Vardast

When AI Should Hand Off to a Human: The Escalation Rules That Save Sales

Category: DM Automation
When AI Should Hand Off to a Human: The Escalation Rules That Save Sales

Most AI assistants don't fail by answering wrong. They fail by never letting go. Here are the six triggers that should always reach a person, and how to hand a conversation over without losing the sale.

It is 11:40 at night. A customer has asked the same question three times, phrased three different ways, and received three polite, correct, completely useless answers. Then she types: "is there an actual person there or not?"

That message is the sound of a sale dying. Not because the assistant was wrong. It wasn't. Because nobody built the exit door.

I keep seeing sellers treat the human handoff as an admission of defeat. Their assistant handles 90% of DMs, so they push it to handle 100%, and the last 10% turns into a room the customer can't get out of. The opposite instinct is just as expensive: a bot that panics at the first hard question, so the seller ends up reading every thread anyway and wonders what they're paying for.

The setups that work do something less dramatic. They decide in advance which conversations belong to a person, and they make the switch feel like an upgrade instead of an apology.

Escalation is a design decision, not an error message

The metric most sellers quote me is "what percentage does the AI handle by itself?" It's the wrong number to be proud of. A conversation that a human closed in two minutes with a photo and a discount is a win for the assistant too, because the assistant did the sorting, the waiting, and the first four messages of context at 2am.

Your assistant's job is the boring, high-volume 80%: price, sizing, stock, shipping time, "is this available in blue". Its second job, the one almost nobody configures, is to recognise the moment a conversation stops being that, and to move it cleanly. Speed is what automation is genuinely good at, and the gap between a nine-second reply and a sixty-minute one is most of the value. Judgment on a refund is not.

The trigger list: six conversations that should never stay with the bot

Write these down and configure them literally. A vague instruction like "escalate if the customer seems unhappy" produces a bot that escalates everything or nothing.

  1. Anger, or anything with emotional weight. Caps lock, a second complaint about the same order, the word "disappointed", a threat to post about it publicly. An automated apology on top of a real grievance makes it worse, every time.
  2. Money that leaves the rulebook. Refunds, damaged goods, a discount bigger than the one you authorised, custom or wholesale pricing, split payments. Anything where the answer costs you money should be spoken by someone who is allowed to spend it.
  3. Repeated confusion. If the customer asks the same thing twice, the second answer will not land either. Two failed attempts on one intent is a hard stop, not a suggestion.
  4. A question your knowledge base has no honest answer to. Collaborations, press, "do you ship to Germany", "can you make it in size 45". A confident guess here is worse than silence, because now you have to un-promise it.
  5. An explicit request. "Can I talk to someone." No negotiation, no "I can help with that too!" That line is where trust goes to die.
  6. Orders far above your average basket. If your typical DM order is one item and someone asks about twelve, that thread is worth a human's twenty minutes. Set a value threshold and let it interrupt you.

The trigger sellers forget is the second one on that list: small refunds. They feel routine, so they get automated, and then a customer with a broken jar of face cream gets a cheerful policy recital. That's a screenshot on someone's story by lunchtime.

How the handoff should read to the customer

Most handoffs fail in one sentence. Here's the sentence: "I'm an AI assistant and I'm unable to help with that. Please contact support."

Three problems. It announces a limitation, it hands over nothing, and it asks the customer to do the work of finding a human. Compare it with this:

"Got it, a refund on the damaged jar. I'm passing this to Sara from our team with your order number and the photo you sent. She'll message you here before 11am tomorrow."

The rules underneath that:

  • Name what's happening next, not what the bot can't do. Nobody needs a status report on your automation.
  • Give a time, and make it one you'll hit. "Shortly" and "as soon as possible" are the two least trustworthy phrases in customer support. "Before 11am" is a promise you can keep and be judged on.
  • Never ask them to repeat themselves. If the human's first message is "Hi! How can I help?", the handoff failed even though the routing worked.
  • Stay on the same channel. Sending an Instagram customer to an email address is how you turn a 15% buyer into a 0% one.

