Opening the form
- Create: From the AI Chatbots dashboard, click Create agent.
- Edit: Click Edit on any existing agent. The form loads with saved values pre-filled.
- Agent identity
- Knowledge base
- Files it can send
- Custom instructions
- Response settings
- Labels
- Notifications
- Follow-ups
- Review (readiness checklist — not a separate saved field)
Agent identity
Defines who the agent is for your team. Customers never see these fields.string
required
An internal label operators see in the inbox and on the assign menu (for example
Sales concierge). Max 60 characters.Customers never see this name.string
required
One sentence the agent works from on every reply — what the job is (pricing, booking, qualifying leads). Max 20,000 characters.Rules and exceptions belong in Custom instructions, not here.Good example:
Qualify inbound leads, answer pricing questions from our published tiers, and book a demo when the customer is ready.“Default” is per WhatsApp number, not a single flag on the agent itself. An agent becomes the org-wide default when every connected number points at it (dashboard Set as default / number assignment).
Knowledge base
Price lists, policies, and product sheets the agent answers from. Knowledge is optional for Save — an agent can still be saved with only Name, Mission, and Instructions — but without knowledge it answers only from the mission and custom instructions.Data sources (file uploads)
file[]
Upload documents Whatsable parses into text for the agent. Supported formats:
Limits: each file must be within the media size cap (typically 10 MB). On create, files stay staged until the first successful save, then upload.Statuses you may see: Processing, Ready, Failed, or Waiting (staged). Failed reasons are closed classes such as unsupported type, timeout, unreadable, or empty — never a raw parser dump.
Written knowledge
string
Short facts that belong in text rather than a file: hours, a delivery price, a promise you never make. Max 200,000 characters.Leave this empty if uploads cover everything. Empty knowledge does not block Save (it only affects the readiness score).
Files the agent can send
Distinct from the knowledge base. Knowledge is what the agent knows. Files are what it can send — the agent never reads the file bytes; it chooses a file only from the description you write. Cap: 10 sendable files per agent.string
required
Required to add a file. Max 1,000 characters. This sentence is the only thing the model uses to decide when to send the attachment.Example:
The 2026 price list, with delivery times for each region — send when a customer asks for pricing in writing.string
Optional display name (max 120 characters). Falls back to the original filename when empty. Editable later without re-uploading.
boolean
default:"true"
Per-file switch. Turn off to keep the file uploaded but stop the agent from sending it, without deleting the row.
Custom instructions
string
required
Rules for every reply this agent sends. The mission says what the job is; these lines say how to do it. Required to Save in the current editor. Max 80,000 characters.Write one rule per line. Typical uses:
- Topics the agent must never discuss
- Questions it must ask before booking or quoting
- Formatting preferences
- When to add or remove a label (name the label exactly as it appears under Labels)
Response settings
Controls how the agent communicates and when it returns after a human reply — not what facts it knows.How should the agent sound?
string
Free-text voice for ordinary replies (max 300 characters). Blank is allowed; when blank at runtime the platform falls back to a friendly default.Four presets fill the input (you can still edit the text afterward):
Follow-up voice is steered separately under Follow-ups → What to say in a follow-up.
How long should the agent wait before it sends a reply?
integer (seconds)
default:"5"
Debounce after the customer writes, before the agent’s answer goes out. Range 0–300 seconds (slider). Default 5.0 is legal — the turn can start immediately. A short wait feels more human and lets several quick messages batch into one reply.This wait does not apply to proactive follow-ups.
Once you reply by hand
enum
default:"never"
What happens after you or a teammate reply manually (including from the WhatsApp Business app). Options:
integer (minutes)
Shown only when the policy is The agent resumes after a set time. Range 1–10,080 minutes (1 minute to 7 days). Required when that policy is selected.The quiet period restarts if you write again. Replies from the Whatsable dashboard vs the mobile app may differ in how the timer is started — see Shared team inbox for operator behaviour in the live chat.
