Wonder why digital takeoff on a raster PDF is tough and near-impossible on a scanned PDF compared to a 3D file?
Why AI can read a 3D file instantly
First, let's look at what information is encoded in 3D drawings and what happens to all that data by the time it's converted into a PDF.
A 3D fabrication package (CAD, STEP or DFX) is expected to have tolerances, revision blocks, title block data, BOM tables, section views, detail views, and notes.
There's a lot of information packed into a 3D file. On top of that, these files have special features that software programs can quickly access to extract details:
1. Every point has coordinates
Every face, edge and corner of the 3D object has its own coordinates (X, Y, Z). This is what's called parametric data, and it's why an object (whether it's a bracket, beam or screw) isn't drawn as a frozen 3D shape. Instead it's defined by mapping how its boundaries connect.
It's easy for AI to analyse the part because it simply queries the database to read the point coordinates and runs the relevant maths equations to answer questions like:
- What is the diameter of the knockout?
- What is the metal thickness?
- What is the bracket angle?
2. The parts come with labels
You squint, tilt your head sideways while wondering "what is it??" Thank goodness the drawing has a title so you can figure it out.
That's what semantic data is, information that tells you the meaning of an object. This could include important information like measurement values, tolerance range, part numbers, materials, finishes or supplier information.
This information is saved as metadata tags that give the drawing context and further manufacturing requirements not directly related to the size or dimension of the object.
3. It knows how the components fit together
Multi-part assemblies with subcomponents have their relationships described by two parallel data trees that link to each other.
- Product Tree: maps the hierarchy of the parts, describing all the different items that make up the final assembly.
- Shape Tree: outlines the geometry of each component and its placement in space in relation to the others.
Why PDFs require analysis, not just data retrieval
With all of this information included in a 3D file, it's fairly straightforward to extract this data for quoting. But when this same drawing is downgraded to a PDF, computer-aided takeoff gets a lot harder.
All of this detail, context and intelligence is stripped away. A vector PDF is just a collection of flat coordinate paths and text blocks. A scanned PDF is even worse, literally a grid of colour values.
On top of this, estimators frequently have to contend with:
- Missing BOMs: no parts list at all, just scattered geometry and a few notes.
- Scattered dimensions: important information is split between views, pages and components.
- Sub-assembly roll-up: multi-part jobs have to be read bottom-up to check completeness and buildability.
- Symbol detection: shorthand symbols can be misread or illegible, causing inaccurate quotes.
Historically, estimators were left with two choices to try and deal with this.
- Print out the drawing and manually determine requirements, which is time consuming and error-prone, or
- Reverse engineer the 3D model to digitally create the bill of materials, which requires expensive labour to do up-front engineering work.
Here's the AI way
Under the hood, AI takeoff operates very closely to how a human estimator works. It uses computer vision to read the measurements and notations, in minutes instead of hours. Because there's no data tree to query, it needs to look at different section views to find details or infer relationships.
It works methodically, cross-checking as it goes, and flagging anything uncertain for a second pair of human eyes to review before a quote goes out. It calibrates to how your shop's drawings and jobs actually work, then reads new ones the same way.
It then turns to historical data to create a list of the manufacturing steps used in similar past quotes, pulling together a detailed bill of materials and rough labour hours.
That's the theory. Here's what it looks like in practice 👇