Deterministic quantities, not AI-generated ones

There are two jobs inside an automated take-off, and they are not the same kind of job. One is reading: working out what a drawing says, which dimension belongs to which wall, what the rooms are, what materials are called up. The other is calculating: turning those facts into quantities of framing, cladding, sheet goods and fixings. QuoteSmiths uses vision AI for the first and a deterministic rules engine for the second, and the line between them is the most important design decision in the product.

Reading a drawing is a perception problem

Building plans are not a clean data source. They arrive as PDFs exported from different tools, and sometimes as a photograph of a printed sheet. Dimension strings sit at odd angles. Room names are abbreviated. Schedules live on a different page from the plan they describe. Reading that reliably is a pattern recognition task, and pattern recognition is what a vision model is for. It tolerates variation in a way a parser written against a template cannot.

So the model is pointed at reading, and only reading. It extracts what the drawing states, and each figure is shown next to the part of the drawing it came from, so it can be checked against the source rather than taken on trust.

• Dimensions, as they appear on the sheet
• Rooms and their boundaries
• Wall runs, openings and the materials called up
• Schedule entries tied back to the plan they belong to

Calculating a quantity is arithmetic

Once the dimensions are known and confirmed, the maths stops being a judgement call. Studs at set centres over a known wall length is a formula. Sheet count for a floor area at a stated waste factor is a formula. Plates per run, fixings per sheet, bracing per wall type: all formulas, and all of them ones an estimator can write down. Nothing in that step benefits from a model's ability to guess.

QuoteSmiths runs it through a deterministic rules engine instead. Geometry-based formulas apply to the confirmed dimensions, waste factors apply per material, and the result costs out against the workspace's own supplier catalogue and SKUs. Change one wall length and the only lines that move are the ones that rule touches.

The same drawing, confirmed the same way, produces the same bill of materials every time.

The confirmation step is where the boundary sits

Between reading and calculating there is a stop. The operator sees every figure the model read and corrects anything wrong before a single quantity is produced. It is deliberate friction, and it does two things at once.

It puts the check where a person can actually perform it. Someone who knows the job can look at a wall length and tell almost immediately whether it is right. Nobody can look at a fixings count on a finished bill of materials and tell whether it is right, because verifying that means redoing the take-off. Catching a misread dimension at the read step is cheap. Catching it after it has propagated through every line that depends on it is not.

It also keeps responsibility where it belongs. The model proposes a reading. A person accepts it. The rules do the rest, and they are the estimator's own rules, not a black box's opinion about how much timber a wall needs.

Auditability is a commercial requirement

The argument for determinism is not tidiness. It is what happens when a number gets challenged. A builder queries a cladding quantity. A client asks why the sheet count moved between revisions. A variation is disputed once the job is well under way. In each case the useful answer has the same shape: this figure came from this rule, applied to this dimension, with this waste factor, on this version of the plan.

A rules engine can produce that answer because the answer exists. A model that generated the quantity directly cannot, because there is no rule to point at, only an output that sounded plausible. And if the same drawing can produce two different quantities on two runs, the number in the quote stops being a fact about the building and becomes a fact about which run happened to be saved. Back-costing fails the same way: comparing estimate against actual only means something if the estimate would reproduce identically today.

What to ask of any estimating tool

This split is not unique to QuoteSmiths, but it is worth checking for, because the phrase "AI estimating" covers both approaches and the marketing rarely separates them.

• Which step is the model doing: reading the drawing, or producing the quantity?
• Can the operator see and correct what was read before anything is calculated?
• Does the same input produce the same output on every run?
• For any line on the bill of materials, can the tool name the formula and the inputs behind it?
• Are the waste factors and rules visible and editable, or buried?

A tool that answers those cleanly can be argued with, corrected and improved. Estimators already carry their own rules, built out of jobs that priced well and jobs that did not. The software's job is to apply them consistently to a drawing someone has actually checked.