AI transparency

1. You are talking to AI

A Web Answerly agent is an AI system, not a person. The chat widget always shows a notice that answers are generated by AI and may not be accurate. When someone from the business takes over a conversation, the chat says so, and the AI stops answering until it is handed back. Businesses must not present their agent as a human.

2. How an answer is made

1. The business gives its agent knowledge: its website, documents, written answers and connected sources. Web Answerly splits these into passages and indexes them for search.
2. When someone asks a question, the agent finds the passages most relevant to it.
3. An AI language model writes an answer from those passages, the conversation so far and the business's instructions, and lists the sources it used.
4. If the passages found do not match the question closely enough, the agent does not guess. It gives the business's fallback message and offers a way to reach a person.

3. Limits you should know about

• Answers are generated automatically. They can be wrong, incomplete or out of date, even when sources are shown. A source shows what the answer drew on; it is not a guarantee.
• An agent knows only what its business has given it, as of the last time it was indexed.
• An agent does not give professional advice. Check anything that matters — medical, legal, financial, safety or similar — with a qualified person.
• Do not type passwords, card numbers, API keys or other secrets into a chat.

4. Controls businesses have

• Trust rules can stop an agent discussing a topic, or limit it to answers the business has verified. A rule only ever restricts the agent.
• Human handoff and takeover: visitors can ask for a person, and the business's team can take over any conversation from a shared inbox.
• Email replies are drafts by default: a person reviews and sends them. A business can turn on automatic replies only for well-grounded answers, and those replies say they were automated.
• Review tools: analytics of unanswered questions, a test lab for checking answers before changes go live, and optional reviews of past conversations. The test lab and reviews use an AI model to grade answers; their results are prompts for a person to look, not verdicts.

5. Actions

An agent can offer buttons, such as booking a meeting, opening a link or sending a request to the business. Where a business connects its own systems, an agent can also look information up or make a change through them. An action that changes something waits for the visitor's confirmation whenever the business's policy requires it, runs at most once, and is recorded. The business decides which actions its agent has and is responsible for what they do.

6. Which AI models are used

• Web Answerly chooses the AI model provider and the default model. Our current provider is OpenAI, which generates answers and creates the embeddings used for search. The current list is on our Subprocessors page, and we give notice before adding a provider.
• A business may be given a choice of models, and can name backup models to answer if its main model fails. Nothing switches to a model the business did not choose; if every chosen model fails, the agent gives its fallback message instead.

7. Voice

If a business turns voice on, speech is turned into text by the visitor's browser or by our model provider, and the text is answered exactly like a typed question. Web Answerly never receives or stores audio, and the microphone is used only after the visitor presses the microphone button.

8. Training

We do not use businesses' content or their visitors' conversations to train AI models. Our model provider states that data sent through its API is not used to train its models unless a customer opts in, and we have not opted in. Our Privacy Policy has the details.

9. Decisions about people

Web Answerly agents answer questions and help people reach the right next step. They are not designed to make decisions that significantly affect people. Our Acceptable Use Policy prohibits using them to decide access to employment, credit, housing, insurance, healthcare, education or government benefits without human review, and prohibits clinical or emergency use.

10. Questions and concerns

If an answer from an agent concerns you, contact the business that runs it. To report an agent you believe is misused, or to ask about this page, email support@webanswerly.com.