What the human needs handed over with it

A handoff is a package, not a doorbell. When the thread lands in front of a person, these should already be there:

  • A two-line summary of what the customer actually wants. The want, not the transcript.
  • Anything the assistant already promised. Prices quoted, stock confirmed, delivery windows mentioned. If your team contradicts the bot in the first reply, you've just taught the customer that neither of you knows anything.
  • The order or product, identified. Not "she's asking about a bag" but the SKU, the colour, the date.
  • Why it escalated, in one word: refund, angry, unknown, wholesale.
  • The full thread, one click away, in the same place you answer everything else, which is the whole argument for keeping every channel in one inbox with one memory.

This is also the part that separates a team that trusts its assistant from one that switches it off after a bad week. Handoffs that arrive without context train your staff to distrust the whole system, and that's usually how sellers end up turning the AI off: not from one dramatic failure, but from twenty small ones nobody logged.

When nobody is awake

Half of these triggers will fire at midnight. That's fine, as long as the assistant stops pretending.

Off-hours escalation has one job: hold the customer without lying to them. Say a person will reply in the morning, say roughly when, and use the remaining minutes to collect what your morning self will need. The order number, the photo, the phone number. A queued conversation with everything attached gets solved in one message at 9am. A queued conversation with "a customer needs help" gets solved in six, spread over two hours.

Handing it back

The messy half nobody plans. Once a human takes a thread, the assistant must go quiet, completely, until someone releases it. A bot that jumps back in mid-negotiation to recite the shipping policy is worse than no bot.

Give the assistant a short cooling period after the human's last message before it resumes routine questions, and make resuming an explicit action. In practice: a person handles the refund, marks the thread done, and the assistant picks up the next unrelated question two days later like nothing happened.

The number to watch

Track your escalation rate weekly, as a percentage of conversations, with the reason attached. Most seller accounts I've seen settle somewhere between 8% and 15%, and where you sit matters less than the direction you're moving.

Climbing past 25% usually means a knowledge gap, not a broken bot. Read the reasons and you'll find the same three unanswered questions over and over. Under 3% is not a triumph either; it usually means the exit door is closed and people are giving up instead of escalating. Either way, every escalation reason is tomorrow's knowledge base entry. That's the loop: the bot escalates, you answer, you write the answer down, the bot stops escalating that one.

If you'd rather not build this rule by rule, Vardast ships with the escalation triggers, the handoff message and the context package already wired up across Instagram, WhatsApp, Telegram and your website. Your assistant answers the routine 80%, and the conversations that need you arrive with the order, the history and the reason already attached.

Build the exit door before you need it. The customer typing "is there an actual person there" is telling you exactly what it cost to skip.

FAQs about handing conversations from AI to a human

What percentage of conversations should an AI assistant escalate?

Most seller accounts settle between 8% and 15% of conversations. Consistently above 25% usually points to gaps in your product knowledge rather than a broken assistant. Under 3% often means the exit door is closed and customers are quietly giving up instead of asking for a person.

Should the AI tell customers it's an AI when it hands off?

Don't build the handoff around that announcement. Say what happens next and when, and who is taking over. Being upfront that customers are talking to an assistant is fine and often required by platform rules, but it belongs at the start of the conversation, not as an excuse at the moment things get hard.

What should happen if a customer asks for a human outside working hours?

Hold them honestly. Tell them a person replies in the morning and give a rough time, then use the remaining messages to collect the order number, photos and phone number your team will need. A queued conversation that arrives complete gets closed with one reply the next day.

How do I stop the AI from jumping back into a conversation a human is handling?

Make the takeover a state, not a suggestion. Once a person replies, the assistant should stay silent until the thread is explicitly marked resolved, plus a short cooling period after the last human message. Anything less produces a bot that interrupts a live negotiation with the shipping policy.

Which conversations are most expensive to automate?

Refunds, damaged goods and anything with anger in it. They look routine, so they get automated first, and a cheerful policy recital on top of a real grievance is what turns one unhappy customer into a public post. Custom pricing and bulk orders belong to a human too, because the answer costs you money.

Does escalating a lot mean the AI assistant isn't working?

No, escalating badly does. An assistant that sorts, answers the routine 80% and hands over the rest with full context is doing its job. Track the reasons weekly and turn each one into a knowledge base entry; the escalation rate drops on its own as the answers get written down.