Grounded and governed answers
Two words describe how the Talven Web Assistant stays exact and honest. Grounded means it answers from what you approved. Governed means rules decide what it may say and do.
- What does grounded mean?
- Can the Talven Web Assistant invent my prices?
- What happens when the Talven Web Assistant does not know the answer?
- What does governed mean?
- How are the Talven Web Assistant's replies checked?
- How do I control what the Talven Web Assistant says?
- How do I know the Talven Web Assistant is doing a good job?
- Is every answer from the Talven Web Assistant always right?
- Do visitors know they are talking to AI?
- Can I test the Talven Web Assistant with hard cases before I go live?
What does grounded mean?
Grounded means the Web Assistant answers from the knowledge you approved, such as articles, images, rules, menus and prices, and checks that knowledge before it answers.
A grounded answer can be traced back to something your business published. In CX Studio Manager you can read everything your Web Assistant knows, edit it in plain words, and see what changed.
It does not answer from the open internet unless you turn web search on. Web search is a setting for the moments your knowledge has no answer.
Last checked
Can the Talven Web Assistant invent my prices?
No. The Web Assistant is built to quote prices, hours and addresses exactly as you set them, and it says it does not know when it has no such fact.
Prices, policies and media stay exactly as you set them. You keep facts that must be exact as structured records, and the Web Assistant shows them on cards without rewording them.
Like all AI, it can still make a mistake. That is why each conversation is graded and why you can correct any answer. See Is every answer always right?.
Last checked
What happens when the Talven Web Assistant does not know the answer?
The Web Assistant says it does not know, does not make something up, and hands the question to your team with the conversation attached.
For example, a guest asks about a date the Web Assistant has no calendar for. It says it does not have that date, passes the question to the front desk, and attaches the conversation. The front desk then confirms the answer directly.
The question also appears in CX Studio Manager among the things it could not answer. You can then add that fact, so the next visitor gets the answer.
Last checked
What does governed mean?
Governed means two layers of rules decide what the Web Assistant may say and do: your own business rules, and a safety and grounding contract that nobody can edit.
- Your rules. You tell your Web Assistant in plain words what it must say and what it must never say. One rule goes on one line. A change is live from the next conversation.
- Talven's contract. Under your rules sits a set of safety and grounding rules. These rules are not editable, which protects your Web Assistant from abuse. Nobody can weaken them under pressure, not even you.
Last checked
How are the Talven Web Assistant's replies checked?
A second AI model, the output judge, reads a reply against the rules before the visitor keeps it, and it can pull the reply back.
The Guardrails page in CX Studio Manager shows three things:
- The status of the judge and what it did in the last 30 days.
- The standing protections. They are always on, and you cannot switch them off.
- Your own guardrail rules, one sentence each.
For each of your rules you choose what happens. Retract pulls the reply before the visitor keeps it. Flag lets the reply through and records it for your team.
Last checked
How do I control what the Talven Web Assistant says?
You control the Web Assistant from CX Studio Manager by editing what it knows, writing its rules in plain words, and correcting any answer you see.
- Edit its knowledge: articles, photos, menus and offers.
- Set its name, colors and greeting.
- Write the rules it follows in chat and in voice mode.
- Write the guardrails it must never break.
- Correct an answer. The correction is live in the next conversation.
Each change is kept as a version, and an activity log shows who did what. Nothing happens behind your back.
Last checked
How do I know the Talven Web Assistant is doing a good job?
CX Studio Manager grades each conversation: did the Web Assistant answer, did it stick to your content, and was the visitor happy or not?
You can read any conversation, see what visitors ask most, and see what it could not answer. You can also test a single article. CX Studio Manager runs test conversations against it to check that the Web Assistant answers from it, and these tests stay out of your analytics.
Last checked
Is every answer from the Talven Web Assistant always right?
No. The Web Assistant uses automated systems, so an answer can be inaccurate or incomplete, and it is not professional advice.
We build the product to keep that risk small: it answers from approved knowledge, it says when it does not know, and replies are checked. We publish no accuracy percentage, because a number would depend on each business's content.
You stay responsible for what you publish through it, as our Terms of Service say. Review what it says, above all when your answers carry legal, medical or financial weight.
Last checked
Do visitors know they are talking to AI?
Yes. A line under the message box is always on and tells visitors that the replies are AI-generated.
You can add a Learn more link to that line. You can also ask visitors to accept terms before they use the Web Assistant, and it can store a record of each consent.
Last checked
Can I test the Talven Web Assistant with hard cases before I go live?
Yes. You can ask your Web Assistant your hardest questions in CX Studio Manager before you add it to your website.
Nothing reaches your visitors until you paste the line of code into your site. A larger organization can bring its hardest questions to a free, private pilot built from its own content. See What is a pilot?.
Last checked