A lot of “AI receptionists” try to never escalate. That looks good in a demo. In a field business it produces invented slots, invented services, and a customer who still has to call.
WhatTheJob treats handoff as a designed state. Queries shows what AI is handling, what is ready to convert, and who needs a person. The customer is told. No fake arrival time. No muted dispatcher.
The useful test is not “did the bot reply.” The useful test is “did anyone promise something the company cannot stand behind.” If the corpus cannot answer, empty context wins. That is slower than a hallucination. It is also how you stay in business.
Asking for a person hands off. A missing address on a job request is not a handoff. The assistant keeps asking.
Why never-escalate is a trap
Demos reward fluency. A fluent answer to “can you come at 3” is easy to generate. A true answer requires your hours, your crew, your catalog, and a person who will send someone. If those are not in the corpus, a fluent 3 p.m. is a lie.
Field customers remember the lie. They wait. They write again. They call. The office then inherits a promise nobody made on purpose. The crew inherits a window that was never scheduled. That is not leverage. That is a second inbox made of apologies.
Never-escalate also hides the desk from the week. If every thread looks “handled,” nobody sees the pile that needs a person. Needs you exists so that pile has a name. It is a first-class queue, not an error toast.
AI describes the same rule in product language. This note is the operational one: when to stay on AI, when to ask, when to hand off, and what the customer hears.
Owners care because a bot that never escalates becomes a reputation problem. Dispatchers care because they inherit promises. Crew care because they drive to a window nobody booked. The designed state is how those three people stay on the same story.
What stays on AI
Catalog and hours questions can stay on the model when the answer lives in your company knowledge. “Do you service that area.” “Are you open Saturday.” “What’s included in a tune-up.” Those are the turns where retrieval should be boring and correct.
A clarification when a field is missing can stay on AI. The thread is a job request. The address is not there yet. The assistant asks for the site. That is not a handoff. That is finishing the brief so a person can later convert it.
A drafted request, with confidence, waiting for a human, is also AI doing its job. Extraction is not escalation. AI drafts the field job. You still approve it is the note for that stack. Convert is still a person.
What stays on AI is bounded by the corpus. Your FAQs. Your hours. Your catalog. That customer’s history in this workspace. Not another company’s thread. Not a price the model invented. Not a loyalty balance it should never touch.
What does not stay on AI
Asking for a person hands off. “Can I talk to someone.” “Put me through.” “This is the third time.” Those are not catalog questions. Software should not answer them with a cheerful FAQ.
A question the corpus cannot answer hands off. If the model cannot ground a reply in your knowledge, it should not improvise. Empty context is better than a made-up fact. The thread goes to a person. The customer is told.
A situation that needs judgment hands off. Two jobs colliding. A site that is unsafe. A customer who is angry in a way a template will make worse. The product is not a personality. It is a path that knows when to stop.
Knowledge answers never create a job. That belongs in both columns. The model can say what you offer. It cannot decide you sold it. A person still presses Convert.
If the model is wrong, handoff is how you recover before the crew is in the van. If the integration fails, the thread is still in Queries. A person can write. The week does not wait for a fluent paragraph to come back.
The customer should hear the handoff
A silent escalation is how trust dies. The customer thinks the bot is still “handling it.” The office thinks the bot is still “handling it.” Nobody is handling it.
When software hands off, the customer is told a person will take the thread. They are not told an arrival time. They are not told a name the model invented. They are not left on read.
On the desk, Queries should make the state obvious:
- AI is handling this turn.
- This draft is ready for a person to convert.
- This thread needs you.
Those are different piles. Mixing them is how a catalog question sits next to an angry callback and both look like “inbox.”
Outbound replies stay labeled. AI assistant when the model wrote it. A teammate’s name when a person wrote it. After handoff, a person writes. The model does not keep talking as if it were the owner.
A thread that should leave the bot
Composite, not a case study. A customer writes: “Can you replace the board on a 14-year-old unit tonight, and what will it cost.” Your catalog has “AC repair.” It does not have that board. Your FAQs do not list a night rate. The customer then writes: “I need to talk to someone. The house is hot.”
The first message is not a clean catalog question. The corpus cannot price a board you do not stock. Software should not invent a number. The second message is a person request. That hands off.
The customer is told a person will take it. A dispatcher opens the raw thread. They can ask for the model, the site, and whether this is tonight or first thing. They can convert a job if it is a job. They can say no if it is not work you do. None of that required a bot that never escalates.
A different thread: “What time do you close on Saturday.” That can stay on AI, from your hours. A third: “AC stopped. 24 King Street.” That is a draft, and a missing window is a clarification, not a handoff.
The office should be able to explain those three outcomes to a new dispatcher in one sitting. Catalog question stays. Person request leaves. Job request becomes a draft. If your current bot treats all three as “I can take care of that,” you do not have a receptionist. You have a delay with a smile.
What the desk does after Needs you
Needs you is not a graveyard. It is a queue a person works.
A dispatcher opens the thread. They see the raw messages, including the one that asked for a person. They see what AI already tried, if it tried anything. They write as themselves. The label on the outbound reply is their name.
Sometimes the next step is still a job. The customer needed a person to trust the window. The dispatcher converts after they talk. Sometimes the next step is a no. You do not offer that board. You do not work tonight. A person can say that without the bot improvising a maybe.
Sometimes the next step is knowledge you should have added last month. Saturday hours were missing. A catalog line was vague. The handoff taught you that. You fix the corpus later. You do not let the model guess in the meantime.
The customer already heard that a person would take it. Do not put them back on the bot for the same question. That is how “handoff” becomes a loop.
What we will not automate
We will not hide escalation so the demo looks smoother. We will not invent a slot to avoid Needs you. We will not let a knowledge answer create a job. We will not keep the model talking after a person owns the thread.
A missing field can stay on AI. A missing truth cannot. If you cannot answer from your FAQs, hours, or catalog, stop. Tell the customer. Give the desk the pile.
That is the whole note, pointed at one temptation. Never-escalate is a product bug with good metrics.
Bring the thread that usually goes wrong, not the one that makes a demo reel. The wrong thread is how you find out whether Needs you is real. If we cannot show that pile on the desk, we are not walking the product. We are walking a sticker.
How this sits on isolation and the path
Handoff is also a security instinct. If the model cannot answer from your corpus, the alternative is not “try another company’s FAQs.” The alternative is a person. An AI with the wrong company's data is a leak is why empty context wins.
The rest of the path does not change. Conversation in. A draft when it is a request. A person who approves. Assign. Crew App. Done. Handoff is the branch that keeps the first half honest.
Contact sales if you want to walk Needs you with a thread that should not stay on a bot. Bring the one that usually goes wrong. That is the one that matters.
If your current week hides those threads in a personal inbox, the first fix is still intake. Handoff cannot save a conversation the company does not own. Read the chat note, then come back here. The pile named Needs you only works when the pile exists in Queries.
Questions
No. Needs you is a designed state. Inventing a slot, a service, or an arrival time so the bot never escalates is the failure.
No. Asking for a person hands off. A missing field is a clarification. The assistant can keep asking. Knowledge answers never create a job.
That a person will take the thread. No fake arrival time. No muted dispatcher. Queries shows what AI is handling, what is ready to convert, and who needs a person.
A person now owns the thread. Catalog and hours questions, from your knowledge, can stay on AI before that point. After handoff, a teammate writes.