> ## Documentation Index
> Fetch the complete documentation index at: https://docs.apollospace.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Retail & e-commerce
> Support, post-sale, order status, cart recovery, and catalog work at volume — with refunds and order-term changes always staying human.
It's 9pm on a sale Tuesday. WhatsApp has fourteen "where's my order?"
messages, a customer wants to swap a t-shirt size, three full carts were
abandoned in the last hour, and a one-star review on the store has gone
unanswered for two days. Each of those messages is a sale cooling off or a
customer souring — and they all landed at once, at the hour when nobody on
the team is watching.
**What leaves your plate:** answering order-status and post-sale questions,
pulling back whoever abandoned their cart, routing returns and exchanges to
the right owner, and keeping the catalog tidy all start happening on their
own — with context, within budget, and with **money and order changes always
waiting on a human**.
## How it works, in one line
Retail is **repetitive volume on top of a record that already exists**: the
order, the customer, the product. The agent invents nothing — it reads the
record, answers from what's there, and only acts alone when the action
touches no one's wallet.
```mermaid theme={null}
flowchart LR
A[Trigger
message, stalled cart,
review, schedule] --> B[Agent pulls context
from Company Brain + CRM + order]
B --> C[Decides + uses tools
Chats, WhatsApp, CRM, Routines]
C --> D{Touches money
or the order's
terms?}
D -->|No| E[Answers / routes / organizes]
D -->|Yes| F[Proposes + calls a human
with full context]
E --> G[Becomes memory
the next interaction already knows]
F --> G
```
## Concrete flows
### 1. Order status and post-sale — answered from the record itself
**Trigger:** a "where's my order?" or a post-sale question lands by
[chat](/en/features/chats) or [WhatsApp](/en/integrations/whatsapp).
The agent identifies the customer, pulls the order record from your store or
ERP integration (via [Composio](/en/integrations/composio)), and cross-checks
it with what the [Company Brain](/en/features/brain) already knows — delivery
windows, shipping policy, how to read the tracking code. It replies on the
spot, in the customer's channel and language: where the package is, when it
arrives, what to do if the box came wrong. Every interaction is logged as an
[activity in the CRM](/en/features/crm), so that customer's history is always
current.
When the answer **isn't** in the record or the Brain — or when the case needs
judgment — it stops, doesn't guess, and **hands off to a human** with
everything chewed: who they are, which order, what it tried.
The win isn't "a bot that answers everything." It's that "where's my
order?" — retail's most repeated and most easily answered question — stops
consuming the team, and a person only steps in when the case truly needs
one.
### 2. Cart recovery — the re-engagement nobody remembers to do
**Trigger:** a [routine](/en/features/rotinas) the agent schedules for itself
— it sweeps carts and checkouts stalled for X hours.
On fire, [Marcus](/en/agents/marcus), the communications specialist, pulls
from the [CRM](/en/features/crm) and the store record who abandoned, drafts a
personalized recovery message — the product left in the cart, the right name,
the store's tone from the [Brain](/en/features/brain) — and sends it on the
customer's channel ([email or WhatsApp](/en/features/outbound)). Replies flow
back into the CRM as activity; anyone who reopened their cart becomes a warm
signal for the sales team.
Cart recovery is external communication at volume. Keep **human approval
before each send** at first — the agent prepares the whole queue, you review
the tone and release it. Autonomy grows like a ratchet: once it's reliable,
you let the batch go. And any **discount or coupon** that changes the price
stays a proposal for a human to approve, never the agent's call.
### 3. Returns and exchange triage — to the right owner, no restart from zero
**Trigger:** a customer requests an exchange, a return, or complains about a
defect.
The agent reads the order, classifies the case — size swap, defective
product, change-of-mind within the window — and opens a
[task](/en/features/boards) already **assigned to the right owner**:
logistics for the pickup, quality for the defect, support for the sensitive
case. The task carries the whole context: order, reason, customer history,
photos that came in the conversation. Whoever takes it doesn't ask "which
order was this again?" — they open the card and see it all.
```mermaid theme={null}
flowchart TD
A[Return/exchange request] --> B[Agent reads the order
+ classifies the case]
B --> C{Type}
C -->|Size swap| D[Task → logistics]
C -->|Defect| E[Task → quality]
C -->|Refund| F[Proposes → human approves]
D --> G[Becomes memory]
E --> G
F --> G
```
The agent runs everything up to the money line. **Refunds, chargebacks, and
any credit it proposes — a person releases.**
### 4. Tidy catalog and review replies — ready to approve
**Trigger:** a new product arrives, a record is incomplete, or a review pings
on the store.
For the catalog, the agent keeps records tidy in hierarchical
[folders](/en/features/pastas) — by category, collection, supplier — flags
records missing a description, photo, or price, and drafts what's missing from
the [Brain](/en/features/brain) and the manufacturer's page
([Scout](/en/agents/scout) researches what's public). It all becomes an
editable [document](/en/features/documentos): the agent assembles, you review.
For reviews, the agent **drafts the reply** — thanks the praise, acknowledges
the problem, points to the next step — in the store's tone, and leaves it
**waiting for your approval** before publishing. A public reply never ships
without a human looking.
**Why this matters:** a messy catalog and an unanswered review are the two
jobs that always slide to "later" — and the two that cost the most in
conversion and reputation. The agent keeps them current as a routine, and
leaves you only the "yes, publish it."
## What stays human
The agent **proposes**, never decides alone, whenever the subject is:
* **Refunds and any money** — chargeback, credit, invoice adjustment,
charge. Every financial move goes through approval.
* **A promise that changes the order's terms** — deadline, free shipping,
out-of-policy exchange, any condition that becomes an obligation to the
customer.
* **Discounts and coupons** — anything that alters the final price.
* **A commercial proposal to a customer** and heavy external, legal, or
regulatory communication.
* **Governance decisions and destructive actions** — deleting records,
canceling orders in bulk.
In those cases the agent prepares everything — the draft, the context, the
recommendation — and **waits for a person to approve.** Every action lands
in the auditable trail, with author and justification, and spend is capped
by [Stars](/en/billing/stars) before the call ever happens.
## Where to start
Wire up [WhatsApp](/en/integrations/whatsapp) and connect your store or
ERP via [Composio](/en/integrations/composio). That's where the agent
reads the order and answers "where's my order?" without depending on
anyone.
Upload your shipping, exchange, and return policy into the
[Company Brain](/en/features/brain), plus the post-sale answers you repeat
every week. Without it, the agent has nothing to pull a first response
from.
Create a [routine](/en/features/rotinas) that sweeps stalled carts and let
the agent assemble the recovery queue — with human approval before each
send at first.
## Next steps
The support and post-sale pattern that holds up high-volume relationships.
Where every customer, order, and interaction stays alive — the history
that doesn't depend on human memory.
The channel where retail happens — support, status, and recovery in the
same place the customer already is.
The cron behind cart recovery and check-ins — the agent fires on the right
day.