Every Aiva Chat deployment has one question at its centre that's rarely asked out loud: when should a human take over?
Get it wrong in one direction and the agent routes nearly everything to a person, so you've added a step to every call without removing any work. Get it wrong in the other direction and the agent handles things it shouldn't, leaving customers who feel they were denied a person exactly when they needed one.
AI escalation rules are the settings that decide which of those happens. In our experience they matter more to a deployment's success than voice quality, speed, or the number of languages supported, because they decide how the hardest conversations end.
Our position, argued below, is that an agent should hand off based on what the caller needs from a person, and should treat its own uncertainty as a gap to fill. If you're evaluating vendors rather than configuring one, the buyer's guide to AI-to-human handoff is shorter and less technical.
What are AI escalation rules?
They're the conditions under which an AI agent stops handling a conversation itself and passes it to a person. In a phone context, that usually means a live transfer to a staff member or department, with a fallback if nobody answers.
Good escalation rules answer three questions:
- When should the conversation move to a person?
- Where should it go, meaning which department or number?
- What happens if that person isn't available?
In Aiva Chat, these live in a few places. The Handoff tab on each assistant has four built-in triggers. Transfer settings define departments, phone numbers, and the hours each number should ring. The Rules tab lets you write situation-specific instructions in plain words, such as "hand off any legal or fraud question to a person."
Escalation rules matter for every business using an AI receptionist, but they matter most where conversations carry emotional weight or legal consequences: clinics, law firms, financial services, property management, and any business that handles complaints by phone.
The principle we use throughout: escalate for who the customer needs, not for what they asked.
How AI-to-human handoff works in Aiva Chat
Step 1: A trigger fires
During a call, the caller asks for a person, sounds upset, raises billing or refunds, mentions an emergency, or says one of a department's trigger phrases. Any rule you've written for a specific situation can also apply.
Step 2: Aiva Chat picks the destination
If the caller names a department, such as "billing," Aiva Chat routes there. Each department has one or more numbers, and each number can be limited to certain days and hours, so after-hours transfers go to whoever is on call.
Step 3: The call is transferred
Aiva Chat connects the caller to the number. While a call is live, staff can open it on the Calls page to see the live transcript and what the AI has done so far, and can also take over or transfer by hand.
Step 4: The fallback runs if nobody answers
If the line is busy, unanswered, or fails, the assistant returns, offers to take a message, records it, and creates a follow-up task. If that isn't possible, the caller can leave a voicemail.
Step 5: The record is kept
The call log holds the recording, transcript, summary, and actions taken. With HubSpot connected, the conversation is logged on the contact, so whoever calls back has the full story.
Benefits of getting escalation rules right
Customers who need a person get one quickly
An upset or frightened caller needs someone who can acknowledge what's happened, more than they need an answer. Fast, reliable handoff protects the relationship at its most fragile moment.
Staff get calls that actually need them
When routine and moderately complex calls stay with the AI, staff receive fewer, more meaningful transfers. They stop feeling that every transfer is pointless.
You stay on the right side of automated-decision rules
In the EU, GDPR Article 22 gives people the right not to be subject to decisions based solely on automated processing that significantly affect them, and in some cases the right to obtain human intervention. An agent that resists handing over is building toward a compliance problem.
The AI's value isn't wasted
Escalating on uncertainty throws away what the agent is good at: managing back-and-forth, gathering context, and completing actions. Escalating on need keeps that value.
You set the built-in triggers per assistant on its Handoff tab, along with the departments and numbers calls transfer to. Our advice is to keep "Asks for a person," "Sounds upset," and "Mentions an emergency" switched on. Getting those wrong has consequences that outweigh any efficiency gain.
When to escalate: the scenarios
Three calls that should reach a person and one that shouldn't (all invented for illustration):
Emotional distress
Problem: A caller says "this is the third time I've called and nobody has fixed it." How Aiva Chat handles it: With "Sounds upset" ticked, Aiva Chat stops trying to solve the issue and transfers to the right person, even if it could technically answer the question. Outcome: The customer reaches someone who can acknowledge the frustration. More on this in how AI handles angry customers.
Billing, legal, and fraud
Problem: A caller disputes a charge and mentions their lawyer. How Aiva Chat handles it: "Asks about billing or refunds" routes the call to billing. A rule on the Rules tab sends legal and fraud questions to a named department. Outcome: Decisions with financial or legal weight are made by people who are accountable for them.
An explicit request
Problem: A caller says "just put me through to someone." How Aiva Chat handles it: With "Asks for a person" ticked, it confirms and transfers. No deflection loop. Outcome: The caller trusts the business more, not less.
The complex but routine call
Problem: A caller wants to reschedule, check directions, and confirm the fee, all in one call. How Aiva Chat handles it: It works through each question in turn. Nothing here needs a person. Outcome: Staff never see the call, and the customer gets everything done at once.
