AI for Design Quality & DFM
How AI first article inspection automates drawing ballooning, GD and T extraction, and AS9102 report prep so quality teams ship FAIs faster and cleaner.
·
⏱
7 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
First article inspection will not disappear, and it should not. It is the gate that proves a process can make a part to its definition before production scales. What AI changes is the cost of preparing the report. Ballooning, dimension capture, and GD and T extraction are the slow, error prone parts of FAI, and they are precisely the parts that drawing understanding can accelerate, taking a job measured in hours down to minutes while leaving the hard judgment calls to a qualified inspector.
The realistic posture is assisted, not autonomous. Let the software read the drawing, draft Form 3, and keep every balloon traceable to its row, then have a person verify the composite frames, the handwritten notes, and the final results. Done that way, AI first article inspection makes AS9102 compliance faster without trading away the rigor that makes the standard worth following.
Ask any quality engineer in aerospace or defense about first article inspection and you will hear the same complaint. The hard part is rarely measuring the part. The hard part is the paperwork that surrounds it: ballooning every dimension on the drawing, transcribing each tolerance and geometric callout into a report, and proving that nothing was missed. A single complex drawing can carry hundreds of characteristics, and each one has to be numbered, captured, and reconciled by hand.
First article inspection, often shortened to FAI, is the formal check that a new or changed part actually matches its engineering definition before full production begins. In aerospace the report format is governed by the SAE AS9102 standard. The work is exacting and slow, which is exactly why AI first article inspection has become one of the most practical places to apply drawing understanding. The same reading of a drawing that a machinist or inspector does in their head can now be assisted by software that extracts dimensions, recognizes GD and T symbols, and drafts the report.
This article walks through how that automation works, where it helps, and where a human inspector still has to stay in the loop.
What first article inspection actually requires
FAI exists to confirm that a production process can make a part to its drawing before that process scales. According to SAE, the AS9102 standard defines the documentation that aerospace suppliers use to record this verification. The report is built around three forms, each with a distinct job.
Form 1, Part Number Accountability, identifies the part or assembly under inspection along with its drawing revisions, serial numbers, and the parts list for an assembly.
Form 2, Product Accountability, records the materials, special processes, and functional tests that the design calls out, listed in the order they appear in the manufacturing process.
Form 3, Characteristic Accountability, lists every design characteristic, meaning each dimension, tolerance, note, and geometric callout, alongside the actual measured result.
Form 3 is where the volume lives. AS9102 requires an inspection drawing or model on which every characteristic is marked with a uniquely numbered balloon, and the balloon numbers must correspond to the characteristic numbers on the form. That single rule is the source of most of the manual effort, because someone has to find, number, and transcribe each callout without skipping one. The standard was last revised as AS9102 Revision C, released by SAE in June 2023, which added guidance for digital product definition and 3D model data and formalized electronic signatures so that a signature on Form 1 locks the full report.
IN PRACTICE
The connection to our PDM and using that as a data source is legit the best thing ever. I found three viable bracket options fitting my exact envelope constraints, in minutes, not days.
Eytan S., R&D Engineer
How AI ballooning reads a drawing
Automated ballooning is the step that turns a flat drawing into a numbered list of characteristics. Older tools relied on optical character recognition alone, scanning the sheet for text and trying to label each value. That works for clean, well spaced dimensions and struggles with everything else. Modern AI ballooning pairs character recognition with a vision model that interprets the drawing the way an inspector does, associating a number with its leader line, its tolerance, and the feature it points to.
The practical payoff is speed. In one published head to head test, an AI ballooning system captured 89 of 95 characteristics, roughly 94 percent, in about 19 minutes against more than six hours for the same drawing done manually. The errors are instructive: the misses included handwritten notes that the text engine misread, GD and T composite frames with ambiguous layout, and a callout buried in a cluttered detail view. That pattern matches what published research on engineering drawing extraction reports, where complex tolerances, modifiers, and non standard text orientation remain the hardest cases.
The takeaway is that AI ballooning removes the tedious bulk of the job while leaving a smaller, clearly defined set of judgment calls for a person. That mirrors the broader shift toward AI design review that catches errors before manufacturing, where the goal is to surface issues early rather than hand the work off entirely.
Extracting dimensions and GD and T correctly
Plain dimensions are the easy part. A linear value with a plus and minus tolerance is structured and predictable. Geometric dimensioning and tolerancing is where extraction gets difficult. Governed by ASME Y14.5, GD and T encodes geometric intent through feature control frames that combine a geometric symbol, a tolerance value, optional modifiers, and one or more datum references. Reading a feature control frame correctly means parsing all of those parts in the right order, not just recognizing the symbols.
This is why naive symbol detection is not enough. A position callout with a material condition modifier and three datums carries very different meaning from a bare profile tolerance, and an extraction tool that drops a modifier produces a report that looks complete but is wrong. The most reliable approach treats the frame as a structured object, capturing symbol, value, modifiers, and datums together, then flags low confidence reads for human review instead of silently guessing.
Getting this right depends on genuinely understanding the drawing, the same capability that underpins AI CAD software for mechanical engineering and the move toward model based definition. When the design is delivered as a 3D model with embedded annotations rather than a 2D sheet, the characteristics can be read directly from the model, which AS9102 Revision C explicitly accommodates.
Generating the AS9102 report and keeping it traceable
Once characteristics are ballooned and extracted, report generation is largely a mapping problem. Each numbered characteristic becomes a row on Form 3, with its nominal value, tolerance, and a slot for the measured result. The part identity and revisions populate Form 1, and the materials and special processes feed Form 2. The value of automating this step is not only speed but traceability, because the link between a balloon on the drawing and a row in the report is created automatically rather than retyped.
Traceability matters because FAI is an auditable record. If a measured result later falls out of tolerance, an auditor needs to walk from the report row back to the exact callout on the drawing. Manual transcription breaks that chain whenever a number is fat fingered or a revision is missed. Automated mapping keeps the balloon, the characteristic, and the form row tied together, and it makes a revision change far easier to handle, since the tool can re extract and compare rather than forcing a full restart.
A common failure mode in real shops is that the source drawing or its revision is hard to locate in the first place, a problem we covered in why PDM search is broken. If the inspection starts from the wrong revision, every downstream form is wrong too, which is why report generation should be anchored to the controlled file in the PDM system, not a copy on someone's desktop.
Where Leo fits: drawing understanding for inspection prep
The capability at the center of AI first article inspection is reading and understanding an engineering drawing: the parts, the PDF, the dimensions, and the geometric callouts. That is the same drawing understanding that powers Leo's design review, where Leo reads an assembly and its drawings to flag problems like a missing critical diameter or dimensions that prevent a part from being assembled. The connection is direct. A system that can spot a drawing missing a diameter a machinist needs is already doing much of the reading an inspector does to prepare an FAI.
Leo is an intelligence layer that sits on top of your existing PDM and PLM, not a replacement for it. Integrations are available for SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM. Because Leo works from the controlled file in those systems, inspection prep starts from the correct, current revision rather than a stray copy, which is exactly the traceability that AS9102 expects. The same engine that supports engineering knowledge management across a team also means the context behind a characteristic, the related parts, prior notes, and design intent, is available while the drawing is being read.
FAQ
Read your drawings before inspection
See how Leo understands your CAD and drawings to speed FAI prep.
Leo is an intelligence layer that connects to your PDM and PLM, reads your drawings and models, and surfaces the characteristics behind every part. Book a demo.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
