AI answers that cite your help center, and hand off when they can't help

How an answer is produced

1. Screen. The question is cleaned and checked for attempts to misuse the bot, such as overriding its instructions or probing for other customers' data. Known abuse gets a fixed reply and never reaches the model.
2. Retrieve. Roper finds the most relevant published articles. With semantic search configured it matches meaning, so a paraphrase still finds the article; otherwise it uses keyword search.
3. Answer. With AI on, the question and the retrieved articles go to the model, which must answer from those articles only. With AI off, the bot replies with the best matching article's opening and a link.
4. Check. The answer is shown only if it cites an article it was given, says something those articles support, is short, and contains no code, links or addresses the articles do not contain.
5. Cite. The articles used appear under the answer as tappable cards linking to your help center.
6. Offer the next step. The latest bot message shows "That helped" and "Talk to a human".

When it hands off to a person

• The customer asks for a person or money back: words like "human", "real person", "refund", "charged twice" or "cancel my account".
• The AI says it cannot help, refuses, or the call fails.
• No article is available to answer from.
• A message contains only attachments, because the bot cannot read files.

A handoff changes the conversation to open and appears at the top of your inbox, with a system message telling the customer they are now talking to the team. If the customer has not given an email, the widget asks for one so you can follow up. The full set of rules is in chatbot human handoff best practices.

Disclosure and safety

• It says it is an AI. The widget shows "You're chatting with an AI assistant, not a person. A human is one tap away." and labels every bot message. When answers come from keyword search only, it says "automated assistant" instead, so the wording matches what it does. Once a person takes over, the notice goes away.
• It never promises refunds, discounts or timelines.
• It has nothing to leak. No secrets, internal notes or other customers' data are ever sent to the model, and the model has no tools to call.

More on the rules that apply to disclosure in should chatbots disclose they are AI?, and on keeping answers grounded in how to stop AI support hallucinations.

What it does not do

• It does not take actions in your product, such as refunds, plan changes or order lookups.
• It does not learn from conversations on its own. It answers from the articles you publish, so improving answers means improving articles.
• It does not guarantee a correct answer. A weak or missing article gives a handoff, not a better guess.
• It does not run workflows or follow routing rules.

Getting good answers

• Write one article per question, with the question as the title and the answer in the first paragraph.
• Use the words customers use, not your internal names.
• Keep articles current. The bot cannot know a price or a step changed unless the article does.
• Read the conversations that hand off and write the missing article.

A longer walk-through is in how to train an AI chatbot on your help center.

Cost

AI answers are included in Pro ($19 a month) and Team. There is no separate fee per resolution. For how that compares with per-resolution pricing elsewhere, see AI customer support pricing.