Somewhere in the pre-production phase—before metal meets machine, before the first part hits the shop floor—costly mistakes hide in plain sight. A misplaced datum reference. A tolerance stack that creates assembly interference. A title-block note pointing to the wrong material spec. These errors, buried in GD&T callouts and engineering drawings, have quietly consumed up to 20 percent of manufacturing costs in some aerospace cases, according to industry consultants—though such figures should be treated as illustrative rather than definitive.
Now a wave of startups believes artificial intelligence can find them first.
The pitch sounds straightforward: automate the tedious work of checking drawings against standards like ASME Y14.5-2018 and ISO 1101, flagging violations before anyone commits to production. Free experienced engineers to focus on design intent rather than hunting for missing dimensions on page twelve of a weldment package. But the reality, as with most automation promises, is more nuanced than the sales deck suggests.
Consider the pressures converging on manufacturers. ETQ's Pulse of Quality survey, published early in 2025, found that 75 percent of manufacturers had experienced a product recall in the previous five years—a figure that crept up from 73 percent the year before. At the same time, a Deloitte and Manufacturing Institute report projects the industry will need 3.8 million workers between 2024 and 2033, with nearly half those positions potentially going unfilled without significant intervention.
Those two trends—rising quality failures and disappearing expertise—are colliding precisely where engineering drawings still govern what gets built.
The Manual Gauntlet
Engineering drawings remain the authoritative production document across most discrete manufacturing environments, even as model-based definition slowly gains ground. Industry surveys have shown companies still lean heavily on 2D PDFs alongside 3D models, particularly in aerospace, medical devices, and pressure vessels, where regulatory frameworks and supplier ecosystems evolved around traditional drawing packages.
The review process? Manual, methodical, and labor-intensive.
A senior engineer or checker walks through each sheet, verifying GD&T frames comply with ASME Y14.5, confirming datum references stay consistent across views, ensuring tolerance stacks won't create assembly interference, checking that notes reference the correct material specs or code clauses. In regulated sectors like aerospace—where AS9102C governs First Article Inspection reports—every single characteristic on the drawing must be verified and documented. One missed callout, one ambiguous datum, and the downstream consequences can range from scrap and rework to a failed source inspection.
And the people who catch those mistakes? They're retiring. Fast.
The same Deloitte report warns that without action, nearly half the projected manufacturing workforce need will go unmet through 2033. ETQ's survey found 70 percent of manufacturers say labor shortages directly impact quality outcomes. Training replacements takes years—a junior engineer can learn to read a GD&T frame easily enough, but understanding when a perpendicularity tolerance creates an unintended stack-up with a profile callout three sheets away requires pattern recognition built over hundreds of reviewed drawings.
Enter the AI Checkers
A cluster of startups thinks they've found the opening.
HERA, which emerged from Y Combinator's summer 2026 batch, launched late in July with a product that reviews engineering drawings against ASME and ISO standards, plus customer-specific rules. Founded by Meera Patel and Noelle So, the Seattle-based startup demos a tool that checks GD&T frames, tolerance stacks across multiple sheets, title-block completeness, and code compliance for sectors like pressure vessels, piping, and weldments. The company frames itself as a "QA layer for physical products," promising reviews in minutes rather than days. Target customers: engineer-to-order manufacturers in aerospace, medical devices, and pressure vessels.
CoLab Software announced an agreement with the Standards Council of Canada in late July 2026 to bring ISO standards under license into its AutoReview product. The deal allows the AI to cite specific ISO clauses when flagging a deviation—a move toward compliance-grade checks with auditable standard references. An earlier blog post from the company demonstrated first-pass drawing checks for materials, bill-of-materials accuracy, tolerances, and GD&T using organization-specific checklists.
Others are staking out adjacent territory. Axial flags ASME Y14.5-2018 and ISO 1101 violations, returning a redlined PDF. RapidDraft offers what it calls "agentic" drawing release and design review that checks completeness, GD&T schemes, and tolerance stacks against internal rules. Paperless Parts—which processes north of 10 million pages of drawings annually for quoting—launched Wingman 2.0 in May 2026 and secured a patent in April for AI that reads drawings and structures GD&T and notes for quoting automation. The company claims Wingman 2.0 extracts "14 times more technical elements per drawing" than prior versions, though naturally such vendor metrics invite scrutiny.
Legacy CAD vendors aren't sitting still. PTC released Creo 13 in June 2026 with AI-powered design guidance. Autodesk's Fusion now includes AI assistants for drawing automation. Siemens NX has long offered Check-Mate for design validation. These established tools—PTC's ModelCHECK, SOLIDWORKS Design Checker—already enforce syntactic and formatting rules based on configurable standards files. The new AI layer adds something different: semantic reading. Interpreting design intent. Spotting cross-sheet contradictions. Flagging GD&T choices that might pass a syntax check but create downstream manufacturing headaches.
The Mechanics Under the Hood
The technical approach blends document vision with domain knowledge, leaning on recent academic work that describes hybrid architectures. Vision transformers parse drawing layouts. Oriented bounding-box detectors isolate GD&T frames and symbols. Vision-language models interpret context and relationships. The output: structured JSON capturing tolerances, datums, surface finishes, material callouts, and notes—data that can feed process planning, First Article forms, or supplier portals.
But the AI has to handle messy real-world PDFs. Hand-marked redlines. Legacy scans with skewed text. Non-standard symbol libraries. Drawings that mix metric and imperial units across views. A 2026 smart-manufacturing AI roadmap emphasized the need for reliability metrics and MLOps practices, especially in regulated industries where an incorrect interpretation can void a certification.
CoLab's blog argued that most tools can read GD&T symbols but miss the functional intent behind datum strategies—a gap that still requires human judgment. Which is perhaps the most honest assessment in this space.
