How we estimate, and when we'd tell you to go elsewhere
Level Best uses AI to help price building work. You should want to know how, because "the computer said £84,000" is not a reason to believe anything.
So here is the whole method, including the parts that are weak.
What the AI is actually for
The AI is good at one thing in particular, and it is not arithmetic. It is good at knowing what a job involves.
Describe a rear extension and it knows there is a foundation, a drain to divert, a steel over the opening, a warm roof build-up, a Part P notification, a plasterer who cannot start until the first fix is done, and a set of bifolds with a ten week lead time that will hold up everything if you order them late. That is the expensive knowledge. It is the reason most people's first budget is wrong: not because they priced the roof badly, but because they forgot the scaffold, the skips, the building control fee and the two weeks of waiting.
So we lean on it hardest for scope. What work is in this job, what order does it go in, what gets forgotten.
What we do about the numbers
Left alone, a language model will produce a confident price for anything you ask it. Confidence is not accuracy, and a plausible number is more dangerous than an obviously silly one.
Three things push against that.
Ranges, not numbers. Every line carries a best case and a worst case. A single figure implies a precision nobody has, including quantity surveyors. When the gap between the two is wide, that is information: it usually means there is an unknown worth closing early, and the app says so.
A check against published bands. We keep a table of what UK work actually costs per square metre, taken from published 2026 cost guides. When an estimate is drafted, we divide it by the size and compare. If a single-storey extension comes out at £900/m² when the guides say £1,700 to £3,000, the app tells you, states what the band does and does not include, and suggests you look for what is missing. The AI never supplies those bands. It classifies the job; the table does the checking.
Live prices for the big items. For the handful of lines carrying most of the money, the estimate is checked against current UK price guides at the time it is drafted, rather than whatever the model happens to remember.
What we exclude, and say so
Our estimates are build costs. They assume VAT is included where VAT applies. They do not include professional fees, planning and building control charges, service connections, finance costs, or anything you have not told us about.
Every report the app produces states that in writing, along with the date, the location the rates assume, and the line that matters most: an estimate is not a quotation. It is a well-informed expectation, and the only real price is one with a builder's name on it.
Where we would tell you to go instead
This is the part most software will not tell you.
If you have drawings and you are about to commit serious money, buy an estimate from an expert like a quantity surveyor.
We are not trying to replace it, and anyone telling you an AI subscription is a substitute for a measured estimate on a £250,000 project is selling you something.
So what are we for?
The stretch before that, and the whole job after it.
Before: you have an idea, no drawings, and no sensible way to find out whether it is a £40,000 idea or a £120,000 one. Nobody serves that moment well. The alternative is a free calculator that multiplies your floor area by a single number and tells you nothing about what it left out.
After: the estimate arrives, and then the job actually has to happen. Quotes come in that do not compare with each other. Something costs more than expected and you need to know what that does to the total. A material has a ten week lead time and nobody wrote it down. Money leaves in a different month than you thought. The estimate, however good, is a snapshot. Running the job is the hard part, and it lasts a year.
That is the bit we care about. The estimating has to be honest enough to be worth having. It does not have to beat a quantity surveyor, and we will not pretend it does.
One more thing we are building
The most useful cost data in the country is not in any guide. It is what people actually paid.
Owners using Level Best can share, anonymously and only if they choose to, what their project really cost: the type of job, its size, the region and the rate. Never a name, never an address. It comes back to everyone as an aggregate, and only once enough projects have reported to mean anything.
Published guides are somebody's model. This will be the real thing, and over time it will make every estimate on the platform better than the guides we started with.