Escalation approaches compared
| Escalate on uncertainty | Escalate on need | Never escalate | |
|---|---|---|---|
| Who reaches staff | Anyone with a non-routine question | Upset callers, high-stakes issues, explicit requests | Nobody |
| Staff workload | High, mostly unnecessary | Lower, mostly meaningful | None, but complaints rise |
| Customer experience | Frequent transfers, repeated stories | Fast help when it matters | Trapped in a loop |
| Compliance exposure | Low | Low | High |
| AI value captured | Minimal | Most | Looks high, hides failures |
| Fix for knowledge gaps | Transfers forever | Fill the gap in information | Wrong answers persist |
Escalating on uncertainty feels safe, and for the first few days of a deployment it can be a reasonable stance while you learn what callers ask. As a permanent setting, it turns the AI into a switchboard that adds a step to every call.
Never escalating looks efficient on a dashboard because handoff rate is zero. The cost shows up elsewhere: repeat calls, angry reviews, and customers asking for a person more than once in a single call.
Escalating on need sits between the two. It hands off the conversations where a person adds something the AI can't, and fixes knowledge gaps by improving the information instead of by transferring.
How to set up escalation rules in Aiva Chat
Step 1: List who callers might need
Write down the situations where a person is genuinely required: complaints, refunds, legal issues, emergencies, and explicit requests. Name the role that should take each one.
Step 2: Create departments and numbers
In the assistant's transfer settings, add departments with names callers would actually say, such as Billing or Support. Add numbers for each, and limit numbers to the hours they're staffed. The guide to business hours and call routing walks through the screens.
Step 3: Tick the built-in triggers
On the Handoff tab, tick "Asks for a person," "Sounds upset," "Asks about billing or refunds," and "Mentions an emergency" as they apply. Add trigger phrases to departments for words like "refund" or "leak."
Step 4: Write situation rules
On the Rules tab, add plain-language rules for anything the triggers don't cover, such as legal questions or requests for a specific manager.
Step 5: Test every path
Call as an upset customer, a billing dispute, an emergency, and someone asking for a person. Then test with nobody answering, to check that the fallback message and task work. Our guide to testing your assistant covers the test tools.
Step 6: Calibrate for two weeks
Review handoff transcripts daily, using the Needs review list and topics in analytics. Adjust triggers in both directions.
Common escalation mistakes
Escalating whenever the agent is unsure
This creates an expensive FAQ lookup. When the agent doesn't know something, add it to the information. Reserve transfers for situations where a person adds something the AI can't.
Turning on "Asks for a person" with nobody to answer
A transfer that rings out is worse than no transfer. Aiva Chat warns you if no transfer number exists, but you also need numbers that are actually staffed during their set hours.
Making callers ask twice
If transcripts show callers asking for a person more than once, the agent is negotiating when it should be transferring. That number should be close to zero.
Treating a low handoff rate as success
A near-zero handoff rate can mean distressed customers are being failed silently. Read it next to repeat calls and resolution, not on its own.
Never revisiting the rules
New services, seasons, and staff changes shift what callers need. Review escalation logs monthly once things settle.
The two failure signatures to watch for
Escalating too eagerly shows up as a high handoff rate concentrated in a few topics, with very short conversations before the transfer. Staff report that transfers feel pointless, which is the most reliable sign and the one nobody thinks to ask about.
Escalating too reluctantly is harder to see, because handoff rate looks good. The signals are repeat calls from the same number within a day, long conversations that don't resolve, and callers asking for a person more than once.
The first failure wastes money. The second loses customers quietly. If you're going to be wrong for a week, be wrong in the first direction.
Escalation rules best practices
- Escalate for who the customer needs. Not for how complex the request is.
- Keep the person, upset, and emergency triggers on. The downside of missing them is too high.
- Staff every transfer number during its hours. Use per-number hours to match reality.
- Fix knowledge gaps with information, not transfers. Use Fix this answer on wrong replies.
- Read handoff rate next to resolution rate. Either alone can mislead. Our KPI guide covers both.
- Test the no-answer fallback. Make sure messages and tasks are created.
- Review daily for two weeks, then monthly. Calibration is ongoing.
Conclusion
If you change one setting after reading this, make it the habit of escalating on need. An agent that transfers whenever it's unsure feels safe but leaves your staff doing the same work as before, plus a handoff step.
The decision that deserves the most thought is which direction to err in during the first weeks. Our advice is to err toward handing off too often, keep the person, upset, and emergency triggers on permanently, and then pull back topic by topic as transcripts show which transfers were unnecessary. Being too eager costs staff time; being too reluctant costs customers you won't hear from again.
Read your handoff rate alongside resolution rate, watch for callers asking twice, and revisit the rules whenever your business changes.
Ready to set handoff rules that hold up?
The best way to calibrate escalation is on your own calls. Set up an assistant, add your departments and transfer numbers, tick the triggers that fit, and review the first two weeks of handoffs.
Before that, call +1 (970) 972-0974 and say "just put me through to someone." How quickly it gives in is the behavior you're configuring. Transfers, live call monitoring, and the handoff analytics come with every plan.
Start a 14-day trial with no card and write your first list of handoff situations, or book a demo to pressure-test that list with us.
Posts from the people building Aiva Chat in Rajkot — how the product works, what we have learned building it, and guides for getting the most out of an AI receptionist.