Documentation
The AI agent
The AI agent reads each incoming message, looks up the relevant order in your store, and writes a reply grounded in what it actually found. You decide how much it is allowed to do on its own — from drafting every reply for approval, to sending and acting without you.
How a reply gets written
When a message arrives, the agent works through roughly these steps:
- Classify. It works out what the message is about — a shipping question, a return, a cancellation, a product question — and how confident it is.
- Gather facts. It calls the tools it needs to answer properly: fetching the order, checking fulfilment and tracking, looking up a product or a refund status. These read your live store, not a cached copy or a static FAQ.
- Route to a specialist where one applies — a focused configuration for a particular kind of request, with its own instructions and its own restricted set of tools.
- Compose the reply from what it found, in your brand voice and following your instructions.
- Send it, or hold it as a draft — depending on the mode you have chosen for that brand.
Because the answer is built from live store data rather than a fixed answer bank, it stays correct when an order ships, a fulfilment updates or a refund lands.
Seeing exactly what it did
Every conversation the AI touched has an AI Activity panel in the customer sidebar. Expand it and you get:
- Category and the confidence score behind the classification.
- AI Turns — how many times it has replied in this conversation.
- A Timeline of the run, in order. Each tool call is listed by name —
get_order_information,get_tracking_information— with how long it took in milliseconds, a tick or cross for success, and the time it ran. You can expand any tool entry to see the arguments it was called with and the result it got back. - The AI Response entry, marked (Draft — pending approval) while it is still waiting on you. Where the model produced reasoning, a Model reasoning section is available.
- Escalated to human, with the reason, if it handed the conversation over.
This is the answer to “why did it say that?”. If a reply is wrong, the timeline usually shows whether the AI misread the request or whether the underlying data it fetched was already wrong.
Draft mode and auto-send
Each brand runs in one of two modes, and you can change it at any time.
Draft mode
The AI writes the reply and stops. It appears in the composer under an AI Draft — Review before sending banner, and the conversation lands in the AI Drafts view. You can:
- Approve — send it as written, or Approve & Close to send and close together.
- Edit then approve — change anything first; your edits are kept.
- Reject — discard it, optionally saying why, which feeds back into improving future drafts.
Important: in draft mode the AI does not carry out actions
Refunds, discounts, cancellations and photo-evidence requests are not run by the AI while a brand is in draft mode — those tools are removed from its toolset. The product tells you this directly above the draft: if the draft promises any of those actions, you need to action them yourself after sending. Read the draft for promises, not just for tone.
Auto-send
The AI sends its reply itself, and — for the tools you have enabled — can carry out the action it describes. Most teams start in draft mode, watch a few hundred conversations, and switch on auto-send once they trust the replies for their common request types.
What the AI can do
The agent has a catalogue of tools, and each one is enabled or disabled by you. Nothing is on by default simply because it exists. Broadly they fall into:
| Kind | Examples |
|---|---|
| Looking things up | Order details, tracking and fulfilment, product information, refund status, return address, order date |
| Managing the conversation | Adding tags, leaving an internal note, snoozing, closing, scheduling a status update |
| Acting on an order | Processing a refund, creating a discount code, cancelling an order, registering a return, requesting photo evidence |
| Handing over | Assigning the conversation to a human team |
Order-affecting tools have their own limits — for example refunds can be capped by category and by amount, so the AI can settle a small damaged-item claim while anything larger waits for a person. See Returns & refunds.
When it hands over to a human
The agent escalates rather than guessing. It hands a conversation to your team when it is not confident, when the customer asks for a person, when the request falls outside what it is configured to handle, when a tool it needs is unavailable, and when a topic is one you have reserved for humans.
Handover reasons are recorded against the conversation, so over time you can see which reasons dominate and decide whether to close those gaps with configuration or leave them with your team.
It will not claim a human is taking over when none is. Whether the AI identifies itself as an AI is a separate, explicit setting per brand — always, only if the customer asks, or never — but the promise of a human handover is tied to an actual handover.
Configuration sets and specialists
A configuration set is a complete AI setup: instructions, enabled tools, specialists and behaviour. Having more than one lets you keep a known-good configuration live while you build the next one.
You can point email and live chat at different configuration sets for the same brand, which is useful because the two channels want different lengths and tones — chat wants short and immediate, email tolerates thorough.
Specialists are focused sub-agents inside a configuration set. Each has its own instructions, its own allowed tools and its own trigger keywords, so your returns specialist can be strict and narrow while your general specialist stays broad. Specialists can be enabled, disabled and prioritised individually.
Reviewing quality over time
The QA section samples handled conversations and scores them against a rubric you define, so quality is measured on a sample of real traffic rather than on the handful of conversations someone happened to notice. Combined with the handover reasons and the reports in Reports, that is how you decide what to change next.
What the AI will not do
- It will not invent order facts. If it cannot retrieve an order it says so or hands over — it does not guess a tracking number.
- It will not carry out refunds, discounts, cancellations or photo-evidence requests while the brand is in draft mode.
- It will not exceed the refund limits you set, and it will not use a tool you have not enabled.
- It will not claim a colleague is picking the conversation up unless it has actually handed over.
- Your customers' data is not used to train AI models. See our privacy policy and Security & privacy.
Getting started with it
- Start the brand in draft mode.
- Enable the read-only tools first — orders, tracking, products — so the AI can answer accurately before it can act.
- Work the AI Drafts queue for a week and reject with reasons where it is wrong.
- Turn on order-affecting tools one at a time, with conservative limits.
- Move to auto-send when the drafts for your common request types are consistently sendable as written.
Questions about a specific reply? Email hello@squidsupport.ai with the ticket number — the AI Activity timeline gives us everything we need to explain what Squid Support did and why.