If an AI answers your phone, give it a job sheet
A decorator I know set up an AI receptionist last spring. Clever bit of kit. Answered every call, never got flustered, worked Sundays.
Three weeks in, it quoted a customer £180 to paint a stairwell.
He’d never have taken that job for £180. The stairwell needed a scaffold tower, two days minimum, and he’d have said no anyway because he stopped doing stairwells after the last one. But the AI didn’t know any of that. It knew he was a decorator. It knew painting costs money. So it did what these things do when they don’t know: it produced something plausible.
The customer had it in writing. He did the job.
The model isn’t the problem
There’s a habit in this industry of blaming the AI. The model hallucinated. The voice agent isn’t smart enough yet. Wait for the next version.
That’s mostly wrong, and it’s an expensive kind of wrong because it means you wait instead of fixing it.
An AI answering your phone has read a large fraction of the internet. What it has never seen is your day rate, your postcode list, the three jobs you swore you’d never take again, or the fact that you’re not VAT registered. It has world knowledge and zero business knowledge, and world knowledge is exactly what makes it dangerous — it’s confident, fluent, and completely making it up.
I’ve been building the electronics for 360° camera systems for four years, and I run a decorating firm. Both jobs teach you the same lesson: a system with a missing input doesn’t fail loudly. It fails plausibly. The reading looks fine. The quote looks fine. You find out three weeks later.
Three things it will get wrong
In roughly this order of cost:
It will invent a price. Not maliciously — it’ll average something reasonable-sounding from context. Reasonable-sounding is not your rate card.
It will take work you don’t do. Every trade has a list. Gas appliances if you’re not registered. Textured ceilings from before 2000. Anything above two storeys without a scaffold. Flat roofs. Insurance jobs. The AI will cheerfully book all of it, because “roofing” is in your description and the job involves a roof.
It will make claims about your credentials. This is the one that should worry you most. Asked “are you Gas Safe registered?”, an AI with no data will guess yes, because a plumber usually is. That’s a statement about your legal standing, made by software, to a customer.
What actually fixes it
You write the rules down once, in a format the AI can read.
That sounds obvious and slightly boring, which is probably why almost nobody does it. What’s changed is that there’s now a proper way to do it rather than everyone inventing their own.
On 12 June 2026 Google Cloud published the Open Knowledge Format — OKF. It’s an open, vendor-neutral spec for representing the knowledge an AI agent needs, as plain markdown files with a bit of YAML at the top. No SDK, no account, no platform. If you can open a text file you can read it; if you can put a folder on a web server you can ship it.
It was built for enterprise data teams — table schemas, metric definitions, incident runbooks. But the shape of the problem is identical for a one-van business. Scattered knowledge that lives in one person’s head, which an AI needs and doesn’t have.
The bit that matters for you
On 24 July, six weeks after the first release, Google shipped v0.2, and it added exactly the thing that makes this usable for a trade.
v0.1 asked what is this? v0.2 asks does anyone stand behind it?
A concept in a v0.2 bundle can carry a verified field recording who confirmed it and when. From that, whatever reads the file works out a trust level: nothing there at all means unverified; a machine confirmed it means machine-confirmed; a person confirmed it means human-reviewed.
That’s a small change with a big consequence. It means the difference between “these are my rates, I checked them, that’s my name on it” and “some software generated this, nobody looked” is a thing your AI can actually see — and filter on — before it opens the file.
There’s a stale_after date too. An absolute date, deliberately, not a countdown. Your prices hold until March; after that the pack flags itself for review instead of quietly quoting last year’s day rate.
The free tool
So I built a job sheet.
Job sheet for your AI phone answerer →
You fill it in once. Business name, trade, the jobs you do, where you work, your rates, your hours, and the jobs you won’t take. It gives you back a tidy OKF bundle you can hand to your AI receptionist, plus a ready-made prompt you can paste straight into whatever you’re already using.
Everything happens in your browser. Nothing you type is sent anywhere or stored. There’s no signup, no email gate, and it’s free permanently — I’m not going to pretend otherwise later.
Three things worth knowing about how it works:
Box 12 is the important one. “Jobs you won’t take.” It’s the field people skip and it’s the one that saves you the most money, because it’s the only thing standing between an AI and a booking you’d have to ring back and cancel. Prefilled with sensible defaults per trade — edit them, they’re guesses about your business and you know better.
The sign-off is real. There’s a tick box at the bottom confirming the details are right. Tick it and your pack is marked human-reviewed, with your name and the date. Leave it unticked and the pack is honestly marked unverified. I’d rather ship that than a signature that never happened, and now the format has somewhere proper to put the distinction.
This will not change how you show up in Google. I want to be blunt about this because you’re going to see tools in the next twelve months that imply otherwise. OKF is not a ranking signal. Google’s search systems do not fetch a bundle from your website and rank you on it. It’s a format for the AI that works for you, not for the search results. Anyone selling it as SEO is selling you something else.
The part I’m most pleased with
OKF v0.2 introduced a concept type called an Attested Computation — a sanctioned way to calculate a value, plus a deterministic check that the sanctioned thing is what actually ran. The agent may fill in the parameters. It may not touch the formula. If it improvises, the check fails and the number doesn’t get used.
Google’s examples are all BigQuery SQL and finance metrics. But that is, structurally, exactly what a quote is. Your pricing formula, fixed and signed. The AI fills in hours and materials. It cannot invent a rate, because the arithmetic is checkable and gets checked.
That’s the £180 stairwell solved properly rather than papered over with a longer prompt — and it’s in the tool now. Put your rates in and the pack comes back with a computations/job-estimate.md holding your formula and the parameters an agent is allowed to fill, plus an attesters/estimate_equality.js that checks any quote against it. Swap a rate, add a term, drop the minimum charge, and the check fails. There’s a simulator in the tool so you can watch it reject a made-up number before you trust it with a customer.
One thing I want to be precise about: applying Attested Computation to a trade quote is my reading of the spec, not something Google has blessed for this use case. Their examples are all data-warehouse metrics. The abstraction is explicitly flexible and I think a quote sits inside it exactly — but that’s an argument, not an endorsement, and you can read the full build spec and disagree with me. The code is on GitHub under Apache 2.0, same licence as Google’s own catalogue, because a knowledge format is only worth anything if more than one person speaks it.
What’s next
A hosted bundle URL, so you get a link to hand your agent instead of having to find somewhere to put the files yourself.
After that, the Surrey Decorators pack goes up as a public sample bundle. Google is cataloguing OKF tools built outside their own org and has asked for sample bundles — and a real pack from a real one-van business seems a more useful contribution than another BigQuery example.
If you run a trade and you’ve got an AI answering calls — or you’re thinking about it — go and fill the sheet in. It takes about ten minutes. And if the defaults for your trade are wrong, tell me, because I only know decorating from the inside.
Jetnor Muhaj is a London-based engineer and technology founder. He is a Senior Electronic Engineer at META Camera Ltd, founder of AI Construction Technologies, and a director of MUHAJ CONSTRUCTION LTD, trading as Surrey Decorators — which is where most of this gets tested.