How AI Reads Aerospace Drawings

A Boeing 737 nacelle bracket drawing crossed my desk last month. Twelve sheets. Forty-seven feature control frames. Sixty-three notes. A flag note on Sheet 1 referenced a Boeing process specification that, when I traced it, pulled in four other specifications, two of which had revision-specific requirements that changed the heat treatment cycle depending on which version was called out. The drawing itself was thirty-eight pages. The full spec stack behind it ran to over four hundred.

That is what aerospace drawings actually look like. Not a single sheet with a few tolerances. A document tree, with the drawing as the root, every note as a branch, every spec callout as another branch, every flag note as a cross-reference that can pull a tolerance from a completely different document. Reading a commercial precision part drawing is hard. Reading an aerospace drawing is a different category of problem.

This chapter is what actually happens when AI tries to read these drawings, OEM by OEM. Not a marketing pitch. A genuine breakdown of what works at ninety-six percent confidence, what falls to seventy on Boeing's flag-note system, and where the entire extraction pipeline still needs a human in the loop. If you read Chapter 03 for the general extraction architecture and Chapter 05 for the GD&T deep dive, this chapter is the next layer down: the OEM-specific dialects that make aerospace its own discipline.

Why aerospace drawings are denser by design

A commercial bracket drawing for a kitchen appliance might have eight dimensions, three tolerance callouts, a surface finish symbol, and a title block. An equivalent bracket drawing for a Boeing nacelle has forty dimensions, eighteen GD&T frames, a dozen surface finish callouts referencing different zones, flag notes pointing to three separate process specifications, a critical characteristic table, a material specification with a revision lock, a heat treatment callout, a coating callout, an inspection method specification, a first article inspection requirement, a serialization requirement, traceability requirements, and a flight safety classification.

This density is not bureaucratic excess. Every additional callout maps to a specific failure mode that aerospace has, over decades, learned to specify against. The reason a Boeing drawing calls out BAC5602-style cadmium plating with a specific revision lock is because, somewhere in the 1980s, a supplier used a different cadmium process and a fastener failed. The reason an Airbus drawing references AIPS 02-01-002 for shot peening is because residual stress in fatigue critical parts is, in fact, the difference between a wing lasting thirty thousand cycles or eighty thousand cycles. The drawings are dense because the lessons are written in fatalities.

For an AI extraction system, this means the problem is not just "extract more text." The problem is to understand the document graph. A surface finish callout is not just a Ra value. It is a Ra value, in a specific zone, applied to a specific feature, modified by a specific note, governed by a specific specification, at a specific revision. Miss any one of those and you have extracted the data incorrectly while appearing to have extracted it correctly.

The four OEM dialects you actually encounter

Across the aerospace work that flows through Mavlon, ninety percent of drawings come from four families: Boeing (and its tier-one supply chain), Airbus (and Premium AEROTEC, Stelia, Spirit AeroSystems), Lockheed Martin (and Northrop Grumman), and Pratt & Whitney (with Rolls-Royce, GE, and Raytheon as cousins). Each has its own conventions. An AI system trained only on one will get the others partially wrong, and "partially wrong" on an aerospace drawing is the same as fully wrong from an inspection standpoint.

Below is what actually distinguishes each, what an AI sees when it opens one, and where the failure modes hide.

Layer 1: Boeing drawings and the PVS flag-note system

Boeing's drawing convention is, on the surface, ANSI Y14.5 compliant. Underneath that, it is its own dialect. The most distinctive feature is the Process Variation Specification system - PVS - which lives entirely in the flag notes.

A Boeing drawing will show, in the upper right of Sheet 1, a column of flag notes labeled with numbers in pentagonal symbols. Note 8N, 9N, 11N, 12N. These look like ordinary notes. They are not. Each pentagonal flag note maps to a controlled process variation that the supplier must execute exactly as specified. A 9N flag, for example, might point to "BAC5602 Type II Class 2" cadmium plating with a post-bake at three hundred seventy-five Fahrenheit for twenty-three hours within four hours of plating.

Here is what makes this hard for AI extraction. The flag note appears in two places on the drawing - once in the note column with the full text, and once next to the feature it applies to, shown only as the pentagonal symbol with the number inside. The AI must read both, link them together, understand that the symbol next to a hole means "apply the full text of Note 9N to this specific hole," and then carry that linkage forward into the extracted data.

