AI for Engineering Productivity
How AI design review checks work against ASME Y14.5, ISO 13485 design controls, and AS9100 in 2026, and how to judge whether a compliance check is trustworthy.
·
⏱
8 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
February 2, 2026 moved medical device design controls from 21 CFR 820.30 into ISO 13485 clause 7.3 and retired the old inspection technique with it. The EU AI Act obligations that were expected alongside it were deferred to December 2027 and August 2028. Neither change alters the underlying problem: the evidence that a design met ASME Y14.5, ISO 13485, or AS9100D is spread across models, drawings, review minutes, and change records that no one person has read together. AI is well suited to reading across that material on every revision and flagging where the requirement, the approval, and the configuration disagree. It is not suited to being the approver. Judge any compliance check on whether every finding carries a citation you can open, which edition of the standard it applied, and whether the check itself can be shown to an auditor.
On February 2, 2026, the rules under a large share of regulated mechanical design work quietly changed. The United States Food and Drug Administration's Quality Management System Regulation took effect, replacing most of the old Quality System Regulation in 21 CFR Part 820 with ISO 13485:2016 incorporated by reference. Design controls did not disappear. They moved into a standard written in different language, audited by a different inspection program.
Aerospace and defense teams already live under AS9100D, which layers configuration management, product safety, and counterfeit part controls on top of ISO 9001:2015. Almost everyone drawing parts works under ASME Y14.5, reaffirmed in 2024 in its 2018 edition. Three different standards, three different vocabularies, one shared reality: the evidence that a design met them is scattered across models, drawings, review notes, and change records that no single person has read end to end.
That gap is where AI design review is genuinely useful, and where it is easiest to oversell. This article covers what a compliance-aware inspection can actually check, where it breaks, and how to judge one before you put it anywhere near an audit trail.
What Actually Changed in 2026, and What Did Not
The clearest change is the medical device one. The Quality Management System Regulation final rule was issued on January 31, 2024 and took effect on February 2, 2026. It amends 21 CFR Part 820 to incorporate ISO 13485:2016 by reference, so a requirement that used to be read out of the American regulation is now read out of the international standard. On the same date the agency retired the Quality System Inspection Technique and moved to the inspection process described in Compliance Program 7382.850.
For a mechanical design team, the practical effect is narrower than the headlines suggest. Design and development planning, inputs, outputs, review, verification, validation, transfer, change control, and the design file are all still required. They now sit in clause 7.3 of ISO 13485 rather than in 21 CFR 820.30, and the process-oriented, risk-based framing of the standard applies to them. Teams that maintained a mapped quality system against both documents have a renaming exercise. Teams that only ever wrote to the old regulation have a translation exercise.
The second change is the one most commonly overstated. The European Union AI Act does classify AI that is a safety component of a device regulated under the Medical Device Regulation or the In Vitro Diagnostic Regulation as high risk. But the Digital Omnibus on AI, given final approval by the Council of the European Union on June 29, 2026, deferred those obligations. Standalone high-risk systems under Annex III moved to December 2, 2027, and high-risk AI embedded in regulated products under Annex I moved to August 2, 2028. If you were told that an AI design assistant triggers AI Act duties in 2026, that is not correct as the timeline now stands. The harmonised standards that would make those duties concrete are still in development.
What did not change is the part that actually costs money. Auditors still ask for objective evidence, and that evidence still has to be found in the design record rather than reconstructed from memory.
IN PRACTICE
With Leo, our team improves design quality, reduces mistakes, and shortens time-to-market.
- Uriel B., Field Warfare and Survivability Specialist
The Three Standards a Design Check Has to Read
The three standards operate at different layers, and conflating them is the most common source of a compliance check that looks thorough and proves nothing.
ASME Y14.5 governs how a requirement is expressed. The 2018 edition, reaffirmed as Y14.5-2018 (R2024), sets the symbols, rules, defaults, and interpretation for dimensioning and tolerancing. It grew to 326 pages from 214 in the 2009 edition, restructured profile, orientation, and form into separate sections, and moved substantially toward model-based definition by pairing many two-dimensional figures with three-dimensional equivalents. A check at this layer answers whether a datum reference frame is fully constrained, whether a feature control frame is complete, and whether a callout means what its author thought it meant. Our guide to engineering drawing standards consistency covers the recurring failure patterns, and model-based definition compared with 2D drawings covers where the annotation actually lives.
ISO 13485 governs whether the design process was followed and evidenced. It does not care what a position tolerance is. It cares that design inputs were documented, that outputs were verified against them, that reviews happened with the right participants at the right stages, and that the design file holds the record.
AS9100D governs configuration and safety on top of ISO 9001:2015. It adds aerospace-specific requirements around configuration management and traceability, product safety, prevention of counterfeit parts, first article inspection, key characteristics, and human factors. Its distinctive demand is that you can show, for any delivered article, exactly which configuration it was built to and what changed since.
A tool that checks only the first layer is a drawing checker. A tool that checks only the second is a document management system. The interesting work sits where a geometric requirement, the review that approved it, and the change that altered it have to be reconciled as one story.
Where Design Compliance Actually Breaks
In practice, findings rarely come from an engineer who does not know the standard. They come from records that drifted apart.
