An AI support chatbot trained on your docsand nothing else.
Point Anchor AI at the help center you already publish, upload the policy PDFs, paste the answers you keep retyping. It answers customers from those sources only. When the answer isn’t there, it says so and hands the conversation to your team with the transcript attached.
If you have written the answer down somewhere, Anchor AI can give it to a customer at two in the morning, in the words you chose. If you haven’t, it will tell you which questions to write up first.
How it works
In four steps
- 01
Teach it
Add the URL of your help center or website. The crawl runs in the background and finishes in minutes. Upload PDFs and documents. Paste short snippets for the answers that live nowhere else: opening hours, delivery zones, the one returns exception.
- 02
Make it yours
Give the agent a name, a greeting and a tone (Friendly, Professional, Casual or Formal). Scope topics tell it what it may answer. Handoff topics tell it what to pass to a person.
- 03
Meet your agent
Ask it your real questions in the Playground before a customer does. Coverage shows the questions your content can't answer yet, so you fix the docs rather than the bot.
- 04
Put it live
One script tag on your site, or share the hosted chat page. Switch it live when the answers read right. If they ever stop reading right, switch it off from the same place.
What it will not do
Invent a price, a policy or a delivery date
If the answer isn't in the content you connected, the agent says so instead of guessing, and offers to bring in a person.
Answer outside its scope
Scope topics are a fence. A question about something you didn't put in bounds goes to your team rather than to a model's general knowledge.
Keep talking when a person is needed
Billing disputes, refunds, an angry customer: whatever you list under handoff topics goes straight to the inbox with the transcript attached.
Train anyone else's model on your content
Your content never trains shared models, and we purge it from our systems on request.
What it refuses to do is as much the product as what it does.
What “trained on your docs” means here
Anchor AI does not fine-tune a model on your content. It reads it. When a customer asks a question, the agent finds the passages in your connected sources that answer it and writes its reply from those passages, and only those. Nothing about your business is baked into a model’s weights. It is looked up, every time.
That is the difference between an agent that answers from your docs and one that was once shown them. Change a price on the page and re-crawl it, edit a snippet, replace a file: the next answer uses the new text. There is no re-training to wait for, and no old version of the fact lingers in the model.
It also means the agent is only as good as the docs, which is what the next section is about.
It tells you what your docs are missing
Every question the content couldn’t answer lands in Worklists → Content gaps, the list of things customers asked that nothing you connected could settle. Next to it, Failure modes separates Missing knowledge (nobody wrote it down) from Wrong or outdated answer (someone did, and it has drifted). The first is a writing job. The second is an editing job.
Before launch, the Playground’s Coverage view does the same job in advance: ask it the questions your team gets, and see which ones it has to decline. The gaps it finds are usually things everyone on the team knew and no one had typed.
What to give it first
Start with what you already publish: the help center or the FAQ page, by URL. Then the policies (returns, delivery, payment) as the PDFs or pages they live in. Then product pages, if the questions are about products.
After that, snippets. A snippet is a few lines you paste in for an answer that exists only in someone’s head: the delivery fee to Tamale, whether Saturday counts as a working day, what to do about a parcel the courier marked delivered. These are the questions that come back every week, and each one takes a minute to write down once.
Keep one source of truth for each fact. When the same price lives on three pages, one of them is eventually wrong, and the agent has no way to know which. Every price change is a release is the routine for keeping the docs, and therefore the agent, right.
What it costs
Anchor AI: one meter
- Starter
- GHS 250 /seat/mo500 credits ≈ 50 conversations
- Growth
- GHS 600 /seat/mo1,500 credits ≈ 150 conversations
- Free trial
- 30 days · 300 credits · no card
A Starter seat is GHS 250 a month with about 50conversations in it. Connecting more content doesn’t change the price. You pay for seats and the conversations they cover, and never for pages or documents.
Common questions
Do I have to write new content before the chatbot is useful?
No. Start with what you already publish (the help center, the FAQ, the policy pages) and let the crawl run. Then read Content gaps after the first week. It lists the questions customers asked that your content couldn't answer, which is your writing list, in order of demand.
What can I train it on?
Three kinds of source: a website or help center crawled by URL, uploaded files such as PDFs, and short snippets you paste in for answers that live nowhere else. The agent answers from those sources only.
Will it make things up?
It answers only from the content you connect. If the answer isn't there, it says so instead of guessing and offers to bring in a person, and the question lands in Content gaps so you can write it up. You can test this yourself in the Playground before any customer sees it.
What happens when my docs change?
Update the source (re-crawl the page, edit the snippet, replace the file) and the next answer uses the new text. There is no re-training. If a fact drifts without anyone noticing, Failure modes flags the conversation as a wrong or outdated answer so you catch it.
Is this the same as training ChatGPT on my data?
No. Nothing is fine-tuned. Anchor AI looks up the relevant passages in your content at answer time and replies from them, so a change to the docs is a change to the answers. Your content never trains shared models and is purged from our systems on request.
How long does setup take?
The crawl runs in the background and finishes in minutes. The four steps (add your knowledge, name the agent and set its tone, test it in the Playground, embed the widget) fit in an afternoon, and every new account starts on a 30-day free trial with no card required.
Keep reading
An AI support chatbot that hands off to a human
Anchor AI answers what your content covers and hands the rest to your team mid-conversation, transcript attached. You draw the line with scope topics and handoff topics; the Escalation report shows where it fell.
AI customer support for businesses in Ghana
Anchor AI is an AI customer support chatbot priced in Ghanaian cedi: GHS 250 a seat a month with about 50 conversations, answers from your own content, human handoff included. 30 days free, no card.
Anchor AI vs. Chatbase
Per-seat credits against per-reply credits that vary by model, and a handoff Chatbase sells from $150 a month.
Anchor AI vs. Intercom Fin
Per-seat credits against $0.99 per outcome, and where Fin still wins.
See it answer from your own content
Point Anchor at your help center and watch it answer your real questions — 30 days free, no card.