A sticker on a generic category is easy to print. It is harder to let a model create work in a business that sends people to other people’s buildings. WhatTheJob classifies the turn, extracts a request, validates it against your catalog, then stops.
That stop is the product. The model extracts. Software validates. You act.
If you came here looking for autonomous dispatch, you are in the wrong note. We do not replace your dispatcher. We do not let a model invent a service you do not offer, a price you did not set, or a crew member who was never assigned. AI is the product page for the same stack. This note is why the stack stops where it stops.
The model extracts. Software validates. You act.
What the model is for
A customer writes a sentence. Sometimes it is a clean brief. More often it is a mix: a problem, a place, a time that is not a time, a photo, a follow-up that changes the first message. A person can read that. A person should not have to retype it into a sheet while the next message arrives.
WhatTheJob uses the model to turn that sentence into fields a desk can review:
- Service, checked against your catalog, not against a guess.
- Customer, when the thread already knows who this is.
- Site, including the messy parts: unit, gate, “the shop around the back.”
- Date and time window, with a score when “tomorrow afternoon” is only mostly sure.
- The problem, in words a crew member can use.
- Priority, when the thread actually said it, not when the model felt dramatic.
Each field carries a confidence score and a note on what is missing. If afternoon is only 82% sure, you see it. If the service is not in your catalog, software rejects the guess. Empty retrieval is better than a made-up fact.
The model can also answer catalog and hours questions from your company knowledge. That is a different turn. It never creates a job. When AI cannot answer, handoff is the product is the note for the turn that should leave AI.
Classify the turn before you extract a job
Not every message is a job request. Software should route the turn before a job is even considered.
A catalog question: “Do you do duct cleaning?” That can stay on AI, from your FAQs and catalog, not from invented facts.
A person request: “Can I talk to someone?” Asking for a person hands off. The customer is told. No fake arrival time. No muted dispatcher.
A job request: “AC stopped. 24 King Street. Tomorrow afternoon?” That is a draft, not a job.
A follow-up that changes the draft: “Actually unit 4, and the shutoff is in the hall closet.” The thread is the source. The draft updates. The job does not exist until a person says it does.
If you skip classification, you get the demo version of AI. Everything looks like work. The office spends the afternoon undoing jobs that were never sold. The crew drives to a site that was a question.
Confidence is part of the draft
A draft without scores is a guess wearing a uniform. The useful part of extraction is not the fields. It is the fields plus how sure the model is, plus what it could not find.
You should be able to see:
- Which field is thin.
- Whether the thinness is “not in the thread yet” or “the thread conflicted with itself.”
- Whether software already rejected a service that is not in the catalog.
The office then does what offices already do, with less retyping. They read the raw thread. They edit what is off. They convert, or they leave it as a conversation. Needs you stays a first-class queue. Missing fields stay visible. Guessing does not.
Outbound replies are labeled by software. AI assistant when the model wrote it. A teammate’s name when a person wrote it. The model does not get to invent who is speaking. That is a small rule with a large trust effect. Customers can tell when a bot is pretending to be the owner.
What the model will not do
The model does not invent prices. The model does not invent loyalty points. The model does not promise an arrival time it cannot keep. The model does not assign a crew member. The model does not create a job.
Those limits are not a later phase. They are the difference between a draft and a dispatch. A business that sends people into other people’s homes cannot treat those as prompt settings.
Deterministic software sits between the model and the record. It checks the catalog. It keeps retrieval inside your company. It refuses empty confidence dressed up as a fact. Security is the page for isolation and the AI corpus. The short version for this note: another company’s history is not context. It is a leak. An AI with the wrong company's data is a leak is the longer version.
If the model is wrong, a person still has the raw thread. They edit. They do not convert. The customer is not told a story nobody approved. If the integration fails, the thread is still there. The week does not depend on a fluent paragraph.
A draft you can stand behind
Composite, not a case study. The thread says: “AC stopped. 24 King Street. Tomorrow afternoon? We had you last summer for the same unit.”
WhatTheJob classifies it as a job request. It drafts AC repair, that site, a window, the problem. It may attach that customer if the workspace already knows the number. Software checks AC repair against your catalog. The afternoon window shows as less than certain. Last summer is history in this workspace, not a reason to invent a part.
A dispatcher opens the thread. They see the raw messages next to the draft. They tighten the window. They convert. They assign. The crew member opens a job, not a paragraph.
Change one fact and the stack should stay honest. If you do not offer AC repair, the draft does not invent it. If the address is missing, the assistant can ask, or a person can ask. Asking for a missing field is not the same as handing off. Handing off is for a person request, or for knowledge the corpus cannot answer.
What a person does at the gate
The gate is not a rubber stamp. It is the moment the company accepts the work.
A dispatcher or owner opens the thread. They read what the customer actually sent, including the follow-up that changed the first message. They look at the draft the way they would look at a sheet someone else typed: useful, not trusted until checked.
They edit the window. They add the unit. They drop a service that is not yours. They convert when they mean it. They leave it as a conversation when they do not. Needs you stays visible so “not yet” has a place to live.
They still assign. That is a second decision. The model does not get a vote. The crew member who opens the job should be able to believe the record, because a person stood behind it.
If two messages conflict, a person decides which one is true. If the customer asked for a person, the gate is already a handoff, not a convert. If the catalog does not contain the work, the honest outcome is no job, not a near-match.
This is slower than a demo that creates jobs from every sentence. It is also the only version of extraction we will sell. A week of invented work is not a productive week. It is a cleanup week with vans attached.
Why we will not sell autonomous
Autonomous sounds like fewer people. In the field it usually means promises nobody can keep.
A model that creates jobs will create the wrong ones. A model that assigns will send the wrong person. A model that prices will invent a number you then have to honor or walk back. A model that texts “we’re on the way” will do it from a confidence score.
We will not put that on a marketing page and call it innovation. The stack is extract, validate, human gate. AI says it in product language. This note says it as a company rule.
If a later phase adds more assist, it still stops before create, assign, and promise. The people who run the path stay in the path.
How this sits on the path
This note is the middle of field service requests start in chat. The conversation is already a company record. Extraction is how the desk stops retyping. The gate is how you stay in business.
After you convert, you still assign. After you assign, the Crew App is the field surface. After the crew marks it done, the customer is told. Optional survey and optional loyalty are switches, not the reason the model exists.
Contact sales if you want to walk Queries, the draft, and the gate with your own catalog, not a generic tour. Bring the messy thread. That is the point of the product.
Bring the one where “tomorrow afternoon” was wrong, or the service was not yours, or the customer asked for a person in the same breath as the brief. If the draft cannot show those scores, and the gate cannot leave the thread as a conversation, we are not looking at this product. We are looking at a sticker on a form that still needs retyping.
Questions
No. The model drafts a request from the conversation. A dispatcher or owner reviews the raw thread, edits what is off, and decides it is a job. That gate is the product.
Service, customer, site, date and time window, the problem, and priority, each with a confidence score and notes on what is missing. Software then checks the draft against your catalog.
No. If the service is not in your catalog, software rejects the guess. Empty retrieval is better than a made-up fact.
No. It does not assign a crew member, invent a price, invent loyalty points, or promise an arrival time it cannot keep. A person still decides.