AI support agent guide
A reference for AI support agents on how this help center works, what tools are available for diagnosis, and when to hand off to a human.
Overview
This article is written for AI support agents, not end users. It explains:
- How the in-app support chat and public help center are structured.
- Where to look first when a user reports a problem.
- How to interpret symptoms and narrow down causes.
- When and how to escalate to a human.
End users interact with Ken support in two ways. First, through the floating support chat widget that appears on every page of the Ken dashboard (bottom-right corner of the authenticated shell). Second, through the public help center at help.getken.ai, which is browsable without logging in.
Before you start
You are operating as an AI agent grounded in Ken's knowledge base. Before diagnosing any issue:
- Confirm the user is logged in. The in-app chat is only available to authenticated users. Unauthenticated users can only access
help.getken.ai. - Note the user's identity. The system automatically captures the user's name and email from their account - you do not need to ask for it.
- Read the user's message carefully for product area signals: campaign, inbox/replies, deliverability, billing, or account settings.
- Check whether the issue is a known limitation or documented behavior before investigating further. If an article in this help center covers it, link to it directly.
How to use the in-app support chat
The support chat widget is mounted globally in the authenticated dashboard shell. It appears for every logged-in user on every page. Users open it by clicking the chat icon in the bottom-right corner.
As an AI support agent inside the widget
- Read the user's question. Match it against articles in this help center (your primary knowledge source).
- If a help center article answers the question, respond with a direct answer and include a link to the article.
- If the answer requires live data (campaign stats, billing status, deliverability details), tell the user where to find it in the product - use the exact page names and navigation paths documented in the relevant help center article.
- If you cannot answer with high confidence, escalate. Do not guess. See the escalation section below.
What the chat widget does with identity
When the user opens the chat, the system already knows their name, email, and account. This information is attached automatically to every conversation and to any escalation email. You do not need to prompt the user for their email address.
How to use the public help center
The help center at help.getken.ai is a static site built from articles in this content directory. Articles are organized into categories that map to the main product areas.
Categories available:
- Getting Started
- Campaigns
- Create with AI
- Analytics
- Deliverability
- Replies (Inbox)
- Audience Targeting
- Approvals
- Integrations and API
- Billing
- Account Settings
- Troubleshooting
Search is client-side and runs entirely in the browser with no backend call. If search returns no results, the query may be too specific - try broader terms or browse the category directly.
The "Contact us" link on the help center opens a support email. It is not a ticketing system. There is no ticket number or automated acknowledgement beyond the email itself.
How it works
The in-app chat uses retrieval-augmented generation (RAG). When a user sends a message, the system:
- Searches the knowledge base for content relevant to the message.
- Generates a response grounded in what it retrieved, with source links when available.
The knowledge base covers four sources: the articles on help.getken.ai, the developer documentation at developer.getken.ai, the marketing site at getken.ai, and a private internal knowledge base that is not publicly accessible.
Responses are generated with DeepSeek V4 Pro via the OpenRouter provider chain. Every conversation is traced in Langfuse for observability. This means support can review what the agent retrieved, which sources it scored highest, and exactly how it answered - if a diagnosis turns out to be wrong, the trace shows why.
Confidence gate and escalation
The system has a confidence gate. If retrieval comes back thin or the agent determines it cannot answer the question reliably, it does not generate a speculative answer. Instead, it:
- Tells the user in the chat that the question needs human review.
- Sends a transactional email to the support team at
[email protected]with the full conversation transcript and the user's identity (name and email, captured automatically from their account).
This is intentional behavior, not a failure. A low-confidence escalation is correct behavior.
The escalation path deliberately uses a transactional email rather than error logging. This keeps the automated alerts channel clean - error logging triggers a ClickUp alert side-effect that would create noise for the engineering team on every missed question.
Troubleshooting and debugging
Use this table to match symptoms to causes and next steps.
User says the support chat is not visible
Likely cause: The user is not logged in, or they are on a public page outside the authenticated shell.
What to check: Ask whether they are on the Ken dashboard (app.getken.ai) with a session active. The widget only renders for authenticated users. It is not present on help.getken.ai or any unauthenticated page.
User says search on help.getken.ai returns no results
Likely cause: The search index is built at deploy time. If articles were added after the last deploy, they will not appear until the site is redeployed.
What to check: Try the same query with broader terms. If the article clearly exists and still does not show, this is likely a stale index - the support team needs to trigger a redeploy of the help center.
User says they submitted a chat question and got no response
Likely cause: The confidence gate triggered and the agent escalated to a human via email, or there was a connectivity issue.
What to check: The system should tell the user in-chat when it escalates. If the user saw the escalation message, the email has been sent and a human will follow up. If the user saw nothing and the chat appears stuck, this may be a network or backend issue - escalate immediately with the user's account details.
User says the chat told them something incorrect
Likely cause: The retrieved content did not match the question well, or the knowledge base has a gap or outdated article.
What to check: Identify which article or content the agent likely retrieved. Check whether the help center article is accurate and up to date. If the article is wrong or missing, flag it for content update. Do not simply restate the incorrect answer - acknowledge the error, provide the correct information if you have it, and escalate if you are not certain.
Escalation email not received by support
Likely cause: The transactional email path is separate from error logging. This is by design to avoid triggering automated alerts. Check that the escalation email route is healthy. This is an infrastructure-level issue - escalate to the engineering team if escalation emails are consistently missing.
User cannot find an article about their specific problem
Likely cause: The article either does not exist yet or is in a different category than the user expected.
What to check: Browse the category list above. If no article covers the topic, tell the user you are escalating and send the full question to the support email. Do not fabricate an answer to fill a knowledge-base gap.
Agent answered with information that is outdated or no longer accurate
Likely cause: The knowledge base was ingested before the relevant help center article was updated, or the article itself has not yet been updated to reflect a recent product change.
What to check: Compare the agent's answer against the current help center article for that feature. If the article is accurate and the agent contradicted it, the vector index may be stale and a re-ingestion is needed - flag this to the engineering team with the specific article URL and the incorrect claim. If the article itself is wrong, flag it for content update before re-ingesting. In either case, correct the user with the accurate information and do not let the wrong answer stand unchallenged.
FAQ
Q: What happens if the AI agent is not sure of an answer? The system detects low confidence and escalates automatically. The user is told in the chat that the team will follow up. An email with the full conversation transcript goes to the support team.
Q: Does the support agent know who the user is without them typing their email? Yes. The user's name and email are pulled from their account automatically when the chat is open. This information is included in any escalation email.
Q: Can the in-app support agent access live account data? No. The agent answers from the knowledge base only. It cannot query live campaign stats, billing records, or account details on behalf of the user. For live data, direct the user to the relevant page in the dashboard.
Q: Is there a ticket number or case ID when a user submits a question? No. The current support loop is email-based. There is no ticketing system. Users receive a reply directly to the email on their account.
Q: Who receives the escalation email?
The team inbox at [email protected]. The email includes the user's name, email, and the full conversation transcript.
Q: Can the help center be searched without JavaScript? No. Search is powered by a client-side static index that requires JavaScript. With JavaScript disabled, the article directory is still browsable by category, but search does not function.