The reason most store owners have not put a chatbot on their site is not cost. It is the fear of what it will say.
A generic assistant bolted onto a store will confidently tell a customer you offer free returns within ninety days. You do not. Now you either honour it or start the conversation by telling someone they were misled by your own website.
Helpie is built around removing that risk rather than apologising for it.
It answers from your content or not at all
Every answer has to be supported by something Helpie actually found on your site. Your pages, your products, your posts, your documentation.
If the assistant cannot point at a source for what it is about to say, the answer does not go out. It offers to raise a ticket instead.
This is the whole design, not a setting. The assistant is not drawing on general knowledge about how online stores usually work, because "usually" is where the wrong answers come from.
It shows its working
Answers carry the sources they came from as clickable links, with the matching passage highlighted so the customer can see where it came from.
That is useful for the customer, who can read the full page if they want the detail. It is more useful for you: when an answer is wrong, you can see immediately which page misled it, and the fix is editing that page rather than wrestling with a prompt.
It asks instead of guessing
When a question is too vague to answer well, the assistant asks one short clarifying question. "Do you mean delivery within Australia, or international?"
If the second attempt still finds nothing, it stops trying and offers the handover. It does not loop, and it does not keep asking questions to look busy.
There is a small thing here that customers notice: when the assistant offers numbered options and the customer replies "2", that is handled instantly and for free, with no AI call at all. Fast, cheap, and it works even when the API is having a bad day.
Handover carries the conversation
When it escalates, the conversation goes with it into the ticket. The customer does not start again, and your agent opens the ticket already knowing what was asked and what the assistant tried.
The handover copy is written for the moment rather than being one canned sentence. If a customer has been going in circles for several turns, the tone acknowledges that instead of cheerfully offering the same form a fourth time.
Refusals that behave
Some questions should not be answered by a chat widget on a shop. Legal advice, medical advice, financial advice, and anything where a customer’s wellbeing depends on being right.
The assistant declines those and routes to a human. It will still surface and link your own content on a sensitive topic, because that content is yours and you published it deliberately. What it will not do is offer an opinion on how the law or a medical question applies to that particular customer.
It also holds the line under pressure. Attempts to talk it into a different persona, extract its instructions, or push it into topics outside your site are blocked before they reach the model at all.
Your own API key
Helpie uses your Anthropic key. You are the account holder, you see the real usage, and you set hard spending caps inside Helpie so a strange week cannot become a strange invoice.
What it costs you to run
Less than people assume, because most support questions are the same twenty questions.
Answers to repeat questions from anonymous visitors are cached, so the second person to ask about your delivery times costs nothing. The greeting is rendered without an API call. The numbered-option shortcut is free. The assistant is designed so the common path is cheap and only genuinely novel questions cost anything.
Off until you are ready
The customer-facing widget is switched off when you install Helpie. You index your content, talk to the assistant privately using the built-in tester, correct whatever it gets wrong, and publish it when you are satisfied.
Nobody’s customers are the test group.