You started using AI to draft replies, but every message still runs through you first: read, edit, approve, send, one at a time. That works until the queue backs up. Then you either fall behind or start hitting send without really checking. An escalation rule fixes that gap.
An escalation rule is a written boundary that tells an AI agent when it can send a message on its own and when it has to stop and wait for a person to check first. It is not a settings toggle you flip once. It is a short, specific list you write down, hand to the agent, and revisit as you learn what actually goes wrong.
What happens when AI sends without a human checking first?
Nothing happens most of the time, which is exactly the problem. A routine reply about business hours or a shipping estimate goes out fine ninety-some times out of a hundred. The failures are rare but expensive: a refund promised that finance never approved, a reply that agrees to a discount nobody authorized, a curt response to a customer who was already furious.
The cost of an unreviewed mistake is not evenly spread. A wrong answer to "what are your hours" barely registers. A wrong answer to "I want my money back" can end up as a chargeback, a public review, or a canceled account. Escalation rules exist to route by cost, not by volume.
Picture a support inbox handling forty messages a day. Maybe two of those forty carry real financial or reputational risk. Reviewing all forty by hand wastes time on the thirty-eight that were never going to go wrong. Reviewing none of them means the two risky ones slip through unchecked. An escalation rule is how the review time drops to the two that actually matter.
How do you decide what an AI agent can send on its own?
Start from the downside, not the upside. For each type of message your business sends, ask what happens if the AI gets it wrong and nobody catches it before the customer reads it. That is a different question than "how good is the draft," because a confident, well-written draft can still be the wrong thing to send.
- List the message types your agent handles today: order status, scheduling, general questions, complaints, quotes.
- For each type, write down the worst realistic outcome if the draft is wrong and goes out unreviewed.
- If the worst outcome costs you money, a customer, or a legal exposure, it escalates.
- If the worst outcome is a follow-up email to correct a small error, it can auto-send.
- Set the rule in writing, not in your head, so anyone running the agent applies it the same way.
Which messages should always wait for a human?
A short list, held to consistently, works better than a long list nobody remembers. These categories should escalate every time, regardless of how confident the draft looks:
- Any message that mentions a refund, credit, or chargeback.
- Any message where the customer references a legal complaint, a regulator, or threatens to sue.
- Any custom price, discount, or contract term that is not on your published price list.
- Any message where the customer's tone reads as angry, or where the same person has written in more than twice.
- Anything about a data, privacy, or security concern.
- Any first-time reply to a customer flagged internally as high-value or high-risk.
Six categories is a workable starting size. Fewer than that and you will miss a real risk; more than that and whoever is reviewing the queue starts skimming past the list instead of checking against it. Write the categories down somewhere the whole team can see, not just in the AI agent's settings, so a new hire covering the queue on a Friday afternoon applies the same rule you would.
Auto-send vs. always-review: a working split
Most small businesses land on a version of this table once they have run an agent for a few weeks:
| Message type | Default rule | Why |
|---|---|---|
| Order status, hours, general FAQ | Auto-send | Low stakes, easy to verify, high volume |
| Scheduling confirmations | Auto-send | Reversible, no money changes hands |
| Refund or credit requests | Always review | Direct financial exposure |
| Custom quotes or discounts | Always review | Sets a price precedent you may not want |
| Angry or repeat complaints | Always review | Wrong tone compounds the problem |
How do you review a queue of AI drafts fast?
Reviewing fast is not the same as reading every word. Most escalated drafts only need you to check three things, in order.
- Read the customer's original message first, not the draft, so you know what they actually asked.
- Scan the draft for a number: a price, a date, a refund amount. Verify that number specifically.
- Check the tone against the customer's tone. A flat reply to an angry customer is the most common failure.
- Approve, edit, or reject. Do not rewrite from scratch unless the draft got the facts wrong.
A queue built this way runs in a few minutes for ten drafts, because you are checking specific risk points, not proofreading.
How often should you revisit the rule list?
An escalation rule is not a one-time setup. It should change on a schedule, not just when something goes wrong.
Check it monthly for the first quarter, then quarterly after that. Two signals mean it needs an update sooner: the review queue is empty most days, which usually means the categories are too broad and safe messages are escalating for no reason, or something got sent that should have stopped, which means a category is missing. Either signal is worth acting on the same week you notice it, not at the next scheduled review.
Where a managed agent changes the math
Writing your own escalation rules works, but someone still has to build the routing, hold the queue, and enforce the rule when volume spikes. PropelClick's customer support agent, ReplyBot, ships with an approval queue built into the setup: routine replies go out, flagged categories stop for you, and the rule list lives in the agent's configuration instead of a shared doc nobody opens.
ReplyBot's setup runs $197 with a $97 monthly fee, and every draft that escalates lands in one queue instead of scattered across inboxes and text threads.
What if the rule slows things down or you get it wrong at first?
Your first rule list will be wrong in some places, and that is normal. You will either escalate too much, which slows you down, or too little, which is how the expensive mistakes happen. Expect to adjust the categories after the first month once you see which drafts actually needed your eyes.
The 30-day guarantee on PropelClick's agents refunds the setup fee if the agent is not doing the job by then. It is not a blanket money-back promise on the monthly fee, and there is no annual contract to walk away from either way, so getting the first version of your escalation rule wrong is not an expensive mistake to fix.
Set your first escalation rule this week
Pick one message type you currently review every single time and ask honestly whether the worst-case outcome justifies it. If it does not, let it auto-send and watch the queue for two weeks. Start with a free AI readiness assessment to see where your business's current workflow already has an escalation point you have not written down yet.
Frequently asked questions
What is an AI escalation rule?
An AI escalation rule is a written boundary that tells an AI agent which messages it can send on its own and which ones must be routed to a person first. It is usually organized by message type, such as refunds, complaints, or custom pricing, rather than by customer or channel.
Should AI ever send a customer message without a human reviewing it first?
Yes, for low-risk, routine messages like order status or scheduling confirmations, where the worst-case outcome of an error is minor and reversible. Anything involving money outside a published price, a legal mention, or an angry customer should escalate to a human every time.
How long should it take to review an escalated AI draft?
A well-scoped review takes under a minute per draft: read the customer's original message, verify any number in the draft, and check the tone. Reviews take longer only when the reviewer rereads the whole thread from scratch instead of checking the specific risk points.
What kinds of messages should never be sent automatically?
Refund and credit requests, anything referencing a legal complaint or regulator, custom prices or discounts outside your published list, and replies to customers who are angry or have written in more than once about the same issue should always wait for a human.
Does a managed AI agent come with an approval queue built in?
PropelClick's ReplyBot includes an approval queue as part of its setup: routine replies send automatically, and messages that match your escalation categories are held in one queue for review rather than scattered across inboxes, texts, and social messages where they are easy to miss.
What happens if my first escalation rule is set wrong?
It is normal for the first version to escalate too much or too little. Most businesses adjust the categories after the first month once they can see which drafts actually needed a human. PropelClick's 30-day guarantee refunds the setup fee if the agent still is not working by then.