If you tried OCR on an engineering drawing any time before 2022, you learned one thing: it does not work. The software would spit out a mess of random characters, misidentified numbers, and completely garbled tolerance symbols. It was, to be blunt, useless for anything beyond the simplest line drawings.
That experience left a scar on the manufacturing industry. When I tell fabricators that AI can now read their 2D technical drawings, the most common response is a polite version of "I will believe it when I see it."
Fair enough. So this chapter is not going to ask you to believe anything. Instead, I am going to explain, in genuine technical detail, exactly what happens when a modern AI system opens a 2D PDF engineering drawing. Step by step. Including where it works well, where it struggles, and why the accuracy numbers mean what they mean.
By the end, you will understand this technology well enough to evaluate it yourself.
First: Why Old OCR Failed
To understand why modern AI works, you need to understand why OCR did not.
Traditional Optical Character Recognition (OCR) was designed to read printed text. Scanned letters, invoices, typed documents. It works by detecting individual characters, matching them against a library of known characters, and stringing them together into text.
Engineering drawings broke OCR in multiple ways:
An engineering drawing is not text. It is a mixture of text, geometric shapes, symbols, lines, arrows, hatching, and spatial relationships. OCR sees all of these as potential "characters" and tries to make sense of them.
Context does not exist. OCR does not know that the number "150.00" sitting between two arrow lines is a dimension. It does not know that the same number in a title block is a weight. It does not know that "Ra 3.2" means surface roughness. It processes every text string identically, without any understanding of what it means or where it sits on the page.
Symbols are not characters. GD&T symbols (flatness, parallelism, position, runout) are geometric shapes, not text characters. OCR either ignores them entirely or misinterprets them as random characters. The position symbol (⊕) might become a "t" or a "+" or nothing at all.
Spatial relationships are invisible. In a drawing, the meaning of a number depends on its position relative to other elements. "25.4" near a hole is a diameter. The same "25.4" along a surface is a length. OCR has no concept of "near a hole" or "along a surface." It just reads text.
Rotated text, leaders, and overlapping annotations. Dimension text in drawings is frequently rotated to align with the dimension line. Leaders (lines with arrows pointing to features) connect text to geometry. OCR struggles with all of this because it expects horizontal text in neat rows.
The fundamental problem was that OCR treats a drawing as if it were a page of text that happens to have some pictures on it. In reality, a drawing is a visual language with its own grammar, syntax, and semantics. Reading it requires understanding the language, not just recognizing the characters.
What Changed: Vision Language Models
The technology that made AI drawing reading possible is called a vision language model (VLM). It emerged from the convergence of two AI capabilities: computer vision (understanding images) and large language models (understanding text and context).
Here is the important difference: a VLM does not read a drawing character by character. It looks at the drawing the way a human does. It processes the entire page as a visual scene, identifying objects, relationships, and meaning simultaneously.
Think about how you read a road sign. You do not process it letter by letter: S... T... O... P. You see the red octagon and instantly understand "stop." The shape, color, and context all contribute to meaning simultaneously.
That is closer to how a VLM reads a drawing. It does not process individual characters and then try to figure out what they mean. It perceives the visual structure, the dimension lines and extension lines, the leader arrows, the tolerance frames, the section markers, and it understands what role each element plays in the drawing's language.
Let me walk through the actual processing pipeline. When a modern AI system opens your engineering drawing, it goes through four distinct layers of analysis.
Layer 1: Visual Segmentation
Breaking the page into regions
The first thing the system does is divide the drawing page into meaningful regions. Not by reading anything, just by looking at the visual structure.
It identifies: the title block (usually a structured grid in the bottom right). The main drawing views (the largest geometric content on the page). Detail views and section views (smaller bounded areas, often labeled). The notes block (text heavy region, usually top left or along the border). Revision tables. Bill of materials. Empty space and borders.
This is important because different regions require different processing strategies. The title block has structured fields with predictable layouts. The main drawing has geometry mixed with annotations. The notes block is primarily text. By segmenting first, the system can apply the right analysis to each region.
This step is fast (fractions of a second) and highly accurate (95%+ for standard drawing layouts). The system recognizes standard drawing formats from ISO, ASME, DIN, and most company specific templates it has been trained on.