In our own pipeline, raw symbol recognition for Boeing flag notes sits at about ninety-four percent. The error mode is almost always the same: the pentagonal symbol gets misclassified as a regular circle balloon, or vice versa. We solved most of this with a dedicated symbol classifier trained on Boeing flag notes specifically, but the residual error is concentrated at low DPI scans and on flag notes that overlap with leader lines.

The harder problem is the linkage. When Note 8N says "applies to all holes marked with this flag," and there are seventeen holes on the drawing, the AI must correctly identify every flagged hole. Miss one and that hole gets manufactured without the correct process. Boeing's own engineers, when they review a supplier's first article report, check for exactly this kind of missed flag note linkage. It is a well-known failure mode.

Beyond the flag note system, Boeing drawings have additional conventions that an AI must learn. The critical characteristic table on Sheet 1 lists dimensions that require one hundred percent inspection rather than sampling. The flight safety symbol - a diamond with a hash mark - marks features whose failure could result in loss of aircraft or life. The material specification box references BMS numbers (Boeing Material Specifications) which, when traced, pull in revision-locked metallurgy that the supplier must source from Boeing-approved mills. A correctly extracted Boeing drawing surfaces all of these as structured fields, not as raw text the inspector has to re-read.

Boeing also uses BDS - Boeing Design Standards - for everything from weld symbols to fastener installations. A note that says "Install per BDS-1480" looks innocuous. Behind that callout is a multi-page standard with specific torque values, sealant requirements, fastener prep, and inspection criteria. An AI system that extracts the text "BDS-1480" without flagging that this is a callout to a controlled installation specification has missed half the information.

Layer 2: Airbus drawings and the AIPS specification stack

Airbus drawings are visually different from Boeing drawings in ways that, at first glance, look like style choices. They are not. The differences encode an entirely different specification system.

Where Boeing uses BAC and BMS callouts, Airbus uses AIPS (Airbus Process Specifications) and AIMS (Airbus Material Specifications). Where Boeing uses pentagonal flag notes, Airbus uses square-framed notes. Where Boeing's title block sits in the lower right, Airbus's title block sits in the lower right but with a fundamentally different field layout, including dedicated fields for design organization approval and production organization approval that map to EASA Part 21 requirements.

The hardest part of Airbus extraction is the AIPS dependency tree. An AIPS callout - say, AIPS 02-01-003 for chemical milling - does not stand alone. It references AIMS material specs, which reference AITM (Airbus Industrial Test Methods) procedures, which reference controlled chemical baths, which reference qualified processor lists. A supplier reading the drawing must trace that entire chain to know what they are actually agreeing to manufacture. An AI system that only extracts the surface-level AIPS number has missed the downstream cascade.

In our pipeline, AIPS callout extraction sits at about ninety-three percent accuracy, but the structured downstream resolution - the ability to say "this AIPS pulls in these AIMS, these AITM, and therefore the supplier needs Nadcap certification in chemical processing" - is closer to seventy-eight percent. The remaining twenty-two percent is concentrated in drawings with revision-locked callouts (AIPS 02-01-003 Rev D specifically, not just AIPS 02-01-003) where the revision number is in a separate field on a separate sheet.

Airbus also uses a tolerance system that differs from Boeing's. Where Boeing relies heavily on GD&T per ASME Y14.5, Airbus uses ISO 1101 GD&T conventions, which look similar but differ in subtle ways: the position of modifiers in the feature control frame, the treatment of independency, the use of all-around versus all-over symbols. An AI trained primarily on ASME conventions will misread ISO conventions about three percent of the time, and that three percent is concentrated in exactly the kind of composite frames that matter most for tolerance stack-up.

On top of all this, Airbus drawings tend to be multilingual. French-language notes appear alongside English notes, sometimes with German notes on parts produced in Germany. The AI must extract all three correctly and recognize that they are translations of the same content, not three separate requirements. We solved this by running a translation alignment pass that catches notes appearing in multiple languages and merging them into a single structured requirement. The pass adds about four hundred milliseconds to extraction time per drawing.

Layer 3: Lockheed Martin and the FRM/PDM convention

Lockheed Martin drawings - particularly those from the Aeronautics and Missiles and Fire Control divisions - use a convention rooted in FRMs (Feature Requirement Manuals) and PDMs (Process Definition Manuals). Northrop Grumman uses a similar system. The visual appearance of these drawings is closer to Boeing than Airbus, but the underlying specification system is its own thing.

