The short answer
An AI receptionist should transfer to a human when the call needs judgment, authority or accountability that a written rule cannot provide. In a service business that comes down to six categories: safety emergencies, angry or distressed customers, pricing exceptions, complaints about completed work, legal threats, and any caller who has asked for a person twice.
Everything else, which is most calls, should stay with the AI: answering, finding out what the caller needs, quoting from the price list, booking the job and confirming it.
As of October 2026, the quality of an AI receptionist setup is decided less by the voice and more by these rules. An AI that transfers too readily is an expensive phone menu. An AI that never transfers will, sooner or later, try to book a routine appointment for someone who is describing a gas smell or threatening to sue. Both failures are rule failures, and both are yours to prevent.
This guide explains how to write escalation rules. It covers the six categories, the difference between warm and cold transfers, what must happen when nobody picks up, after-hours rules, a decision table to adapt, and a weekly review that keeps the rules honest. The platform referenced is Run with Jarvis, with capability claims limited to its published plan list.
Why escalation rules are the owner's job
An AI receptionist follows instructions. It does not know that your biggest commercial client gets special treatment, that a certain kind of leak is always urgent in your trade, or that you would rather take a call at dinner than lose a customer of fifteen years. You know those things. They have to be written down.
The vendor cannot write them for you. A vendor can supply sensible defaults. The specifics depend on your trade, your team and your tolerance for interruption.
Your staff are already following unwritten versions. Ask your office manager which calls she walks straight to you. That list is the first draft of your rules.
Writing them improves the business with or without AI. A new human hire needs the same rules. So does the after-hours on-call technician. This is one of the first tasks covered in which jobs to delegate to an AI employee first: you hand over the routine work and keep the judgment, and the escalation rules are the line between them.
The test for any rule
Before the categories, here is a simple test for whether a situation belongs with a person.
Does it need a decision outside the price list and the schedule? Discounts, refunds, exceptions, promises about results.
Could a wrong answer cause harm? Physical danger, property damage, a legal or financial consequence.
Is the caller's emotional state the main issue? Fear, anger or distress that a scripted answer will make worse.
Has the caller asked for a person? At some point the request itself is the reason.
If the answer to any of these is yes, the call should reach a human or produce a callback from one. If all four are no, the AI should handle it.
The six categories that must reach a person
1. Safety emergencies
What it covers. Anything where someone could be hurt or property is being actively damaged. The specifics depend on your trade: a gas smell, sparking or burning from an electrical panel, a child or pet locked in a vehicle, flooding near electrical equipment, no heat in dangerous cold with vulnerable people in the home, a carbon monoxide alarm sounding.
What the AI should do. Recognize it from the caller's words, not wait for the word "emergency". Give only the safety direction you have approved in advance, such as contacting emergency services or the utility and leaving the building. Capture the address and a callback number immediately, because the call may drop. Alert the on-call person at once.
What the AI should not do. Diagnose. Reassure the caller that it is probably fine. Continue a normal booking script as though nothing unusual was said.
Writing the rule. List the trigger phrases and situations for your trade. Be generous. A false alarm that reaches a person costs a minute. A real one that does not costs far more. For general workplace and hazard information relevant to your trade, the Occupational Safety and Health Administration is the place to check, and your approved safety wording should be reviewed by someone qualified in your field.
2. Angry or distressed customers
What it covers. A caller who is shouting, crying, swearing, or plainly at the end of their patience. Also the quieter version: clipped answers and a statement like "this is the third time I've called."
Why a person. An upset customer is not looking for information. They want to be heard by someone with the power to do something. An AI that continues calmly through its script reads as indifference.
What the AI should do. Acknowledge briefly and without arguing, collect the essentials, and hand off. If no one is available, promise a callback from a named role within a stated time, and make sure that promise is recorded.
What helps. On Run with Jarvis Pro, sentiment and intent analysis is part of the plan, which makes it easier to find these calls in review and to see whether the handoff happened when it should have. Repeated frustrated contacts are also an early sign of a customer about to leave, a pattern described in the guide to detecting at-risk customers.
