How Blaze answers: grounded, cited and honest about what it doesn’t know

An AI agent in customer service has one job: get the customer the right answer, or get them to a person who can. Sounding confident isn’t the same as being right, and a wrong answer delivered politely is still a wrong answer.

We built Blaze around a simple rule: Blaze only answers from content you’ve connected, and it shows its sources. When your content doesn’t cover a question, Blaze says so and brings in your team. This post walks through what happens between a customer pressing send and Blaze’s answer appearing, and the decisions we made along the way.

1. Read the question first

When a customer writes, Blaze waits a moment, in case they’re still typing a second message, and then reads the conversation so far. A fast first pass works out:

• which language the customer is writing in;
• what they’re asking about, and whether the message makes sense on its own or depends on earlier messages;
• whether the message looks like an attempt to give Blaze new instructions, such as “ignore your rules and…”.

Some questions shouldn’t be answered by an AI improvising at all. For those, you can write a custom answer: the exact reply you want sent when a customer asks about, say, refunds or legal terms. Custom answers are checked before anything else.

2. Find the right passages

Blaze searches the knowledge you’ve turned on for it: Help Center articles, internal articles, pages from your docs website, uploaded PDF and Word files, and snippets.

Search is hybrid. Semantic search finds passages that mean the same thing as the question, even when they use different words, so “invoice address” finds an article about “billing details”. Full-text search catches exact terms, like an error code or a feature name, that semantic search can blur. Results from both are combined, filtered to the content this customer is allowed to see, and boosted when they match the customer’s language.

Only passages that clear a relevance threshold are kept, and they’re trimmed to a fixed budget so the answer stays focused. Each one gets a number, which is how citations work later.

3. Write an answer, with sources

Blaze then asks the language model to answer using only those numbered passages. The response is structured, not free text. It contains:

• the answer itself, written in your tone and at the length you chose;
• the numbers of the passages it relied on;
• a confidence level;
• whether a person should take over;
• an optional clarifying question, for when the request could mean two different things.

The answer streams into the Messenger as it’s written, so customers aren’t left watching a spinner. The cited articles appear underneath as links, so customers can read further and your team can check exactly what Blaze relied on.

If the passages don’t contain the answer, Blaze doesn’t guess. It tells the customer it couldn’t find the answer and passes the conversation to your team. Every question it couldn’t answer is collected as a content gap, so you know which article to write next.

Your guidance shapes all of this: plain-language rules such as “be warm and concise”, “ask for the order number before answering a refund question” or “billing disputes go straight to the Billing team”. Rules apply to every answer, in the order you set.

4. Check that it actually helped

Sending an answer isn’t the same as solving the problem. After each answer, Blaze asks whether it helped, with two buttons: That helped 👍 and Talk to a person 👤.

• If the customer confirms, the conversation is marked resolved and closed.
• If they go quiet, Blaze waits for the inactivity period you set, then counts the conversation as an assumed resolution.
• If they ask for a person, or write again with a follow-up Blaze can’t answer, Blaze hands off.

Confirmed and assumed resolutions are reported separately, so you can see how much Blaze really resolves.

5. Hand off with the full story

A handoff should never mean the customer starting over. When Blaze steps aside, it:

• routes the conversation using your rules, for example to the team that owns billing;
• tells the customer, in their language, that a person will reply, and when, if you’re outside office hours;
• leaves an internal note for the teammate summarizing what the customer asked, what Blaze already shared and what’s still open.

If a teammate takes over while Blaze is still writing, Blaze drops its answer and leaves the conversation to them. Customers never get two replies at once.

Safety by design

Customers can type anything into a chat box, so Blaze treats every customer message as content to answer, never as instructions to follow. Its prompts are built to resist injected instructions; it politely declines requests that are off-topic, and repeated attempts are handed to your team. Links in answers are limited to your own sources.

When AI isn’t available, because of a provider outage or because your workspace has used its monthly allowance, Blaze steps aside and the conversation goes to your team. Your customers always get a reply.

Test before you launch

The testing playground lets you ask Blaze anything as a specific customer, before you turn it on. An inspector shows the passages it used, the guidance it applied and how confident it was. Change a rule or an article, ask again, and compare.

Your data stays yours

We never train AI models on your data. Blaze uses an enterprise AI service on paid terms that don’t allow your prompts or answers to be used to train its models. Copilot, the assistant for your teammates, can learn from past resolved conversations, but contact details, payment numbers and secrets are removed before it does.

How Blaze is counted

Each plan includes a monthly allowance of Blaze AI conversations: conversations in which Blaze sent at least one answer that month. A conversation counts once a month, however many messages it takes, and there’s never a separate fee per resolution. The Free plan includes 50 a month; paid plans start at 1,000.

If you’d like to see Blaze answer from your own content, start for free and connect your Help Center or docs website. It takes a few minutes, and you can try everything in the playground before a single customer sees it.