Layer 2: Object Detection
Identifying what each element is
Within each region, the system now identifies individual objects and classifies them. This is where it starts to understand what it is looking at.
Dimension annotations. The system recognizes the combination of extension lines, dimension lines, and text as a dimension callout. It identifies the type: linear, angular, radial, or diameter. It reads the dimension value and associates it with the specific feature it measures.
Tolerance annotations. When a dimension has a tolerance attached (plus/minus values, limit dimensions, or a general tolerance note), the system identifies and extracts both the nominal dimension and the tolerance specification.
GD&T frames. The system detects the rectangular feature control frame and parses its contents: the geometric characteristic symbol, the tolerance value, any material condition modifiers (MMC, LMC, RFS), and the datum references. This is one of the hardest parsing tasks because GD&T frames pack dense information into a small visual space.
Surface finish symbols. The check mark or fork symbol with associated Ra/Rz values and any additional parameters.
Weld symbols. Arrow, reference line, tail, and the various elements that specify weld type, size, contour, process, and testing requirements.
Section and detail markers. Letters, numbers, and cutting plane lines that link views together.
Leaders and callouts. Lines with arrows pointing to features, with associated text containing notes, part numbers, or specifications.
Each detected object gets classified (what type of annotation is this?) and its content gets extracted (what does it say?). The system processes all objects on a page simultaneously, not sequentially.
Layer 3: Context Understanding
Understanding what each element means
This is the layer that separates modern AI from old OCR. And it is the hardest to explain, because it involves the kind of reasoning that feels intuitively obvious to a human but is extraordinarily difficult for a machine.
Consider this: the text "150.00" appears in three places on a drawing page. In one location, it is a linear dimension between two extension lines. In another, it is a weight value in the title block. In a third, it is part of a note: "minimum batch size 150."
Old OCR would extract "150.00" three times and have no idea what any of them meant. The VLM understands all three because it processes them in context:
The first "150.00" sits between extension lines connected to a geometric feature, with a dimension line and arrows. Context: this is a dimension with a value of 150mm.
The second "150.00" sits in a title block field labeled "Weight" or "Masse." Context: this is the part weight in grams or kilograms.
The third "150.00" is part of a natural language sentence in the notes block. Context: this is a production parameter, not a geometric measurement.
The system resolves these ambiguities by combining visual context (where is the text on the page, what graphical elements surround it) with semantic context (what does this text mean given the type of region it is in).
This context understanding also handles the multilingual challenge. A drawing with dimension text in the views but notes in French and title block fields in German presents no special problem, because the system understands the visual language of the drawing independently of the written language of the annotations.
Layer 4: Relationship Mapping
Connecting everything together
The final layer connects individual extractions into a coherent data model. This means understanding that:
A specific tolerance callout applies to a specific dimension on a specific feature.
A GD&T frame references specific datums, which are themselves features identified elsewhere on the drawing.
A surface finish symbol on View A corresponds to the same surface shown from a different angle in Section B–B.
A note saying "all dimensions in mm unless otherwise stated" modifies the interpretation of every dimension on the drawing.
A detail view at 5:1 scale shows the same feature as the main view, just magnified. Dimensions in the detail view should be consistent with the main view.
This relationship mapping produces a structured output that is not just a flat list of extracted values, but a connected data model where each data point has context: what it is, where it came from, what it applies to, and how it relates to other data points.
What This Looks Like in Practice
Let me make this concrete.
Imagine a 7 page drawing for a machined aluminum bracket. Material: EN AW 7075 T6. General tolerance: ISO 2768 mK. The drawing has a main orthographic view, two cross sections, one detail view at 5:1 scale, and two pages of notes and referenced specifications.
What the AI extracts in minutes:
From the title block: part number, revision level, material spec (EN AW 7075 T6), weight (0.34 kg), general tolerance class (ISO 2768 mK), surface finish default (Ra 3.2), projection method (first angle), and drawing standard.
From the views: 87 dimensions (linear, angular, radial, and diameters), each associated with their specific feature. 12 specific tolerance callouts overriding the general tolerance. 8 GD&T frames including 3 positional tolerances, 2 flatness callouts, 2 parallelism callouts, and 1 total runout. 6 surface finish callouts (4 at Ra 3.2, 1 at Ra 1.6 on a sealing surface, 1 at Ra 0.8 on a bearing bore).