3. Pricing exceptions
What it covers. Any request to depart from the price list: a discount, a price match, a waived fee, a custom quote for unusual work, a dispute about what was quoted.
Why a person. Quoting the standard price is a rule. Changing it is a business decision about what a customer and a job are worth. If the AI can negotiate, callers will learn that quickly.
What the AI should do. Quote the standard price clearly. If the caller pushes for an exception, say that a specific person makes those decisions and arrange the handoff or callback. It should still offer to book at the standard price, because many callers ask for a discount and then book anyway.
The boundary to write down. Standing offers that you have approved are not exceptions. If you give a set discount to seniors or returning customers, put it in the instructions and let the AI apply it. Exceptions are everything you have not pre-approved.
4. Complaints about completed work
What it covers. The repair did not hold. The technician left a mess. Something was damaged. The customer was charged more than expected. The work is under warranty and the problem is back.
Why a person. A complaint needs someone who can look at the job record, decide what went wrong and commit to a remedy. That is accountability, and it cannot be delegated to software.
What the AI should do. Take the complaint without defending the company or admitting fault. Record what the customer says happened, which job it relates to, and what they are asking for. Route it to the person who handles complaints, as a priority.
A useful distinction. A warranty return visit with no anger attached can often be booked directly by the AI. A warranty return with an angry customer is a complaint. Write both rules.
5. Legal threats
What it covers. Any mention of a lawyer, a lawsuit, a regulator, a licensing board, the attorney general, a chargeback as a threat, or a news outlet.
Why a person. What gets said next matters, and it should be said by the owner or someone the owner has designated. The AI's only job is to stop the call from getting worse.
What the AI should do. Stay polite, avoid arguing or making any statement about fault, capture the details, and escalate to the owner immediately. No offers, no explanations.
Writing the rule. Keep the trigger list broad and the response narrow. This is the one category where "the owner, and nobody else" is the right routing.
6. Anyone who asks for a human twice
What it covers. The caller says "can I speak to someone?" or "representative" or "is this a real person?"
Why twice. The first request is often a reflex. A good AI receptionist should answer honestly that it is an AI assistant, say it can help with booking and questions, and ask what the caller needs. Many callers continue happily at that point. If the caller asks again, the request is deliberate, and continuing to resist it is the fastest way to turn a neutral caller into an angry one.
What the AI should do. On the second request, stop offering to help and start the handoff. If no one is free, say so and arrange the callback.
What it should never do. Claim to be a person. It is dishonest, and callers who discover it do not forgive it.
A decision table to adapt
This table is a starting point. Change the rows to fit your trade and your team.
| Situation | AI handles | Transfer to a person | Take message and schedule callback |
|---|---|---|---|
| New customer wants to book a standard job | Yes | No | No |
| Caller asks the price of a listed service | Yes | No | No |
| Existing customer wants to reschedule | Yes | No | No |
| Question about hours, service area or payment methods | Yes | No | No |
| Safety emergency in your trade | Approved safety direction only | Yes, immediately, to on-call | If transfer fails, urgent callback and alert |
| Urgent but not dangerous, such as no hot water | Books the earliest slot | Only if no slot is acceptable | If outside dispatch hours |
| Caller is angry or distressed | Acknowledges and collects essentials | Yes, warm, during staffed hours | After hours, with a stated time frame |
| Request for a discount or price match | Quotes standard price, offers to book | If the caller insists and staff are available | Otherwise, callback from the decision-maker |
| Complaint about completed work | Records the details | Yes, warm, to whoever owns complaints | If that person is unavailable |
| Warranty return, caller is calm | Books the return visit | No | No |
| Mention of a lawyer, regulator or lawsuit | Captures details only | Yes, to the owner | If the owner is unavailable, urgent callback |
| Caller asks for a person once | States it is an AI, offers help | No | No |
| Caller asks for a person a second time | Stops and hands off | Yes | If nobody is available |
| Commercial account with a contract question | Identifies the account | Yes, to the account owner | If unavailable |
| Vendor, sales or recruiting call | Takes a brief message | No | Low-priority message only |
| Job applicant | Collects contact details | No | Message to whoever hires |
Two things to notice. First, most rows have the AI doing something even when the call goes to a person. A handoff with the details already captured is far better than one without. Second, the right-hand column is never empty for a serious call. There is always a fallback.
