import { Steps } from '@astrojs/starlight/components';
import Figure from '@/components/Figure.astro';

**Who it's for:** any business that wants to solve customers' repeat questions without a person having to
be available all day, keeping your team's time for what genuinely needs a person.

## What the end customer sees

The customer writes on WhatsApp, Telegram or your website widget and gets answers based on the
documentation you uploaded — never made-up information. If they ask to talk to a person, or an escalation
rule your admin configured fires (certain keywords, or too many turns without resolving the question), the
AI says so plainly ("I'll let a member of the team know…") and stops answering from that message on. They
are never left unanswered or unexplained, on any channel, day or night: the AI answers 24/7, and if
something escalates outside your working hours, it waits until someone on your team picks it up.

When a person takes control, the customer gets a notice ("you're now talking to `{your name}`, a member of
our team") — they never change who they're talking to without knowing it. And if that person decides to
hand the conversation back to the AI without closing it, the customer gets the hand-back notice too.

## Steps in the product

<Steps>

1. **Upload documents to your knowledge base.** Drag your files (PDF, DOCX, MD, TXT, CSV or HTML) into
   **Knowledge base → Documents**; optionally tag them to control which agent can use them — untagged
   documents get "General" automatically. See
   **[Knowledge and documents](/en/guides/conocimiento-documentos/)**.
2. **Test what the AI would retrieve, before a real customer sees it.** In the **Playground** tab, write a
   sample question and pick an agent (or "Global") to see exactly which document fragments it would use to
   answer. If information is missing, go back and add documents.
3. **Adjust your agent's escalation rules.** In **AI agents**, configure which keywords or how many
   unresolved turns trigger a hand-off to a person. See **[AI agents](/en/guides/agentes-ia/)**.
4. **A customer writes and the AI answers from your documents**, or escalates if the customer asks or a
   rule fires — the conversation shows up in **Queue** (or "All") waiting for someone on your team.

   <Figure
     src="/captures/inbox-conversacion.png"
     alt="An open conversation still handled by the AI: Unclassified/No analysis labels in the header, a Take over button, a notice that the AI is handling it, and a disabled composer, with the contact card on the right"
     caption="While the AI answers, the composer stays disabled — the Take over button is all you need once a case escalates."
   />

5. **Open the conversation and press "Take over".** The composer becomes enabled, the conversation switches
   to HUMAN mode, and the AI stops stepping in until you resolve it or hand it back yourself. See
   **[Conversations](/en/guides/conversaciones/)**.
6. **Reply to the customer.** Read the history first so you don't repeat questions the AI already asked.
7. **Press "Hand back to AI" or "Resolve"**, depending on the case: handing back keeps the conversation
   open but the AI answers again; resolving closes it entirely.

</Steps>

## What to configure

- **Knowledge base**: documents uploaded, tagged if you have several agents with different scopes.
- **Escalation rules**: keywords or the unresolved-turns threshold, per agent.
- **Channel owners**: who gets notified (bell, email, and WhatsApp if you turn it on) when a conversation
  on that instance escalates.

## Limits and honesty

- The AI **never invents** facts about your business: if it's not in your knowledge base, it doesn't know
  it.
- An escalation always lands **QUEUED**, visible to anyone on your team — it isn't auto-assigned to one
  specific person unless your instance has channel owners configured.
- The Playground is a test view: it doesn't generate a chat reply, it just shows you which fragments the
  AI would retrieve.

## See also

- **[Knowledge and documents](/en/guides/conocimiento-documentos/)** — upload, tag and test in the
  Playground.
- **[Conversations](/en/guides/conversaciones/)** — take over, hand back to AI and resolve.
- **[AI agents](/en/guides/agentes-ia/)** — escalation rules, field by field.
- **[Lead capture via WhatsApp and web](/en/use-cases/captacion-leads-whatsapp-web/)** — the step before
  this one, while the conversation is still a fresh opportunity.