An operating system
for the physical world.
Software has eaten the industries where the work is information. It has barely touched the ones where somebody has to go to a building, diagnose a real thing, price it, schedule it, do it and get paid. Gable Pro is the operating system for that work — and it is ours, so everything this site claims runs here first: answers that show where they came from, prices worked out by code, and anything that costs money waiting until a person approves it. Open it and check.
Live at gable.pro, open it straight away.
Two halves, one brain.
Software has eaten the industries where the work is information. It has barely touched the ones where the work is physical — where somebody has to go to a building, diagnose a real thing, price it, schedule it, do it and get paid. Gable Pro is an operating system for that work.
It is two halves over one brain. One half takes a problem — described, or photographed — and works out what is likely wrong, what it should cost, and who should do it. The other half is what the business doing the work runs on: customers, jobs, its own pricebook, estimates, the schedule, the invoice and the payout. The same agent loop drives both, so the trade knowledge is one asset and a fix lands on both sides at once.
Home services is where it runs today and where the model was proved, because it is the hardest version of the problem: the highest variety of jobs, the smallest businesses, and the least tolerance for a wrong number. The parts that make it work reach beyond a house. A diagnosis, a defensible price, a schedule that keeps every hour to one visit, an invoice held within its estimate, and an approval before anything irreversible — those are the same in a plant, on a site, or across a portfolio of buildings.
It is also why the rest of this site can be specific. Gable Pro is shipped software with real customers rather than a prototype, and every safeguard we describe here was built for it first, in front of real people spending real money, and proved there before it was pointed at anything else.
-
Gable Pro, the demand side
Describe the problem, or photograph it. It works out what is likely wrong from a library of checked guides, says what the work should cost, and finds somebody vetted to do it. Every answer shows the guide it came from, and the price is computed rather than written by a model.
-
Gable HQ, the operating side
What the business doing the work runs on: customers, jobs and a pipeline, its own pricebook, estimates a customer signs, a calendar that keeps every hour to one visit, invoices, payouts, and a shared inbox that reads a new enquiry into a job. Its own front door at hq.gable.pro.
-
It runs unattended
A durable job queue does the work that has to happen on time rather than on demand: the morning brief, appointment reminders, estimate follow-ups and overdue-invoice nudges. The business is still running when the office is shut.
-
One brain underneath
Both halves run the same agent loop behind a different profile, so the trade knowledge is one asset and a fix lands on both sides at once.
From the first call to the last visit.
Past the jobs board and the invoice, Gable Pro carries the rest of what a trade business does in a week: prices it can stand behind, the rules in force where the job is, a page customers find it on, plans that bring them back, jobs too big for one visit, the crews it hires in, and the phone.
Cost guides
A public guide for each kind of job at gable.pro/guides: what usually causes it, when to call a pro, and an itemized estimate the pricing code works out when the page loads. Every figure on a guide comes from that code, and a test enforces it. Where the job is known, labor and materials are adjusted for its state.
Building codes and permits
The assistant can say which residential code a state has adopted, whether local amendments are allowed, who the authority is, and whether a kind of work usually needs a permit, with a link to the source. Clearances, spans and sizes it leaves to the code book and the local inspector, where they belong.
A page customers find you on
Every business gets a public page built from seven sections: about, services, gallery, booking, a quote form, where it works, and contact. They can be reordered, hidden and resized, with eight colour themes. A quote request lands as a job on the board, and online booking offers only slots the real calendar has free.
Service plans
Recurring work on a schedule: every month, every second Tuesday, the 15th. Each visit opens as a real job up to two weeks ahead. A plan is billed either per visit or up front, and a paid-up-front plan marks each visit covered until a refund stops it.
Projects
A job too big for one visit becomes a chain of phases, each waiting on the one before it. The quoted, billed and paid totals are summed from the phases, and one combined estimate covers the whole project, and it goes out once every phase is priced. Changing a signed estimate produces a change order beside it, and the signed version stays as it was.