The distinctive feature of Lockheed drawings is the heavy use of controlled key characteristics - labeled KC1, KC2, KC3 in the feature control frame area. These map to specific Statistical Process Control requirements that the supplier must execute. A KC1 characteristic, for example, requires Cpk greater than one point three three with one hundred percent inspection until process capability is demonstrated. A KC2 requires Cpk greater than one point zero zero with sampling per AS9100 procedures. A KC3 is a flagged-but-non-statistical characteristic.

An AI extracting a Lockheed drawing must not only identify the KC callout but understand the cascading inspection requirement that follows. A KC1 hole position tolerance is not just a position tolerance. It is a position tolerance with a CMM inspection requirement, a sample-size requirement of one hundred percent until thirty consecutive parts demonstrate capability, a documented SPC plan requirement, and a first article report field where the Cpk number must be reported. Miss the KC1 and you have extracted a dimension. Catch it and you have extracted a manufacturing contract.

Lockheed drawings also use a flag note system similar to Boeing's but with different symbols - typically hexagonal rather than pentagonal. The dimensional and tolerance conventions follow ASME Y14.5, so the GD&T extraction logic from Chapter 05 applies. But the program-specific manuals - F-35 has its own, F-22 had its own, the missile programs each have their own - add layers of requirements that show up as specification callouts in the notes column.

In Mavlon's pipeline, Lockheed KC callouts extract at about ninety-five percent. The error mode is similar to Boeing's flag notes: the KC symbol gets misclassified or its linkage to the specific feature gets lost. We have not seen the AIPS-style downstream cascade problem on Lockheed drawings to the same degree, because the FRMs tend to be self-contained rather than referencing other manuals.

Layer 4: Pratt & Whitney engine drawings

Engine drawings are their own world. The geometry is rotational - most parts are turbine blades, vanes, disks, cases, seals, or housings with complex internal cooling passages. The materials are exotic - single-crystal superalloys, ceramic matrix composites, titanium aluminides. The tolerances are tight in absolute terms but also tight relative to the dimensional scale, which means feature control frames stack up in ways that surface-mount drawings rarely do.

Pratt & Whitney uses PWA (Pratt & Whitney Aircraft) specifications throughout. A PWA-1480 callout points to a specific single-crystal nickel alloy with controlled grain orientation. A PWA process specification might call out an electron-beam welding procedure with controlled atmosphere, controlled cooling, and post-weld heat treatment within strict time windows. An AI system reading a PWA drawing must extract not just the PWA number but the revision, the class, the type, and any callout-specific notes that modify the base specification.

The hardest extraction problem on engine drawings is not the text. It is the section views. A turbine blade drawing typically has six to twelve section cuts showing internal cooling passage geometry. Each section is labeled (A-A, B-B, C-C) and each has its own set of dimensions, GD&T frames, and surface finish callouts. The AI must associate the section dimensions with the correct feature on the main view, which requires reading the section labels, locating the section cut on the main view, and understanding the geometric correspondence.

Section view association is where engine drawing extraction currently struggles most. Our raw section identification sits at ninety-one percent. Dimension-to-section linkage drops to about eighty-three. Below that, the correct association of a tolerance on a section view to the specific feature on the main view - the information a CMM programmer actually needs - sits around seventy eight. That gap is where human review still adds the most value on engine drawings.

Engine drawings also use a specialized notation for cooling-hole patterns. A blade might have one hundred forty cooling holes, each with a specific angle, depth, and breakout location. The drawing does not show one hundred forty individual holes. It shows a few representative holes with a table or a pattern callout that defines the rest. An AI must read the pattern definition and reconstruct the full hole population, which means understanding rotational arrays, helical arrays, and conditional patterns ("eighteen holes equally spaced except where interfered by the fillet, then offset by five degrees").

The cross-cutting problem: specifications behind the drawing

Every aerospace OEM has its own specification library, and every aerospace drawing pulls from that library. An AI system that extracts the drawing without understanding the specifications has extracted half the information.

Consider a Boeing drawing that calls out "BAC5602 Type II Class 2." The drawing tells you cadmium plating is required. The specification tells you the bath chemistry, the current density, the thickness range, the hydrogen embrittlement relief bake, the post-bake hold time, the inspection methods, and the acceptance criteria. The supplier cannot quote the part without knowing whether they have a Nadcap-certified plating cell that can hit the BAC5602 Type II Class 2 requirements. The estimator cannot price the part without knowing the plating cost. The quality engineer cannot plan the inspection without knowing the BAC5602 inspection methods.

