Help Answers Grounded in Reviewed Articles

AI chat assistants are an appealing way to offer help inside a product. A customer types a question in plain language and gets an answer in seconds, without searching a help center. The risk is that a general AI assistant will happily invent instructions that sound right and are not. "Click Settings, then Advanced, then Export" is useless, or harmful, if there is no Advanced tab.

When we built help into the admin area of our platform, we designed it so the assistant cannot invent instructions. Instead of writing answers, it finds them.

Select, do not compose

When a customer asks a question in the help bubble, the AI's job is to choose the single most relevant article from our reviewed Help Center. The customer then sees that article's approved text, along with a citation linking to the full article.

The AI does not write a new answer in its own words. Every instruction the customer reads was written and reviewed by a person, and the citation lets them check the source.

Only articles the customer may see

The pool of eligible articles is limited to general help and articles written for that customer's own account. Content written for any other organization is excluded. A help assistant must respect the same boundaries as the rest of the platform, which we describe in tenant-first thinking.

When there is no good answer

Sometimes no article answers the question, or the question touches something sensitive like billing or account security. In those cases the assistant does not guess. It points the customer to the relevant section of the Help Center and offers to open a support ticket.

If the customer wants help from a person, the assistant creates a single support ticket on their behalf, using their signed-in identity and account, and keeps the conversation open with a link to that ticket. Duplicate tickets for the same question are avoided.

Support conversations do not train the assistant automatically

When our team answers a support ticket, that answer might be useful to others. But ticket content often includes private details, and it was written for one person's situation. So ticket replies are never used automatically. A reply can be suggested as a candidate for a new article, with private details removed, but a person must turn it into a draft, review it again and publish it explicitly before the help assistant can use it.

Small details that make it trustworthy

A few design choices help the help feel reliable:

• Conversations persist. An open conversation stays visible as the customer moves between pages, until they close it.
• Typing status is honest. A visible indicator shows when an answer is being prepared, so silence is not mistaken for failure.
• Costs are clear. For customer accounts, the small AI step that selects an article uses the account's AI credits under the same transparent billing as other AI features.

Improving answers means improving articles

Because the assistant only shows reviewed articles, the way to improve its answers is to improve the Help Center. When customers ask questions that no article answers well, those gaps become visible, and our team writes or updates articles to fill them. Each new article goes through the same review before the assistant can use it.

That creates a healthy loop. Better articles help customers who read the Help Center directly, help the assistant choose better answers and reduce the number of tickets our team has to answer one by one. Nothing in that loop depends on trusting the AI to be right about how the product works.

Help that knows where you are

The help bubble also links to contextual help: the articles most relevant to the page the customer is on. Someone on the payment settings page sees payment articles first; someone in the page builder sees builder articles. That context makes the right answer quicker to find, whether or not the customer types a question.

Why grounded help is better help

An assistant that only shows reviewed content is less flashy than one that writes fluent paragraphs about anything. It is also much more trustworthy. Customers can rely on what it shows, our support team can improve answers by improving articles and nobody has to wonder whether the instructions were made up. That fits a broader rule we hold for AI features: refuse to invent facts.

Frequently asked questions

Does the help assistant make up answers?

No. It selects a reviewed Help Center article and shows that article's approved text with a citation.

What if the help assistant cannot answer my question?

It points you to the Help Center and can open a support ticket for you, so a person can help.

Are my support tickets used to train AI?

Not automatically. A reply may become a help article only after private details are removed and a person reviews and publishes it.