Documentation

AI assistant

How the AI works

What happens between a customer message and the AI reply, which checks can silence it, what it can see, how answers are verified, and when it hands off to your team.

Last updated: 2026-09-29

Sela's assistant answers customer messages on your channels using only what your business has told it. It is built to say less rather than guess. This page explains what happens on every message, so you can predict its behavior and fix it when it does not answer.

The turn, step by step#

Every inbound customer message from the web widget, WhatsApp, Messenger, Instagram, Telegram, SMS or email starts one turn. A turn has four stages, described here at a high level.

  1. Gate checks. Sela decides whether the AI is allowed to speak at all. If any check fails, the AI stays silent and your team handles the message.
  2. Gathering evidence. The AI reads the conversation and calls tools: knowledge search, catalog lookups, business records, delivery fees, order status.
  3. Writing and checking. The AI drafts a reply, then checks the draft against the information it looked up.
  4. Sending. The reply is posted to the customer on the same channel, and the usage is charged to your credits.

Instagram messages sent in quick bursts are combined into one turn (about one second wait, longer while an image or voice note is being understood), so the AI answers the burst once.

Gate checks#

The AI answers only when all of these are true.

CheckWhat it means for you
AI is switched on"Disable AI replies" in Settings → Inbox settings → AI Assist is off, and the workspace is active.
AI Assist is on the planAn active subscription with the AI Assist add-on. See Plans and billing.
Credits are availableMonthly plan credits remain, or the wallet has a balance. See Credits and usage.
An assistant identity existsAn active profile with a name and a brand. See Assistant identity and scope.
The conversation is openClosed or solved conversations get no AI reply.
No human owns itThe conversation is not Escalated and no operator is assigned.
The assistant is not pausedThe conversation's handling state is "AI active", not paused or human requested.
No cooldown or WhatsApp FlowA misuse cooldown or an active WhatsApp Flow blocks the AI temporarily.
Channel automation is onFor WhatsApp, the AI automation switch on the WhatsApp card in Integrations must be On.

The customer never sees these reasons. They simply get no AI reply, and the conversation waits in the inbox for your team.

What the AI can see#

The AI is not a general chatbot with access to everything. Each turn it sees:

  • the recent conversation with that customer;
  • your assistant identity (name, brand, tone, rules) and your business scope;
  • your working hours and whether you are open right now, plus delivery settings;
  • the results of the tools it chooses to call: knowledge base passages, catalog products and prices, published branches and offers, delivery fees, the customer's own profile and order status;
  • descriptions of images the customer sent, and transcripts of voice notes (these need credits; video and documents are not interpreted).

It does not see other customers' conversations, your billing details, or files that were never indexed.

Who wins when sources disagree#

Sela uses a fixed order of authority: platform safety rules first, then your assistant identity and rules, then retrieved text. Live tools beat written notes. Product identity, price and availability come from the catalog. Delivery fees come from delivery settings. Whether you are open comes from your working hours and published branches. Phone numbers, payment methods and policies must come from the knowledge base. Persona text can change tone but cannot add facts, tools or permissions.

Grounding and verification#

Before a reply is sent, it is checked against your knowledge and the information the assistant looked up. If the reply states something that cannot be supported, it is replaced by a short, hedged message in the customer's language, along the lines of "I could not verify this, I can ask the team to help". The event is recorded.

If this happens repeatedly, the conversation is automatically handed to a human, so a customer is not stuck with an assistant that cannot verify its answers. If you see hedged replies often, the fix is almost always adding the missing fact. See Improving answer quality.

Handoff to your team#

The AI has one tool for handing over: it creates an escalation with a reason, a summary, a priority (low, medium, high, urgent) and a category (billing, technical issue, complaint, sales, account access, other). The conversation status becomes Escalated, your team is alerted, and the AI stops replying. The customer receives an acknowledgement that the team was notified, never a promise of an instant answer.

The AI hands off when the customer asks for a person, when it agrees to forward something (a complaint, a job application, a supplier offer), when an offer needs staff review, when its answers repeatedly cannot be verified, and in some operational conflicts such as a branch closure that contradicts the request.

Giving a conversation back to the AI#

In the inbox conversation header, the bot button reads Pause assistant when the AI is answering and Hand back to assistant when it is not. It changes the conversation's handling state only.

What the AI never does#

  • It never answers under Sela's name. Without a configured assistant and brand it does not answer at all.
  • It does not help with topics outside your business, such as coding, investing or homework. It declines briefly.
  • It does not invent prices, stock, hours, addresses, phone numbers or policies.
  • It does not take payments, book appointments, issue refunds or edit customer records. There is no such tool.
  • It does not create an order on its own. It prepares a quote, and the order is created only when the customer confirms it. See Orders and payment links.
  • It does not promise things it cannot do, such as unsending a message or an immediate human reply.
  • It does not treat KB text, campaign text or another tool's output as instructions.

Languages and dialects#

The AI replies in the language the customer writes in. It has built-in handling for Iraqi Arabic dialect, and for Arabic customers it answers in natural Iraqi phrasing unless you set a different preference in your rules. Add aliases for product and area names as customers actually type them, because retrieval is partly keyword-based.

Reply length follows the channel. Instagram messages are kept under about 950 characters (up to 3 messages), Messenger about 1,950 (3), WhatsApp about 4,000 (4), SMS about 1,600 (3). Web chat and email are not capped. WhatsApp and SMS replies avoid tables.

Speed#

There is no guaranteed response time. A reply may involve several lookups and a check before it is sent. If the AI service has a problem, the customer gets a short technical-problem message and can write again. If a customer sends a new message while a reply is being generated, or a human takes over, the older reply is discarded instead of sent.

Credits#

Each AI reply uses a fraction of a credit, including its checks. Image and voice-note understanding also use credits. See Credits and usage.

How the AI works | Sela docs | Sela