Mavlon's pipeline ingests the major aerospace specification libraries - BAC, BMS, BPS for Boeing; AIPS, AIMS, AITM for Airbus; PWA for Pratt & Whitney; PS, MS for Lockheed; and the cross-industry SAE, AMS, MIL-STD specifications that all OEMs share. When the AI extracts a specification callout, it links to the controlled document and surfaces the relevant requirements as structured fields on the quote.

Specification library coverage is one of the largest moats in aerospace drawing AI. A new entrant can build a generic extractor in a few months. Building the specification graph - accurately cross-referenced, revision-locked, kept current as OEMs update their libraries - takes years. This is one of the reasons general drawing AI tools struggle with aerospace and why aerospace-focused tools (Mavlon included) invest disproportionately in the specification layer.

Honest accuracy table, by OEM

Here is what we actually see in our own pipeline. These numbers are measured against expert-annotated test sets of fifty to one hundred drawings per OEM. They are not generalized - they reflect Mavlon's current state in May 2026 - but they give an honest sense of where AI drawing extraction stands on aerospace work today.

For Boeing drawings, raw text and dimension extraction sits at ninety-six percent. GD&T frame parsing at ninety-three percent. Flag note symbol recognition at ninety-four percent. Flag note linkage to the correct feature at eighty-six percent. Specification callout extraction at ninety-five percent. Specification graph resolution (the full downstream dependency chain) at eighty-two percent. Composite extraction quality - meaning a drawing extracted end-to-end with all linkages correct - at seventy-eight percent.

For Airbus drawings, raw text and dimension extraction at ninety four percent. GD&T per ISO 1101 at ninety-one percent. AIPS callout extraction at ninety-three percent. AIPS graph resolution at seventy-eight percent. Multilingual note alignment at ninety-seven percent. Composite extraction quality at seventy-four percent.

For Lockheed and Northrop drawings, raw text and dimension extraction at ninety-five percent. GD&T parsing at ninety-three percent. KC characteristic identification at ninety-five percent. KC-to-SPC requirement linkage at eighty-nine percent. Composite extraction quality at eighty-one percent.

For Pratt & Whitney engine drawings, raw text extraction at ninety-three percent. PWA spec callout at ninety-four percent. Section view identification at ninety-one percent. Section-to-feature linkage at eighty-three percent. Cooling pattern reconstruction at seventy-six percent. Composite extraction quality at seventy one percent.

Engine drawings sit at the bottom of this table because section view association and cooling pattern reconstruction are the two hardest residual problems in aerospace drawing AI. The composite number is the one that matters for shops: it represents the percentage of drawings where an AI extraction would pass a downstream quality review without rework.

Where AI struggles with aerospace specifically

Five places, in order of how often they cause real problems.

First, low-DPI legacy scans. A 1978 drawing scanned at one hundred fifty DPI in 2003 and re-released in 2019 looks like a degraded photocopy of a photocopy. The optical resolution is below what modern OCR needs. Flag note symbols become illegible. Tolerance values drop digits. We see this most often on F-15, F-16, and legacy 737 derivatives that have been in production for decades. The fix is not more AI; it is asking the supplier to request a higher-resolution drawing release from the OEM.

Second, hand-redlined revisions. An OEM engineer makes a change to a released drawing by drawing a circle around the old value and writing the new value next to it. The redline appears on the drawing but the title block revision letter has not yet been updated. The AI does not know whether to honor the printed value or the handwritten redline. Our pipeline flags every detected redline for human confirmation. We have not found a confident automated solution for this.

Third, sheet-spanning notes. A note on Sheet 4 that says "Applies to all features on Sheet 1, 3, and 5" requires the AI to maintain cross-sheet awareness. We solved most of this by treating the entire drawing package as a single document with cross-references, but very long packages (twenty or more sheets) still occasionally lose linkage on rarely-applied notes.

Fourth, specification revision conflicts. The drawing calls out "BAC5602 Type II Class 2." The drawing was released in 2018 against BAC5602 Revision M. The current BAC5602 is Revision Q. The supplier needs to know which revision applies. The standard answer is "whatever was current at drawing release," but some OEMs explicitly require current revision, and some require the revision specifically called out on the drawing. The AI cannot resolve this without external context. We surface the conflict and ask the supplier to confirm.

Fifth, security-classified or partially redacted drawings. Some defense work arrives with portions of the drawing redacted - coordinates blacked out, certain features hidden. The AI must extract what is visible without hallucinating what is hidden. Modern vision-language models are remarkably good at not confabulating in this case, but we still add a redaction-detection pass that flags any black-rectangle regions for human attention.