From the notes: heat treatment requirement (T6 condition), edge break requirement (0.3 to 0.5mm all sharp edges), material certification requirement (EN 10204 3.1), marking requirement (part number electro etched), and packaging specification.
From referenced standards: ISO 2768 mK, EN 10204, ISO 1302 (surface finish), EN 573 (aluminum alloy designation).
Total: approximately 180 discrete data points, structured, categorized, and connected.
An experienced estimator doing this manually would take 60 to 90 minutes and would typically capture 60% to 70% of these data points, focusing on the ones most relevant to pricing.
Honest Accuracy Numbers
I am not going to give you a single accuracy number and call it a day. Accuracy varies significantly by extraction type, drawing quality, and complexity. Here are real numbers from production use on actual fabrication shop drawings.
| Extraction Type | Accuracy | Notes |
|---|---|---|
| Linear dimensions | 90% to 96% | Highest on clean digital PDFs, lower on scanned documents |
| Angular & radial dimensions | 85% to 92% | Slightly lower due to symbol recognition (degree, radius, diameter) |
| General tolerance identification | 92% to 97% | The standard class (mK, fH, etc.) is almost always read correctly |
| Specific bilateral tolerances | 85% to 93% | Plus/minus format is well handled |
| GD&T: simple frames | 80% to 90% | Single characteristic, single datum reference |
| GD&T: complex frames | 65% to 80% | Multiple datums, modifiers, composite frames |
| Surface finish (Ra values) | 85% to 92% | Symbol recognition + value extraction |
| Material spec from title block | 93% to 98% | Structured fields are easiest to read |
| Title block fields | 90% to 96% | Part number, revision, weight |
| Notes (English) | 90% to 95% | Standard manufacturing notes |
| Notes (multilingual) | 85% to 92% | Technical vocabulary in FR, DE, IT |
| Referenced standards | 88% to 95% | ISO, ASME, EN, MIL spec numbers |
| Weld symbols | 75% to 85% | Complex weld symbols with multiple elements are harder |
The weighted average across all extraction types on a typical fabrication drawing is approximately 85% to 90%.
What "85% accuracy" actually means in practice
This number confuses people, so let me unpack it.
On a drawing with 180 data points at 85% accuracy, approximately 153 data points are extracted correctly and 27 are either missed, misread, or partially wrong.
Those 27 errors are not randomly distributed. They cluster in predictable areas:
Complex GD&T frames (composite frames with multiple tolerance zones, multiple datum references with modifiers). These account for a disproportionate share of errors because they are the most visually dense and semantically complex annotations on the drawing.
Overlapping or crowded annotations. When dimension text overlaps with other dimension text or sits on top of geometric lines, extraction accuracy drops.
Non standard layouts. Drawings that deviate significantly from ISO or ASME conventions (unusual title block formats, non standard annotation styles) produce more errors.
Low quality source documents. Scanned paper drawings, especially older ones with fading or bleed through, are harder to process than clean digital PDFs.
The key insight is that your estimator reviews the extraction, not recreates it. Reviewing 180 pre extracted data points and correcting 27 errors takes approximately 8 to 15 minutes. Reading the same drawing from scratch and extracting 180 data points manually takes 60 to 90 minutes.
85% accuracy does not mean 15% of the value is lost. It means 85% of the work is eliminated and the remaining 15% is a focused correction task instead of a from scratch reading task. That is a 6x to 8x productivity improvement on the extraction phase alone.
Where AI Still Struggles (And Probably Will for a While)
I want to be honest about the current limitations because overselling this technology destroys trust.
Handwritten annotations
Customers sometimes annotate drawings by hand: circling a dimension, writing "CRITICAL" next to a feature, adding notes in pen. Printed text on a digital drawing? High accuracy. Handwriting on a scanned copy? Much lower. The variance in handwriting styles makes this fundamentally harder than printed text recognition.
Implicit requirements
Some requirements are not written anywhere on the drawing. They are implied by standards, by customer practice, or by industry convention. "If the material is 7075 T6 and the part goes into an aerospace assembly, anodize per MIL-A-8625." That inference requires domain knowledge that current AI systems handle inconsistently.
