REFERENCE · SIDEKICK
A guide for the team

Sidekick, always listening.

Sidekick is the conversation-intelligence layer that runs over every conversation in Atender. It reads what’s being said in real time, understands it, and routes that understanding to whoever — or whatever — is on the other end. For human agents, it shows up as a panel of suggestions and matching knowledge. For the system itself, it works in the background — tagging, extracting structured data, spotting incident matches.

i
Not a panel. A layer.
The slide-out panel agents see is one of Sidekick’s outputs, not the whole of it. Tagging, summarisation, entity extraction, and incident matching run for every conversation — whether a human is looking or not. As Agent Stacks take on more conversations, the same understanding feeds them too.
The core idea

What Sidekick understands

Every customer message kicks off a fresh read. Sidekick extracts three kinds of understanding, and that understanding is the same regardless of who’s handling the conversation. What changes is where it goes next.

1Three reads, every message — what Sidekick pulls out of every inbound
≡
Topic
What the customer is asking about, what’s already been tried, where things stand.
#
Entities
Structured data inside the messages — order IDs, serial numbers, emails, phone numbers, anything you teach it to recognise.
◇
Patterns
Relationships across conversations — this looks like an active incident, this contact has three other open threads, this should be tagged.
That same understanding is then routed two ways ↓
For humans

Surfaced in the panel

When an agent opens a conversation, Sidekick appears as a slide-out panel — a stack of modules and, when something is up, a hint card on top. The set of modules grows over time; what’s here is what ships today.

1Panel modules — drag-and-drop, hide what you don’t need
Summary orient on open

Two to four sentences, no fluff

What the customer is asking about, what’s already been tried, and where things stand. No greeting, no sign-off. Just the orientation an agent needs before reading the thread.

Suggested response click to insert

A drafted answer, with sources cited

When Sidekick is confident, it drafts a reply based on the conversation, the Handbook, and the Knowledge Base. Sources are clickable footnotes. One click drops the draft into the reply editor — edit as needed or send as-is.

Handbook articles internal — agent only

Internal procedures the agent needs

How to handle cancellations, what the return policy is, escalation steps. The agent sees the procedure; the customer never does.

Knowledge articles customer-facing

Public articles to share with the customer

Matching pages from your Knowledge Base. Link them, quote them, or open them inline. The same content that a customer would find by searching the public site.

Relevant links references

External and internal references

Documentation pages, help-center articles, internal tools — surfaced from the procedures Sidekick matched. A small useful pile, not a search result page.

More, over time module slots

The panel is alive

Modules are independent units. New ones are added as the product grows; existing ones can be reordered or hidden per tenant. Don’t treat the list above as fixed.

And on top of the panel, two cards that pop when something material is happening ↓
2Hint cards — sit above the modules when relevant
Possible incident match link · subscribe · reply

Recognise the outage in one place

When a customer’s description matches an active incident, the card shows the incident, severity, status, and a confidence score. Two actions: Link & subscribe attaches the conversation to the incident and queues an automatic update when it resolves; Insert reply drops a templated acknowledgement into the editor.

Multiple active conversations same contact · merge

See the other open threads in one click

When the same contact already has other conversations open, the card lists them with their channel and a link to jump in. A merge action is available when consolidating is the right move.

The hint cards aren’t modules. They appear and disappear based on what’s actually happening in the conversation. An incident card only shows if a real match is found; the duplicate card only shows if the contact has other active threads. The panel is for help-on-tap; the hint cards are for help-when-it-matters.
For the system

Quiet workers in the background

Most of what Sidekick does is invisible. Workers run on every conversation regardless of whether a human is reading the panel — they tag, extract, populate fields, and feed the rest of the system.

1Background workers — always on, never wait for a human
Detection rules pattern · AI · regex

Capture structured data without asking the agent

Each rule scans messages for one kind of value. Three methods to choose from: pattern builder (prefix, suffix, length, character set — pre-made templates for serial numbers, emails, phone numbers, order IDs), AI (describe what to look for in plain language), or regex for the cases where you want a raw expression. Matches highlight inline in the conversation, and high-confidence matches auto-save to a custom field on the contact, the conversation, or both.

Auto-tagging embedding-based

Tag conversations by meaning, not keywords

Sidekick reads the conversation and matches it against tags you’ve opted in for AI tagging. Matching is semantic — a refund question gets the refund tag even when the word “refund” isn’t in the message. A confidence threshold and a max-candidates cap keep noise low.

One slider, two outcomes. The confidence threshold turns a probability into a decision: above the line, a value auto-saves and is highlighted green; below, it’s flagged amber for the agent to confirm. Move the slider to trade volume for accuracy.
All of it configured in one place ↓
2What you tune — four tabs in Settings · Sidekick
⚙
General
Master switch, incident discovery, inline highlights, auto-save, confidence threshold.
⋮⋮
Panel layout
Drag modules into order. Hide what you don’t need.
▦
Detection rules
Build, edit, reorder, and map rules to custom fields.
◷
Auto-tagging
Toggle the worker, tune its threshold, opt tags in.
At a glance

One layer, two audiences

Every conversation rolls through the same understanding pass. The output then splits two ways — surfaced in the panel for human agents, and routed into the system in the background. As Sidekick grows, the new modules and workers slot into one of those two branches.

Sidekick
What it understands three reads, every message
Topic
Conversation summary
Entities
Order IDs Serial numbers Emails Custom values
Patterns
Incident match Duplicate threads Tag candidates
The same read powers everything that follows.
For humans the panel
Modules
Summary Suggested response Handbook Knowledge Links
Hint cards
Incident match Duplicate convos
Drag-orderable. The set is alive — new modules slot in here.
For the system background workers
Workers
Detection rules Auto-tagging
Side effects
Custom field fills Inline highlights Tags applied Incidents linked
Runs whether or not a human is reading. Same read, different audience.