Why this matters for AS9102 Form 3

Every aerospace part requires a First Article Inspection report. For dimensional and tolerance characteristics, that is the AS9102 Form 3 - a balloon-by-balloon, characteristic-by-characteristic record of what was specified versus what was measured. A typical aerospace bracket FAI has eighty to two hundred characteristics. A typical engine component FAI has three hundred to eight hundred.

Filling out Form 3 manually takes between three and twelve hours depending on drawing density. The bottleneck is not measurement - the CMM produces measurements quickly. The bottleneck is transcription. An inspector reads a dimension off the drawing, finds the corresponding bubble number, writes the spec value, the measured value, the tolerance band, the result. Repeat eighty times. Repeat three hundred times for an engine part. Errors accumulate. Studies of first-pass FAI submission rates put the rejection rate at around fifteen to twenty percent in commercial aerospace and higher in defense.

When AI extraction works correctly on an aerospace drawing, the Form 3 fills itself. Bubble numbers map to characteristics. Characteristics carry their full GD&T, datum references, flag note linkages, KC classifications, and specification callouts. The inspector measures, the system records, the form populates. First-pass acceptance rates climb from eighty percent toward ninety-five.

This is why aerospace drawing AI is not, in the end, about drawing AI. It is about FAI acceleration, supplier scorecard improvement, and the long compounding effect described in Chapter 04. The AI that reads the drawing is the foundation. The AS9102 Form 3 autopopulation is the visible win. The compounding intelligence across a shop's historical FAI library is the moat.

The next twelve months

Three things will change in aerospace drawing AI between now and mid-2027.

First, the specification graph will become a commodity. Today, building the BAC-BMS-BPS-AIPS-AIMS-PWA-MIL-STD cross-reference library is one of the largest hidden moats. Within twelve months, OEM-published specifications will be ingested by every serious aerospace AI vendor and the graph will be table-stakes. The differentiator will move to revision tracking, version reconciliation, and the supplier-specific approved processor list.

Second, multimodal models will close most of the section-view gap on engine drawings. The hardest current problem - correctly associating a tolerance on Section A-A with the specific feature on the main view - is exactly the kind of spatial reasoning task that next-generation vision models are improving on quarterly. We expect engine drawing composite accuracy to climb from seventy-one toward eighty-five within a year.

Third, redline handling will be automated for OEMs that publish digital revision packages. Boeing and Airbus are both moving toward fully digital MBD (Model-Based Definition) releases, in which redlines are replaced with controlled revision artifacts that AI can read directly. For the next three to five years, though, the hybrid world of paper-style drawings with embedded redlines will persist, and human-in-the-loop redline review will remain necessary.

Everything else - the basic OEM dialect recognition, the specification linkage, the GD&T parsing, the flag note association - is already at production quality for the majority of drawings. What separates AI vendors today is not whether the extraction works on a clean Boeing drawing. It is how the extraction degrades on the hard fifteen percent: the legacy scans, the engine sections, the multilingual Airbus packages, the revision conflicts. That degradation curve is where shops should evaluate vendors.

What to ask any aerospace drawing AI vendor

If you are evaluating drawing AI for an aerospace shop, the questions that actually matter are not the ones in the marketing deck. The questions that matter are the ones below.

Ask: "Show me your composite extraction accuracy on a sample of my own drawings, not yours." A vendor that has trained on Boeing will excel on Boeing. The question is whether they degrade gracefully on Airbus, Lockheed, and Pratt & Whitney.

Ask: "What is your flag note linkage accuracy?" Raw symbol recognition is easy. Linking a flag note to the specific feature it applies to is hard. Ask for the number, not the demo.

Ask: "Which specification libraries do you maintain, and at what revision currency?" If the vendor cannot tell you whether they are tracking BAC5602 at Revision Q or Revision M, they are not maintaining the spec graph. They are extracting strings.

Ask: "How do you handle redlines, and how do you handle revision-locked specification callouts?" If the answer is "the AI handles it," walk away. The correct answer is "we surface the conflict to a human reviewer with the relevant context."

Ask: "What is your section-view-to-feature linkage accuracy on engine drawings?" If the vendor has not measured this, they are not a serious aerospace drawing AI vendor. If the number is below seventy-five percent, expect significant human review on engine work.

Aerospace drawing AI is not, in 2026, a solved problem. It is further along than it was in 2024, and dramatically further along than it was in 2022. The shops that integrate honestly-evaluated drawing AI today - with full awareness of where it works and where it does not - will compound an advantage that, by the time the residual problems are solved, will be hard for laggards to catch.