Real-world deployments show time savings, at least in vendor case studies. Scalian reported in March 2026 that AI reduced the time to check First Article Inspection reports from roughly three hours to five to ten minutes in aerospace applications. Coffee Inc. published a case study claiming auto-ballooning cut annotation time from two to four hours per sheet down to less than five minutes, with AS9102 and PPAP reports auto-generated from the result. A February 2026 practitioner blog detailed multiple FAI failures caused by missed coating specs and incomplete characteristic verification—underscoring the stakes.
Standards, Compliance, and the Auditor's Eye
Compliance frameworks shape how these tools are built and sold, maybe more than market demand does.
AS9102C, published in June 2023, governs First Article Inspection in aerospace and requires 100 percent verification of drawing characteristics. Adoption by major primes accelerated through 2024 and 2025. AS9145 extends advanced product quality planning and production part approval process requirements across the aerospace supply chain; major primes increasingly mandated it in supplier agreements throughout 2024 and 2025. Drawing correctness feeds directly into FAI acceptance or rejection.
Medical device manufacturers operate under the FDA's Quality Management System Regulation, which became effective in February 2026 and aligns with ISO 13485:2016. That framework demands rigorous design controls, change management, and traceability. In pressure vessels and piping, jurisdictional codes require compliance with ASME BPVC and B31 series standards; inspectors verify that drawing notes match code clauses and that welding symbols reference approved procedures.
The CoLab and Standards Council of Canada agreement signals a notable shift—toward licensed-standards integration. Rather than training AI on generic GD&T patterns, vendors are embedding the actual text of ISO and ASME standards so the tool can cite clause numbers in its markups. That improves auditability in regulated environments where an inspector or certification body needs to trace a flagged violation back to a specific standard requirement.
HERA markets an "Evidence Chain" concept—tracking who flagged what, which standard clause was violated, what version of the drawing was reviewed. The company says prints never leave customer facilities, addressing concerns about controlled technical data in defense and aerospace. Paperless Parts highlights FedRAMP and CMMC boundary considerations for AI systems handling CUI or ITAR-controlled drawings.
All very reasonable. And all very necessary in sectors where liability and certification are on the line.
Market Momentum, or Market Hype?
Market forecasts reflect urgency, or at least vendor enthusiasm. ABI Research pegged the AI-powered quality management market at $11.4 billion in a May 2026 analysis. Grand View Research valued the machine vision quality assurance and inspection segment at roughly $9.95 billion in 2025, projecting a 13.1 percent compound annual growth rate through 2033. Multiple syndicated reports from 2025 and 2026 forecast double-digit growth for AI visual inspection tools through the early 2030s, though the numbers vary widely across firms—as they tend to.
The broader regulatory and customer pressures are real enough. Northrop Grumman issued a supplier letter in July 2025 requiring all First Article Inspection Reports be submitted through Net-Inspect, a platform that auto-generates AS9102 forms from ballooned drawings. The FDA's Quality Management System Regulation raised the bar for design documentation and traceability. The 2025 and 2026 editions of ASME BPVC and B31.3 codes are now in force, propagating into drawing notes and inspection criteria that suppliers must interpret correctly.
A CoLab survey of 250 manufacturing leaders, conducted around mid-2026, claimed 96 percent value design standards, but nearly half aren't applying them consistently. The same survey suggested roughly 75 percent of drawing review could be automated when grounded in organization-specific rules. That aligns with industry expectations that AI will handle first-pass checks—completeness, syntax, simple GD&T, cross-sheet consistency—while experts focus on design intent, design-for-manufacturing trade-offs, and the judgment calls that can't yet be codified.
Perhaps can't ever be codified, though no startup is going to say that out loud.
What Comes Next
The competitive landscape will likely stratify along familiar lines. Standalone AI checkers like HERA, CoLab, and Axial compete for early adopters willing to embrace new workflows. Legacy CAD vendors embed AI into tools engineers already use, which may accelerate adoption in shops with entrenched PTC, Siemens, or SOLIDWORKS deployments. Quoting and supplier platforms like Paperless Parts pull drawing intelligence into the RFQ phase, normalizing AI "reading" before production even starts.
Integration with PLM, ERP, and QMS systems will determine how much friction these tools actually remove. Structured outputs from drawing readers should flow into routing, process planning, SPC charts, and supplier portals, reducing re-keying and the mismatch risk that comes with manual data transfer. Net-Inspect's adoption by primes like Northrop Grumman shows how auto-ballooning and characteristic extraction are becoming table stakes for aerospace suppliers. The same logic applies upstream: if the drawing data is already structured and verified, downstream processes get cleaner inputs.
Generalization remains the challenge. Academic work from recent years reports strong precision and recall on curated datasets, but messy scans, non-standard conventions, and ambiguous annotations still trip up the models. A 2026 quality assurance paper on industrial AI emphasized the need for human-in-the-loop validation, especially in sectors where liability and certification are on the line. Organizations will likely keep human approvals even as AI handles the grunt work—which is probably wise.
The broader shift, vendors insist, is from reactive to preventive quality. BCG's May 2026 report on AI-powered factories argued that integrating AI, automation, and digital systems reshapes cost structures and competitiveness. Deloitte's tech trends guide for CFOs around the same time emphasized preparing for agentic AI by investing in data management and governance.
Catching a GD&T error before the first part ships is cheaper than catching it during First Article Inspection. Far cheaper than catching it during a recall. If AI can compress that feedback loop from weeks to minutes, the math changes for manufacturers deciding whether to automate pre-production review or keep hiring checkers they can't find.
The question isn't whether these tools will improve. It's whether they'll improve fast enough, and whether the humans who still know how to read those drawings will stick around long enough to teach them.
