is answered from the new text. No fine-tune, no deploy.
Each assistant picks its own model, and each reply costs credits by tier. Most support volume is well served by the cheapest tier; spend the budget where the questions are hard.
| Workload | Sensible pick | Cost per reply | Why |
|---|---|---|---|
| High-volume FAQ deflection | Gemini 2.5 Flash Lite, GPT-4o Mini | 1 credit | Fast and cheap; plenty for questions your docs already answer well. |
| Nuanced technical support | Claude Sonnet 5, GPT-5.4 Mini | 2 credits | Better at holding a multi-step debugging thread together. |
| Long, messy knowledge bases | Gemini 3.1 Pro, GPT-5.5 | 3 credits | Million-token context windows and stronger synthesis across pages. |
Invite people per assistant with one of four roles. Nobody needs the billing password to fix a bad answer, and a viewer can read every transcript without being able to change a thing.
Owner
Everything, including billing, settings and deletion.
Admin
Manage the team and edit training and content.
Editor
Add and correct training data, read transcripts.
Viewer
Read-only access to conversations and analytics.
It is a chat assistant trained on your product’s own documentation, help centre and changelog, sitting on your docs site, marketing pages or inside the product itself. For a SaaS company it does three jobs that a generic widget does not: it unblocks new users during onboarding when activation is decided, it answers technical and API questions accurately enough to be trusted, and it tells you which documentation you have not written yet.
That is arguably where it earns the most. Support tickets are visible and get counted; a trial user who could not get the integration working on day one is invisible and simply does not come back. Putting the assistant where new users get stuck — the setup guide, the first-run screen, the API quickstart — catches the questions that never become tickets.
Docs first, because that is where the content already lives and where the questions are most repetitive. The marketing site next, since pre-sales questions about limits and security stall evaluations. Inside the authenticated product last — it is the most valuable placement and the one where you most want live account context, which means wiring up API tools rather than just crawling pages.
Only through tools you connect. Crawled documentation knows what your product does in general; it does not know this user’s plan, usage or account state. Describe the relevant endpoint as an API tool and the assistant can call it during the conversation — that is the difference between “here is how limits work” and “you are at 82% of yours”.
Search matches keywords and hands back documents. Retrieval-augmented generation matches meaning, pulls the relevant passages and has a model compose an answer from them. A user who asks “why am I getting a 429” gets the rate-limit explanation even though the page never uses the word 429 in the heading.
It can explain your API from your reference docs, and it can call your API if you configure an API tool — a plain-English description of an endpoint that CustomerBot turns into a callable tool schema. That is how you get answers about a specific account rather than the general case.
It falls back rather than guessing, and the question is recorded. The analytics view surfaces the questions with no good source behind them, which is effectively a prioritised backlog for your documentation team.
Yes — each assistant has its own team with owner, admin, editor and viewer roles. Support can read transcripts, a docs writer can edit the training set, and nobody has to share a login.
It detects the language of the question and replies in it, drawing on your English documentation. That covers 95+ languages without you maintaining a translated help centre.
CustomerBot runs as a hosted service and routes model calls through OpenRouter, so you can pick from more than forty models across OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and others per assistant. There is no self-hosted edition today.
The free plan is enough to crawl your documentation and put the assistant in front of a real question before you decide anything.