Learn how to write an escalation playbook for chat support: triggers, tiers, ownership, handoff messages and review habits, plus how to set it up.
Every support team has a moment when a chat goes sideways: an angry customer, a refund beyond the bot's remit, a question nobody has answered before. If the next step lives only in one manager's head, the customer waits and your team improvises. This guide shows how to write an escalation playbook that is short enough to follow under pressure, and how to turn it into working rules for website chat, WhatsApp and Messenger.
What an escalation playbook is (and is not)
An escalation playbook is a written answer to one question: when a conversation needs more than the first responder can give, who takes it, how fast, and with what information?
It is not a policy manual or a list of every possible complaint. It is a decision aid. A good one fits on two pages and can be used by a new hire on day one.
A useful playbook has five parts:
- Triggers: what causes an escalation.
- Tiers: who receives it at each level.
- Ownership and timing: who is on duty and what the customer is told.
- Handoff package: what context travels with the conversation.
- Review loop: how you learn from escalations and shrink the list.
Step 1: Define your triggers
Triggers should be observable. "The customer seems upset" is a feeling; "the customer asks for a human twice" is a signal anyone can check.
Group triggers into four families:
Customer request
- The customer asks for a person, a manager or a complaint process.
- The customer repeats the same question after two answers.
Risk and money
- Refunds, chargebacks, cancellations or price exceptions.
- Anything involving legal threats, safety, health or personal data requests.
Knowledge gaps
- The agent has no answer in its knowledge base.
- The customer disputes an answer and cites a document you cannot verify.
Emotion and tone
- Strong negative language, shouting, repeated complaints.
- A VIP or high-value account, if you track that.
Step 2: Design the tiers
Most small and mid-sized teams need only three tiers. More than that and people argue about who owns what.
| Tier | Who | Typical scope |
|---|---|---|
| 0 | AI agent | FAQs, order and policy questions, lead capture |
| 1 | Frontline human | Takeovers, tone-sensitive cases, simple exceptions |
| 2 | Specialist or manager | Refunds above a limit, technical faults, complaints, legal or privacy matters |
For each tier, write down three things: what they can decide alone, what they must pass up, and the maximum time before the case moves on.
- 01Customer message
- 02AI agent answers
- 03Trigger fires, Tier 1 takes over
- 04Tier 2 for exceptions
Write decision limits as numbers you own
Rather than "small refunds," write your own limit, such as "Tier 1 may approve refunds up to the amount in our refund policy; anything above goes to Tier 2." Use whatever figure your business has set. The point is that nobody has to guess.
Step 3: Assign ownership and timing
Escalations fail most often at the gap between "flagged" and "picked up." Two things close that gap: a rota and a response target.
- Rota: list who is on duty by day and timezone. If you serve customers across US, UK and Asia-Pacific hours, a follow-the-sun rota avoids a single overloaded person.
- Out-of-hours rule: decide what happens when nobody is on shift. Usually the answer is "tell the customer honestly when a person will reply, capture contact details, and queue the case for the next shift."
- Response targets: set a target per tier, but only publish to customers what you can keep.
Never promise a wait time the rota cannot support. A customer told "within 5 minutes" at 3 a.m. will remember the broken promise longer than a slower, honest answer.
Step 4: Specify the handoff package
A human who inherits a chat should never ask "How can I help?" again. Define the minimum context every handoff carries:
- The full conversation so far
- The trigger that fired
- The customer's name and contact details, if given
- What the AI already tried or told them
- Any order, ticket or account reference
Then write the customer-facing line. It should say three things: you are connecting them to a person, what happens next, and what they do not need to repeat. For example:
Channel notes
- Website chat: the customer is watching the window, so a short holding message matters most.
- WhatsApp: customers may return hours later, so state when to expect a reply. Remember that WhatsApp business messaging has its own rules about replying within a time window after the customer's last message.
- Messenger: similar expectations to WhatsApp; keep replies brief and confirm next steps.