The requirement and its evidence live in different systems. A design input sits in a requirements document, the tolerance that satisfies it sits in a model, and the verification that proves it sits in an inspection report filed by part number rather than by requirement. Nobody linked them at the time because linking them was manual. Building that connection deliberately is the subject of AI for requirements traceability.
A change propagated to the model but not to the record. The revision went through, the drawing updated, and the design review minutes still describe the previous geometry. Under AS9100D configuration management this is the finding, not the geometry.
Annotation is technically legal and practically ambiguous. An underconstrained datum reference frame, a profile tolerance with no basic dimensions behind it, or a surface finish callout with no inspection method attached will pass a casual review and produce a supplier question or a nonconformance later.
The standard moved and the template did not. Title blocks and drawing templates still invoking the 2009 edition are common years after a team adopted the 2018 rules in practice, which means the document says one thing and the design intent is another.
The knowledge is in a person, not a system. The engineer who knows why a particular tolerance was tightened in 2019 is the record. When they leave, the justification leaves with them, and the next reviewer either re-derives it or defers to it.
None of these are exotic. All five are cheap to catch early and expensive to catch during an audit or a first article inspection. Broader production-side checking is covered in AI compliance checking for manufacturing.
What Compliance-Aware Inspection Looks Like in Practice
The useful framing is inspection rather than approval. An AI layer does not sign off a design and does not replace a quality function. It reads the same material a careful reviewer would read, does it on every revision instead of on the ones someone had time for, and shows its working.
This is the value driver behind Leo. Leo is an AI assistant for mechanical engineers trained on more than one million pages of standards, books, and technical articles, and it connects to an organisation's own knowledge base: product data management, product lifecycle management, local and network directories, and enterprise resource planning. Leo offers integrations with leading PDM and PLM platforms including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM, among others. That matters for compliance work specifically, because the evidence an auditor wants is already in those systems and the reason it is hard to produce is that no one can read across them quickly.
Concretely, a compliance-aware inspection layer over that material can flag a feature control frame whose datum reference frame does not fully constrain the part, surface the design review record associated with a given revision and note where the minutes describe superseded geometry, retrieve the prior calculation or standard clause that justified a tolerance and cite it, and identify the parts in a family that were built to a configuration the current drawing no longer describes.
Every one of those outputs should arrive with a citation the reviewer can open. That is not a nicety in regulated work. An assertion without a source is not evidence, and a design record built from unsourced assertions is worse than no automation at all. Leo is SOC-2 certified and GDPR compliant, no AI is trained on customer data, and customer intellectual property remains protected, which is the baseline before this kind of material goes anywhere near an external service.
How to Evaluate an AI Compliance Check Before You Trust It
Compliance is the one area where a confident wrong answer is more damaging than no answer. Five questions separate a check you can rely on from one you cannot.
Does every finding carry a citation you can open? Ask for the clause, the document, and the revision. If the tool can only assert, it can only be a prompt for a human to go and verify, which is a much smaller claim than it usually makes.
Which edition of each standard is it reading? Y14.5-2009 and Y14.5-2018 disagree in ways that change interpretation. A tool that cannot tell you which edition it applied is not auditable.
Does it read your record or only the file in front of it? A checker that sees one drawing cannot detect the configuration and review drift that produces most findings. It has to reach into the systems where history lives.
What happens to your data? Design files in regulated programs are frequently export controlled or subject to customer confidentiality. Certification status, training policy, and data residency are procurement questions, not technical details.
Can you show the check itself to an auditor? An automated review that cannot be described, versioned, and evidenced becomes its own finding. If the check runs on every revision, its output is part of the design record and should be treated that way, in the same way a nonconformance report workflow is.
Run these questions against a real part from a real program, ideally one where you already know what an auditor found. A tool that rediscovers a known finding, with a citation, on material it had not seen before is worth continuing with. A tool that produces a long list of plausible observations you cannot trace is a review burden wearing the costume of a review.
FAQ
United States Food and Drug Administration, Quality Management System Regulation final rule amending 21 CFR Part 820, issued January 31, 2024 and effective February 2, 2026, incorporating ISO 13485:2016 by reference; the Quality System Inspection Technique was retired on the same date in favour of the Inspection of Medical Device Manufacturers Compliance Program 7382.850.
Council of the European Union, Digital Omnibus on AI, final approval June 29, 2026: obligations for standalone high-risk AI systems under Annex III deferred to December 2, 2027, and for high-risk AI embedded in regulated products under Annex I, which includes devices regulated under the Medical Device Regulation and the In Vitro Diagnostic Regulation, deferred to August 2, 2028.
ASME Y14.5-2018 (R2024), Dimensioning and Tolerancing, American Society of Mechanical Engineers; 326 pages, superseding ASME Y14.5-2009.
SAE International and the International Aerospace Quality Group, AS9100D, Quality Management Systems, Requirements for Aviation, Space and Defense Organizations, incorporating ISO 9001:2015 and adding requirements for configuration management, product safety, counterfeit part prevention, first article inspection, key characteristics, and human factors.
See compliance evidence in context
Find the clause, the review, and the change behind any design decision
Leo reads across your PDM, PLM, and network directories and answers design questions with a citation you can open, so the evidence exists before the audit asks for it.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