Drawing interpretation
The AI reads what is on the drawing. It does not (yet) interpret what the designer intended but did not explicitly state. An experienced estimator looks at a drawing and thinks "this tolerance does not make sense here, the designer probably meant it to apply to the other surface." That kind of interpretive judgment is beyond current AI.
Very old or degraded drawings
Scanned blueprints from the 1980s with poor contrast, fold lines, and handwritten modifications. The system can extract some data, but accuracy drops significantly. Some of these drawings are genuinely hard even for human readers.
Assembly relationships
When a drawing package includes an assembly view with multiple parts, understanding which dimensions belong to which part, and how parts relate to each other geometrically, is significantly harder than reading a single part drawing.
If a drawing was created digitally (CAD to PDF) within the last 10 years and follows ISO or ASME conventions, expect 85% to 95% extraction accuracy. If it was scanned from paper, expect 70% to 85%. If it includes heavy hand annotations, expect 60% to 75%. In all cases, human review is required. The difference is whether that review takes 10 minutes or 90.
The Review Step: Why It Matters
No serious AI drawing extraction system should operate without human review. This is not a limitation to apologize for. It is a design decision.
The human review step serves three purposes:
Error correction. The estimator scans the extraction table against the drawing. Errors jump out quickly because the estimator is looking at structured data next to the source document, not building the structure from scratch. This is the difference between proofreading and writing.
Judgment overlay. The AI extracts what is on the drawing. The estimator adds what they know about the drawing. "This tolerance is unusually tight for this application. The customer probably does not need it. I should call them." That judgment cannot be automated.
Confidence calibration. Over time, the estimator develops a sense of where the AI is reliable and where it is not. They learn to pay extra attention to GD&T frames and weld symbols. They learn to trust the title block extraction. This calibration makes the review process faster with each drawing.
The goal is not to remove the human. The goal is to change what the human does. Instead of 90 minutes of reading and typing, they spend 10 minutes of reviewing and judging. The skill level required is the same. The tedium level is dramatically lower.
The Email Layer
I have focused on drawing extraction because it is the most technically challenging part. But a complete RFQ parsing system also reads the email itself.
Email parsing is a significantly easier problem for AI. Emails are primarily text, structured in recognizable patterns (greeting, body, signature), and contain named entities (companies, people, dates, quantities) that language models are very good at extracting.
Accuracy on email parsing runs 95%+ for structured data (customer name, quantities, delivery dates, material specs, quality standards). It handles multilingual emails (English, French, German, Italian) with minimal accuracy degradation because the technical vocabulary is similar across languages.
The combination of high accuracy email parsing with 85%+ drawing extraction gives the system enough data to deliver genuinely useful structured output. The email provides the what (customer, quantity, deadline) and the drawing provides the how (geometry, tolerances, materials). Together, they give the estimator everything needed to start pricing.
How This Technology Will Evolve
I want to end this chapter with an honest forward look, because understanding where the technology is heading helps you make better investment decisions.
Accuracy will improve steadily, not dramatically. Going from 85% to 90% is achievable in the near term. Going from 90% to 95% is harder. Going from 95% to 99% is very hard. The diminishing returns curve is real. But each percentage point of accuracy improvement reduces review time meaningfully.
Complex GD&T parsing will improve the most. This is currently the weakest area and the one receiving the most research attention. Expect meaningful improvements in composite frame parsing and datum reference resolution within the next 12 to 18 months.
Domain specific training will matter. A system trained on thousands of sheet metal fabrication drawings will outperform a general purpose system on sheet metal drawings. Specialization drives accuracy. Expect vertical specific models (sheet metal, machining, welding, aerospace) to emerge.
The review loop will create a flywheel. Every time an estimator corrects an extraction error, that correction can be used to improve the system. Shops that process high volumes of drawings will see accuracy improvements faster because their correction data feeds model improvement. This is the compound effect we explore in Chapter 04.
The technology is not perfect. But it is good enough to transform your front office today. And it is getting better every month.
Stop reading.
Start seeing.
Drop in your hardest drawing - aerospace, sheet metal, precision machining. Watch every dimension, GD&T frame, and tolerance get extracted in minutes.
Upload a Drawing