Step 5: Build the alert path
A handoff nobody sees is not a handoff. For each tier, decide where the alert lands: a team chat channel, email, or a webhook into your ticketing tool. Choose the place your team actually watches during the day, not the place that looks tidiest in a diagram.
Add a backup. If an alert is not acknowledged within your target, who gets nudged? Put the name or role in the playbook.
Step 6: Review and shrink
Escalation is a cost. Your goal is not zero escalations but the right ones.
Once a month, sample a batch of escalated chats and sort them:
- Should have been answered by the AI: add the missing knowledge or fix the answer.
- Correctly escalated: leave it alone.
- Escalated too late: tighten the trigger.
- Escalated needlessly: loosen it.
In this made-up sample, 14 of 40 cases point to content you can add, which is the cheapest improvement available. Repeat each month and the fixable share should fall.
Doing this in AI SNS CHAT
AI SNS CHAT supports most of this playbook directly. Here is how the pieces map.
Triggers and Actions. Escalations is one of the built-in Actions under the integrations area. Integrations need the Hobby plan or higher, and the number of Actions per agent depends on your plan (0, 5, 8, 12 or 20 from Free upward). Persona and guardrails on the agent let you write in plain language when it should hand over, such as refund requests or repeated questions.
Rota and routing. Under Helpdesk → Scheduling you set shifts per person and timezone, and handoffs are auto-assigned to whoever is on shift. Helpdesk needs Standard or higher. When a handoff happens, the server reads the rota and adds the "what happens next" line to the customer message itself. It never invents a wait time, which matches the honesty rule above.
Alerts. Handoff alerts can go to Slack, Microsoft Teams, LINE WORKS, WhatsApp, email or a webhook. Pick the one your Tier 1 and Tier 2 staff already watch.
Taking over. In Activity → Chat logs (all channels except email) and in Inbox (email only), you can take over a chat, and the bot pauses while you reply. Live translation helps when a customer writes in another language, and saved replies speed up common answers.
Learning loop. You can edit a reply into the ideal answer to train the agent, which turns each escalation into a fix. In Chat logs you can bulk select, flag, escalate, resolve, filter and export conversations, and thumbs up/down ratings help you find weak answers for your monthly review.
Accountability. The workspace activity log (Workspace settings → Activity, for owners and admins) records creates, edits and deletes with name, email, feature and time, which is useful when you change triggers and want to know who changed what.
Rehearse first. Use the Playground to test the escalation wording before it reaches customers. If you are unsure where a setting lives, Synthia, the AI fleet manager, can navigate to the screen and highlight the control. Plans and limits are on pricing, and the full list is on features.
A one-page template to copy
Use this outline as your first draft:
- Purpose: one sentence.
- Triggers: five to eight, each observable.
- Tiers: who, what they can decide, time limit.
- Rota: who is on duty, out-of-hours rule.
- Handoff package: the context list and the customer message.
- Alert path: where alerts go, backup contact.
- Review: monthly sample, owner, what changes.
Conclusion
A playbook works when it is short, specific and owned by a named person. Start with a few triggers, three tiers and an honest handoff message, then let real conversations tell you what to add. Revisit it monthly, and your team will spend less time deciding who should act and more time actually helping.
Frequently asked questions
How many escalation triggers should I start with?
Five to eight is plenty. Choose ones that are observable, such as a customer asking for a person twice or a refund request. Add more only when your monthly review shows a real gap.
Should the AI agent ever refuse to escalate?
Only if your policy says so for narrow cases, and even then it should offer another route, such as taking contact details. When a customer asks for a human, make the path to one clear rather than forcing them to argue with the bot.
What should the customer see when nobody is on shift?
An honest message saying a person will follow up, plus a way to leave contact details. In AI SNS CHAT the handoff line is built from the rota and does not invent a wait time, so it stays truthful outside hours.
Do I need Helpdesk Scheduling to run a playbook?
No, a small team can start with alerts sent to a shared channel and a written rota. Scheduling, which needs Standard or higher, becomes valuable when you have several people across timezones and want handoffs assigned automatically.