Labels
Lets the agent add or remove conversation labels you grant it. Powers default on.boolean
default:"true"
When on, the agent may tag conversations according to your label instructions and grants.
boolean
default:"true"
When on, the agent may remove granted labels when instructions say so.
integer | null
Shown when add is enabled. Range 1–5. Leave unset (Use the maximum) to use the platform max of 5.
integer | null
Same pattern as the add cap: 1–5, or unset for the platform max of 5.
enum
default:"all"
string
Policy text the model reads every turn (max 10,000 characters). Describe the moment in plain language.Example:
Add Hot lead when they ask to hold stock. Remove it once they pay.Notifications
When a customer wants a person, the agent steps back, the thread can be left unassigned, and a teammate can pick it up. Keywords catch explicit phrases; the situations box catches the rest. There is no separate notification toggle on this form (including no “Human Agent Request Notification” control). Configure keywords and situations here; team alerting follows product handoff behaviour.Which words mean they want a person?
string[]
Words or short phrases that, when present in a customer message, signal human assistance. Matching is case-insensitive.Limits: up to 25 keywords, each up to 120 characters.Common examples:
agent, human, speak to someone, real person, talk to a personAdd rows with Add keyword; remove with the control on each chip.Besides those words, when should it hand over?
string
Plain-language situational escalation — anger, a large refund, a safety issue — even when no keyword matched. Max 10,000 characters. This is policy for the model, not a WhatsApp message body.Example:
Hand over when the customer is frustrated or angry; when discussing refunds over $500; when a product safety issue is reported; when three exchanges have not resolved the issue.Leave blank if you only want keyword-based handoff.Keywords and situations work together. Keywords catch explicit asks; situations catch contextual ones.
Follow-ups
Outbound chase when a conversation goes quiet after your side (or the agent) sent a WhatsApp message. Cadence fields appear only while follow-ups are enabled; turning the master switch off hides them in the UI but stored values are kept for when you turn it back on.Master switch
boolean
default:"false"
Master switch. Off by default. When on, an unanswered outbound message can start a follow-up sequence inside the budget and window you set.
Cadence
Enter delays as days + hours + minutes (they add up to a single minute total stored by the API).integer (minutes)
default:"240"
Delay from the unanswered outbound before the first follow-up. Range 1–43,200 minutes (30 days). Default 4 hours. Must be strictly less than Stop following up after.
integer (minutes)
default:"7200"
Hard stop window from that same unanswered outbound. Range 2–129,600 minutes (90 days). Default 120 hours (5 days).Every follow-up in the stretch must go out before this deadline.
Some on-screen validation copy may still mention a 30-day window maximum. The platform contract allows up to 90 days.
integer
default:"3"
Budget for one quiet stretch, not a lifetime cap. Range 1–15. Default 3. If the customer replies and later goes quiet again, the budget can restart for the new stretch.
integer (minutes)
default:"0"
Minimum spacing between one follow-up and the next. Range 0–43,200 minutes. Default 0.
enum
default:"front_loaded"
A slot that would land in quiet hours or on a quiet day is skipped, not moved to the next open slot.
Quiet hours, quiet days, and timezone
integer pair (0–23) | empty
Quiet-hour start and end (hours 0–23). Leave both empty to allow any hour. Setting only one side clears both on save.
day toggles
Whole calendar days with no follow-ups (Sunday–Saturday), evaluated in the follow-up timezone.
IANA timezone
default:"UTC"
Timezone for quiet hours and quiet days (for example
Europe/Madrid). Curated list plus stored or browser zones. Default UTC.Follow-up content and risky toggles
string
Used only when composing a follow-up, not ordinary replies. Max 12,000 characters. Say what to offer, what never to promise, and not to repeat the last message.
boolean
default:"true"
On by default. The agent can still follow up when a customer goes quiet after your side replied — from the inbox, the WhatsApp phone app, or your own API integration.
boolean
default:"false"
Off by default.
boolean
default:"false"
Off by default. Outside the customer 24-hour service window, WhatsApp only allows an approved template. With this on, the agent may propose a new template for Meta review on your WhatsApp account — enabling it is that consent. Templates it wrote are marked in your template list.
integer
default:"2"
Shown only when template authoring is on. Range 0–20. Default 2.
Saving the configuration
Click Save when ready. The editor blocks save until:- Name is not empty
- Mission is not empty
- Instructions are not empty
Editing an agent does not rewrite conversations already in progress. Changes apply to the next message processed after save.
Related tools (not config fields)
- Improve my agent — helps organize mission, knowledge, instructions, handoff, and follow-up text before you apply them to the form.
- Test your agent — sandbox preview; nothing is sent to WhatsApp.
- Review — live readiness checklist before save.
Next steps
AI Chatbots Overview
Dashboard actions, plan requirements, and how an agent processes messages end-to-end.
Shared team inbox
How operators take over a chat, resume policies, and live inbox behaviour.
Embedding Process
Connect a WhatsApp number to your workspace before an agent can go live.