Warm transfer vs cold transfer
How the call is handed over matters almost as much as whether it is.
Cold transfer. The caller is connected to another line with no introduction. The person who answers knows nothing. The caller explains everything again. For an already-upset customer, repeating the story is fuel.
Warm transfer. The person receiving the call is told who is on the line and why before they speak to the customer. They pick up with "Hi Mrs. Alvarez, I understand the repair from Tuesday is leaking again. I'm sorry about that." The conversation starts in a completely different place.
When each fits. Use warm transfers for emergencies, upset customers, complaints, legal threats and anything complex. Cold transfers are acceptable for simple routing where context adds nothing, such as a vendor asking for accounts payable.
What makes a warm transfer possible. The receiving person needs the context in a form they can absorb in a few seconds: the caller's name, the related job, and one sentence on the issue. On Run with Jarvis Pro, whisper, barge and transfer are part of the plan. A whisper lets the person receiving the call hear a brief introduction that the caller does not, which is the mechanism behind a warm handoff. The full use of those three tools is covered in the guide to live call coaching with whisper, barge and transfer.
Tell the caller what is happening. "I'm going to connect you with our service manager, and I'll let her know what you've told me so you don't have to repeat it." That sentence lowers the temperature before the transfer even begins.
When nobody picks up
This is where most escalation setups fail. The rule says "transfer to the owner". The owner is under a sink. The call rings out and lands in voicemail.
From the caller's side, that is the worst outcome available. They were told they would reach a person. They reached a beep.
The rule: no escalation ends in plain voicemail. Every transfer needs a defined fallback.
Step 1: a second person. For serious categories, name a backup. If the owner does not answer, try the service manager. Keep the chain short. Two attempts is enough. A caller should not sit through four rings on four phones.
Step 2: the AI comes back. If no one answers, the AI should return to the caller and be honest: nobody is free at this moment.
Step 3: a callback ticket with details. The AI creates a callback item containing the caller's name and number, the job or customer record, what the issue is in the caller's words, how urgent it is, and what the caller was promised. On Run with Jarvis Pro, callback scheduling is part of the plan, so the promise becomes a scheduled item and not a note someone has to remember.
Step 4: a specific promise. "Someone will call you back" is not a promise. "Our service manager will call you within the hour" is. Only promise a time frame you have decided in advance that you can keep.
Step 5: an alert the team will see. The callback has to reach a human who will act on it. A message nobody reads until morning is voicemail with extra steps.
Step 6: close the loop. Someone marks the callback as done. Open callbacks older than your promised time frame should be visible to the owner every day.
For emergencies, add one more line: if the transfer fails, the AI repeats the approved safety direction before ending the call.
After-hours rules
Escalation rules change when the office is closed, and they should be written separately.
Decide what counts as urgent at night. During the day, a complaint goes straight to a person. At 2 a.m., it probably becomes a first-thing-in-the-morning callback. A safety emergency is urgent at any hour. Draw the line clearly for each category.
Name the on-call person and the schedule. "The on-call technician" must resolve to a specific phone on a specific night. An out-of-date rotation is the most common cause of an after-hours escalation reaching nobody.
Let the AI book more on its own. After hours, the AI should resolve as much as it can without waking anyone: book the first morning slot, quote the standard price, explain any after-hours fee you charge. Reserve the wake-up call for true emergencies and for callers willing to pay for immediate service, if you offer it.
Be honest about response times. If nobody will see a non-urgent callback until 7 a.m., the AI should say so. Callers accept "first thing tomorrow" far better than a vague promise that turns into silence.
Protect the on-call person. If every upset caller is transferred at midnight, your on-call technician will stop answering. Tight after-hours rules are what keep the urgent path reliable.
