AI agents
An AI agent reads new messages in the inboxes you choose and answers them in the voice you describe. You decide whether it sends on its own, suggests a reply for you to send, or only writes drafts.
How it works
Section titled “How it works”When a customer writes, Pinlyx finds the agent that covers that inbox. It waits until the customer stops sending for a few seconds, so a burst of short messages gets one answer instead of several. Then it checks, in this order:
- Is the AI paused for this contact? Then it stays silent.
- Does the message fit the agent’s trigger (every message, first contact, keyword or rules)? If not, it skips the message.
- Does the message contain a stop keyword, is it quiet hours, or has the agent hit its hourly or daily cap? Then it skips the message.
- Does the message contain a handoff keyword, or does the customer ask for a person? Then it hands the chat to your team.
- Does one of the agent’s rules match? Then the rule decides.
- Otherwise the agent writes a reply from its persona, its knowledge base and what it has learned. If you allow it, it can still hand off when it is unsure or the customer sounds upset.
Every outcome is recorded on the agent’s Activity tab. The number of agents you can create depends on your plan, see Plans and limits.
Before you start
Section titled “Before you start”- Add your AI provider key under Settings › AI assistant. See AI usage and keys.
- Connect at least one channel. See Channels.
Create an agent
Section titled “Create an agent”-
Open AI agents and click New agent.
-
On Start, choose how to begin:
- Learn from your chats studies conversations on your connected social accounts and writes a persona and knowledge notes in your voice. Pick the accounts, choose an Analysis depth (Standard, Deep or Everything), click Analyze my chats, then Use this voice. Nothing is sent during the analysis.
- Start from a template: Blank, Sales assistant, Customer support or Lead qualifier.
-
On Basics, enter a Name, an optional Description (only your team sees it) and the Persona. The persona is the brief the agent follows in every reply: who it speaks for, how it sounds, what it must never say and when to hand off.
-
On Channels, turn on the channels the agent may answer in. Under Account scope, pick specific inboxes, or pick none to cover every account in that channel, including ones you add later.
-
On Knowledge (optional), add the facts it should answer from. See Knowledge base.
-
On Behaviour, set When should the agent reply? and How should it reply? (see the tables below).
-
On Safety (optional), review the pre-filled values: Handoff keywords, Stop keywords, Max replies / hour, Max replies / day, Reply delay (sec) and quiet hours.
-
Click Create agent. The agent is saved as a Draft and opens on its own page.
When it answers
Section titled “When it answers”| Option | What it does |
|---|---|
| Every incoming message | Considers every message on its channels |
| First contact only | Replies only to the first message from a new contact |
| Only when keyword matches | Replies only when one of your keywords appears |
| Use rules | Only your rules reply, the AI itself stays off |
How replies go out
Section titled “How replies go out”| Mode | What happens |
|---|---|
| Auto-send | The agent sends the reply on its own |
| Suggest | The reply is drafted in the conversation and you send it |
| Draft only | The reply is saved silently for review, with no inbox notification |
New agents start in Suggest. Keep them there until the drafts read the way you would write.
Channels it can answer
Section titled “Channels it can answer”| Channel | What it covers |
|---|---|
| Telegram | Direct messages to your Telegram accounts |
| Social inbox | Instagram, Facebook and WhatsApp messages |
| Live chat | Chats from the widget on your website |
| X | Direct messages on X. Replies arrive as suggestions you send yourself |
| Emails to your connected mailboxes, answered through an inbox rule |
For email, the agent does not watch the mailbox by itself. Create an inbox rule with the Auto-reply action, set Reply with to AI agent and pick the agent.
For live chat, the widget’s AI replies settings choose which agent answers. Auto (highest priority) lets your top live chat agent reply. See Live chat.
The Channel routing tab on the AI agents page shows which agent answers each connected inbox. An inbox with no agent gets no AI reply at all. When several agents cover the same inbox, the first one that matches replies.
The agent page
Section titled “The agent page”| Tab | What you do there |
|---|---|
| Overview | Its setup in plain words, replies in the last 24 hours, handoff rate and latest replies |
| Identity & tone | Persona, when it answers, how replies go out, and Model and creativity |
| Knowledge | The entries it answers from |
| Channels | Channels and account scope |
| Handoff & rules | Handoff conditions, keywords, pace, quiet hours and rules |
| Alerts | Situations to watch for and who to notify |
| Learning | What it learned from replies you type by hand |
| Activity | Every run, with approve and discard for pending replies |
| Test | A playground that never reaches real contacts |
Model and creativity lets one agent use a different provider or model than your default.
Pace and quiet hours
Section titled “Pace and quiet hours”- Max / hour and Max / day cap how many replies the agent sends or suggests. When it reaches a cap, it skips new messages until the window moves on.
- Reply delay (s) holds an automatic Telegram reply for that many seconds before it is sent, so answers do not look instant. On the other channels a reply goes out as soon as it is ready.
- Quiet hours stop replies between two times. They use UTC. Leave both empty to answer around the clock.
Rules run before the AI, top to bottom, and the first one that matches decides. Each rule has a Condition and a Then action.
| Condition | Action |
|---|---|
| Message contains keyword | Send a fixed reply |
| Message matches regex | Move contact to stage |
| Contact pipeline stage is | Add a tag |
| Channel is | Hand off to human |
| Block / mute auto-reply |
Move contact to stage and Add a tag apply and let the agent continue. Send a fixed reply, Hand off to human and Block / mute auto-reply end the run without asking the AI.
Test and tweak
Section titled “Test and tweak”-
Open the agent and go to Test.
-
Type a sample customer message and click Send. The agent answers with its persona, knowledge and rules. Each reply shows the tokens it used.
-
Click Good reply to save it as an example the agent should match.
-
Click Tweak when something is wrong. Select or type The wrong part, explain What’s wrong, and why?, optionally write What should it have said instead?, and click Save to memory.
-
Click Try again to see the difference.
Saved lessons appear under Agent memory next to the playground. The most relevant ones shape every new reply, in tests and in live chats. Click Forget this lesson to remove one.
You can teach the agent from live chats too. Under an AI reply in a conversation, click Good reply, or Was this reply wrong? to describe the mistake and the correct answer.
Learning from your own replies
Section titled “Learning from your own replies”On the Business plan, the Learning tab studies every reply you type by hand in the agent’s inboxes. Wording and approach are picked up automatically, and facts wait for your approval. See Knowledge base.
Go live
Section titled “Go live”-
Run a few realistic messages on Test, including a refund request and an angry customer.
-
Check Handoff & rules: handoff keywords, stop keywords and the switches for when to hand off. See Human handoff.
-
Click Go live. The status changes to Live.
-
Watch the Activity tab. With Suggest, approve good replies with Approve & send and remove the rest with Discard.
-
Switch How replies go out to Auto-send only when the suggestions no longer need edits.
Click Pause at any time to stop the agent everywhere.
Good to know
Section titled “Good to know”- On X, agents do not send on their own. Use Suggest and send the reply yourself.
- The Activity tab explains silence. A Skipped run shows the reason, for example quiet hours, a stop keyword or the reply cap.
- Deleting an agent cannot be undone. Pause it instead if you may need it again.
- Replies need a working key. If your provider key is missing or rejected, runs show as Failed.