Subcontractors
A directory of the crews a business hires in, with trade, rate, and insurance and license expiry dates. Work goes out per job with a scope and a price, and the sub accepts it straight from a link, account-free. The link shows the site and the scope and keeps the customer’s details private. Work goes only to subs whose insurance and license are current.
Phone answering
A business on the Pro plan buys a number inside Gable, and the line answers itself. It says it is automated, takes the caller’s name, number, address, the problem and how urgent it is, escalates an emergency at once, and transfers anyone who asks for a person. Prices and arrival times stay with the business to give. Every call becomes a lead with a full transcript for somebody to call back.
Four things it does, all running for real.
One system, in the order a job moves through it: work out the problem, price it, send someone, record what happened.
Gable Pro itself is the evidence rather than the product on sale. What a buyer engages us for is the engine underneath it, or a service line.
The most expensive step is the one before the work.
A request arrives as three sentences and a blurry photo. Somebody guesses what it is and who to send, and finds out on arrival whether they guessed right: wrong trade, part missing from the van, second visit booked, and a week gone on a ten-minute fix.
How diagnosis works What it refuses to answer, how a photo is read, and where it runs.
Look it up, check it, answer, or admit the gap.
What matters is what it refuses to do. Every part of an answer comes out of the guides, and it commits to a trade or a price only once a fixed check says the guide it found is really about your problem.
-
It searches two ways at once
One search matches your words, the other your meaning. It looks at whole guides and at single paragraphs, so one paragraph that is exactly right can surface a long guide that would otherwise have been missed.
two searches -
A guide about something else is thrown out
Unless a guide shares something distinctive with what you asked, it is rejected however well it scored. This is what stops a confident answer about the wrong thing.
fixed check -
Answer, or keep looking, or say so
Covered by the guides, you get an answer showing where it came from. Outside them, it asks one narrowing question, then searches the web under a quality check, and tells you when that fails too.
in that order
It runs where your data already lives.
Inside your own boundary
Your own cloud, your own building, or a network cut off from the internet. If the search service is unreachable it does less rather than breaking, and which AI provider it uses is a setting you can change.
Your documents alongside ours
Building manuals, an equipment list, warranties, or the repair history of a whole portfolio load in beside our guides, so an answer can quote your paperwork where it exists.
Your data stays out of training
Your documents, photos and requests are kept out of the training of any model anyone else gets. How long anything is kept is a setting, and by default what you send is discarded once the answer is produced.
The AI explains the price. Code writes the number.
An AI is good at reading a description of a job and bad at arithmetic you intend to bill from, and the mistake is invisible, because a made-up total looks exactly like a real one. So the AI only decides which priced recipe applies. Code does the arithmetic, and the same details always give the same number.
How pricing works The four ways it prices a job, and how it reads a contractor’s quote.
Four ways to price a job, best one first.
Something that can only price jobs it has seen is useless on a real one. Something that always invents a number is dangerous. So it works down four options in order and tells you which one answered.
-
Your price list
If the business has said what it charges for this work, that price wins. Their real price, sized to the job, rather than ours with a markup.
basis: pricebook -
Our own price list
Worked-out costs for common jobs: hours of labour, quantities of materials and fees, each scaling with area, length or how many, at each quality of finish.
basis: catalog -
Worked out from current prices
For jobs outside our list, it builds a costing from current real prices, keeps it only if it passes a quality check, and reuses it next time.
basis: learned -
A rough range, clearly labelled, or an honest blank
Last resort: a range for the whole job, labelled unmistakably as rough. If that is out of reach, it says so rather than filling the space.
basis: ballpark
Is this quote reasonable?
Upload the PDF, the Word file, or a photo of the paper. It reads out what the quote covers, and only that, then prices the same work itself and tells you where the quoted total sits against its own figure.
The answer is arithmetic rather than an opinion. Fixed code decides whether you are being overcharged.
Send a quote to reviewA status column that stays true.