The broader approach to nights and weekends, including pricing and staffing, is in the after-hours calls playbook.
How to write the rules down
A rule the AI can follow has four parts.
Trigger. What the caller says or what situation is described. Use examples in the caller's language. "My house smells like gas" is more useful than "gas emergency".
Action. What the AI does first: safety wording, acknowledgment, collecting details.
Route. Who receives the call, by role, with a named backup.
Fallback. What happens if the route fails: callback ticket, time frame, alert.
A written example for one category:
- Trigger: caller describes a previous job that has failed or was done badly, or uses words such as "still broken", "came back", "not fixed", "made it worse".
- Action: apologize for the trouble without assigning fault, confirm the customer and the job, ask what happened and what they would like done.
- Route: warm transfer to the service manager during office hours.
- Fallback: if no answer, create a priority callback for the service manager, promise a call within two business hours, alert the owner.
Write one of these for each category. Six categories, a page or two in total.
Keep the list short. Six to ten rules. A list of forty will contradict itself.
Use plain words. If a new office hire could not follow the rule, the AI will struggle with it too.
Decide the gray areas in advance. What about a caller who is merely annoyed? A customer who wants a small discount? Pick a side and write it down. You can change it after the first review.
Reviewing escalations every week
Rules written once and never checked drift out of date. A short weekly review keeps them working. Allow twenty minutes.
Pull the week's escalated calls. Every call that was transferred or turned into a callback.
For each one, ask whether it matched a rule. If yes, the rule worked. If no, find out why the AI escalated. Usually the instructions had a gap and the AI did not know how to answer.
Then look for the misses. Sample calls that were not escalated, especially long ones and ones where the caller sounded unhappy. Should any have gone to a person? On Pro, AI transcription and call recording with playback make this quick, since you can read first and listen only where needed.
Check the fallbacks. How many transfers were not answered? How many callbacks were completed within the promised time? An unanswered transfer rate that creeps up is a staffing or routing problem, not an AI problem.
Change one or two things. Add a missing answer so a routine question stops escalating. Tighten or loosen a trigger. Update the on-call list. Small weekly changes beat a large quarterly rewrite.
Keep a simple log. Date, what changed, why. In three months it will show you how the rules matured, and it is useful for training staff.
This review fits naturally alongside a general call quality check. If you already score calls, add an escalation question to the sheet, as described in the phone call quality scorecard. The first month deserves closer attention, with reviews every few days, as laid out in the first 30 days of an AI receptionist.
Common mistakes
No fallback. The transfer rings out and the caller gets voicemail. Covered above, and the most damaging error on the list.
Escalating everything. If the AI hands off at the first sign of difficulty, your staff are still answering the phone all day, now with a delay.
Escalating nothing. A setup with no human path will eventually mishandle an emergency or a complaint in a way that costs more than the system saves.
Rules by name, not role. "Transfer to Mike" breaks the week Mike is on vacation.
A stale on-call list. Check it every week.
Cold transfers for upset callers. Making an angry customer repeat the story.
Vague promises. "Someone will get back to you" with no time frame and no owner.
Letting the AI negotiate. Price exceptions decided by software become the new price list.
Pretending to be human. It always comes out, and it colors everything after.
Never reviewing. The rules that fit your business in month one will not fit it in month six.
Where to start
Sit down with whoever answers your phone today and list the calls they would never handle alone. Sort that list into the six categories. For each, write the trigger, the action, the route and the fallback. Then write a second, tighter version for after hours.
Test the rules before customers do. Call your own line and play each scenario: the emergency, the angry customer, the discount request, the request for a human. Confirm that every path ends with a person or a recorded callback, never a dead end.
If you are thinking about how small businesses structure roles and responsibilities more generally, the Small Business Administration has planning resources that are a reasonable companion to this exercise.
Run with Jarvis plans are on the pricing page: Core $500, Pro $750 and Elite $1,200 per month, with zero setup fees, unlimited users and month-to-month terms. To walk through escalation rules for your trade and how transfers and callbacks would be set up, get in touch.
