Skip to main content

Build an AI Assistant

An AI Assistant is a Knowledge-Base-powered bot that answers questions in natural language. It uses retrieval-augmented generation (RAG): when a contact sends a message, the Assistant searches the KB folders you've connected, picks the most relevant chunks, and generates a grounded reply.

This guide walks the full journey: create → template → KB → test → deploy.

Before you start​

You'll need:

  • A SendSeven account with the AI Assistant feature enabled (all paid plans).
  • At least one connected channel (WhatsApp, Telegram, Messenger, Instagram, Email) or a website widget where the bot will live.
  • Some content for the Knowledge Base — a website URL, a few PDFs, or a list of FAQs. You can also start blank and add content later.

Step 1: Create the Assistant​

  1. Log in to SendSeven.
  2. Open Automation from the sidebar.
  3. Click Create Bot.

You'll see a template gallery.

Step 2: Pick a template​

Templates are starting points — they pre-fill the personality prompt, greeting message, and recommended settings. You can edit everything afterwards. All AI Assistant templates produce the same kind of bot under the hood; the difference is just the default prompt.

TemplateDefault behavior
Customer SupportFriendly, helpful tone. Escalates politely to a human when confidence is low. Good for general support inboxes.
SalesMore benefit-oriented language. Offers to book a demo or hand off to a sales rep.
FAQConcise answers, high confidence threshold, escalation disabled. Good for self-service knowledge bases.
Email ReplyTuned for asynchronous email replies — longer, more formal, signature-style closings.
CustomBlank slate. You write the personality prompt from scratch.
tip

If you're not sure, start with Customer Support. You can tweak the prompt and confidence threshold once you see how it behaves in real conversations.

Click the template card you want and continue to configuration.

Step 3: Configure the basics​

The configuration form is organized as a small set of tabs. The minimum required fields all live on the Bot Profile tab — a unified tab that merges what used to be the separate Identity, Personality, and General tabs:

  • Name — internal label (e.g., "Support EN", "Sales DE"). Not shown to contacts.
  • Display name — what contacts see in the chat header. Defaults to the bot name. We recommend using something human, like "Sam from SendSeven".
  • Personality prompt — the system prompt that shapes the Assistant's voice and behavior. The template pre-fills this; tweak it to match your brand.
  • Greeting message — sent when a new conversation opens. Skip for email-reply mode.
  • Response language — leave on Auto-detect unless you want to force a specific language.

Live Test Bot panel​

The Bot Profile tab includes a Test Bot panel that lets you chat with the Assistant as you edit. On desktop it docks to the right of the editor; on mobile it collapses into an expandable section at the bottom. Edits to the personality prompt and greeting message are reflected immediately on the next test message — no save round-trip required.

For tuning advice on the personality prompt, see Prompt Engineering.

AI disclosure (EU AI Act)​

If you operate in the EU, Article 50 of the EU AI Act requires disclosing to users that they are talking to AI. SendSeven Assistants include an AI disclosure message field — leave the default on unless you have a specific reason to remove it.

Step 4: Connect Knowledge Base folders​

Open the Knowledge Base tab on the bot editor.

The KB is organized into folders, each of which is a separate Vertex AI RAG corpus. You attach one or more folders to the Assistant; at query time the Assistant searches all attached folders in parallel and combines the results.

Folder typeWhen to use
GeneralYour default folder. Product docs, policies, generic FAQs.
WebsitesContent imported via website crawl. Refreshed automatically.
Ticket SummariesAuto-generated from resolved support tickets (when ticket-summarization is enabled).
Custom folderPer-product, per-region, or per-language splits. Each becomes its own corpus.

Empty state​

If you've never added KB content, the editor will show an empty-state card:

This Assistant has no Knowledge Base content. It will only respond using its personality prompt. [Add documents] [Crawl a website]

You have two options:

  • Add documents — upload PDFs, DOCX, TXT, or Markdown files; or paste text FAQs and manuals directly.
  • Crawl a website — give the crawler a starting URL and (optionally) a sitemap. The crawler imports pages into the Websites folder. Recrawl daily / weekly / monthly via the schedule control.

See the Knowledge Base guide for the complete content-strategy guidance.

note

You can attach the same KB folder to multiple Assistants. There's no duplication cost — the corpus is shared.

Step 5: Tune behavior​

Open the Behavior tab.

  • Confidence threshold (0.00–1.00, default 0.70) — minimum confidence required to send a reply without escalating. Lower = bot answers more often but with less certainty. Raise it for FAQ bots, lower it for sales.
  • KB-only mode — if on, the Assistant refuses to answer anything not supported by KB sources. If off, it falls back to general knowledge from the underlying LLM. Use on for legal/medical/compliance bots, off for everything else.
  • Escalation — when on, the bot offers to hand off to a human after max_failed_attempts low-confidence answers in a row, or when the user types one of the escalation keywords (default: "human", "agent", "support").

See Prompt Engineering for tuning recipes.

Step 6: Test in the playground​

Open the Test tab and chat with the Assistant. Each test message shows:

  • The bot's reply
  • The confidence score (0.00–1.00)
  • The KB chunks that were retrieved, with their relevance scores
  • Whether the bot would have escalated in a real conversation
  • Token usage

Use this to verify:

  1. The bot answers known questions correctly.
  2. The bot escalates (or politely declines) for things outside the KB.
  3. The tone matches your brand.

If answers are off, the fix is almost always in the KB content (Step 4) or the personality prompt (Step 3). Confidence threshold and KB-only mode are secondary levers.

Step 7: Publish your bot​

The wizard now ends with a "Publish your bot" step — this is where deployment happens. (Previously the wizard opened with deployment and ended with activation rules; the order was flipped so you can shape the bot's personality and KB before deciding where it lives.)

The Publish step replaces what used to be a separate Activation Rules front-loaded step. From here you have two deployment paths:

Option A — Website widget​

If you want the Assistant to power live chat on your site, attach it to a widget. The Publish step lets you:

  • Attach to a new widget — creates a fresh Support widget pre-wired to this bot. You'll get the embed snippet on the next screen.
  • Attach to an existing widget — pick from your existing support widgets. (Newsletter widget types are no longer offered in this picker — newsletter widgets don't run bots.)

If you pick an existing widget that already has a bot attached, the wizard prompts you to either:

  • Replace the existing bot — the previous bot is detached from this widget. It remains as a configured bot and can be reused elsewhere.
  • Create a new widget instead — opens the new-widget shortcut so both bots stay live on their own widgets.

Option B — Channel (WhatsApp, Telegram, Messenger, Instagram, Email)​

Open the Activation Rules tab on the bot editor and add a rule:

  • Pick the channels the bot should be active on (by channel ID, not by channel type — so you can have one Assistant on your DE WhatsApp and a different one on your EN WhatsApp).
  • Optionally restrict to a schedule (e.g., weekdays 09:00–17:00 Europe/Berlin).
  • Optionally set Require no agents online if the bot should only fire outside human-staffed hours.
  • Set priority if multiple bots could match — lower number wins.

See Deploy to Channels for the full activation-rules reference.

Step 8: Activate the bot​

Toggle the Active switch on the bot. The bot will start handling incoming messages on the next inbound conversation that matches its activation rules.

warning

Activation rules are evaluated at message time, not at activation time. Bots will not retroactively claim conversations that started before they were activated.

What's next​