Every repair system has a status field, and in every one it is out of date, because keeping it right is unpaid admin that loses to the real work. Here the status is a result of what happened:
There is very little to click, on purpose.
How operations work A calendar that keeps every hour to one visit, quote through to payment, and who can see what.
Each hour holds one booking.
Beyond “unlikely”. The database itself refuses two visits that overlap, so a phone call and an email arriving at the same moment, or two people booking in two browser tabs, end up with one visit at 2pm and a refusal for the other. The app checks as well, but only so the error message is a useful one.
The paperwork is where the money leaks: a quote that went unchased, a deposit taken twice, a bill for work the customer turned down. Each is made impossible here rather than written up as a procedure.
See it fail correctly| Gaps as well as slots | It reports each free stretch and how long it lasts, so a six-hour job in a week of four-hour gaps is split across two mornings instead of coming back empty. A shortened list always says it was shortened. |
|---|---|
| Quotes that expire | A quote has a date on it, the customer is told that date, and the approve button stops working after it. One live quote per job, older ones kept rather than deleted. |
A job title is just a starting point. Permissions are what count.
Access follows permissions rather than someone’s job title. It checks whether that person has the specific permission, which is what lets you give one extra permission to one person instead of inventing a whole new role for them.
How much you can see is part of the permission
“Can see jobs” and “can see their own jobs” are two different permissions, so the code checks exactly the permission it means to. An engineer sees their own work; the person booking sees the whole company’s.
Contact details are a separate permission
When details are hidden the screen says so, because a phone number that is missing and one that is hidden must look different. Inboxes and phone lines are granted one person at a time, including to the owner.
Staff who need only a record
Someone on the crew can be scheduled, assigned work and tracked as a plain record instead of an account. The people who need a login get one and the rest are records, which is what makes a big crew affordable to run.
The one rule everything else turns on.
One assistant with a set of narrow, well-defined tools, rather than a crowd of AI specialists handing work to each other, which costs more every turn and leaves the question “so which one decided this?” open when someone audits it.
The assistant can say anything about what the business is able to do. It can commit the business only to what somebody agreed to. The guard is more than an instruction in a prompt, because a prompt is advice and advice gets followed most of the time. It is how the code is built: the tools that would change something only suggest the change. Each checks what it was asked to do, writes a one sentence summary, files it as a suggestion, and stops.
How the platform works How much it may do alone, what happens when things go wrong, and connecting your systems.
A dial a business owner can actually think about.
“Let it do things I could undo myself” is a sentence an owner can hold in their head. “Let it use tool number seven” is opaque to them. So how much it may do alone is set by reversibility, and it can be set per line: the overflow phone answers on its own while the main line still asks first.
See the dial in practice| 0 · Watching | It says what it would have done and stops there. For building trust early, while the team reads along. |
|---|---|
| 1 · Asks first | The default. Everything arrives as a card you approve or cancel. A week at this setting is the fastest way to learn what you would have let it do. |
| 2 · Acts if it can be undone | Changes inside your own records happen on their own. Anything that leaves the building still asks. |
| 3 · Acts on its own | Including sending things out. Deliberately hard to turn on, and only an owner can do it. |
| Irreversible | A quote, an invoice, a message to a customer. Once sent, they stay sent, so they are refused at every level below full independence. |
The failures are assumed rather than hoped against.
The system knows what actually happened: which tools ran and what came back. So when a reply claims more than the record supports, that is caught rather than trusted.
-
Claiming something was done when the record shows otherwise
The record is empty of runs and suggestions, yet the reply says a change was made. Checked against the record, then corrected before you see it. Same for “here are three contractors near you” over an empty result.
-
Do less, rather than stop
Lose search, or email, or payments, or the database, and each one takes away that one thing and leaves the rest working, with the screen saying what is missing. Every call out has a time limit, because something that just hangs stays silent about its failure.
This is running software. The assistant, the rule that it asks before it acts, the dial for how much it may do alone, and every turn recorded, all in production in Gable HQ, running a home services business end to end.