What a PE Must Verify Before Stamping an AI-Generated Drawing

A licensed PE can stamp AI-generated engineering drawings only when the work meets the same requirements for competence, responsible charge, and verification as hand-drafted work. An engineer who approves AI-produced structural calculations or drawings still has the same legal liability. Before stamping, the PE must verify dimensional accuracy, applicable code requirements, discipline coordination, and documented assumptions through a structured review built into production.

Can engineers stamp AI-generated drawings with less legal responsibility than hand-drafted ones?

A PE’s stamp is a full legal attestation of responsible charge. NCEES Model Rules define Responsible Charge as direct control and personal supervision, and NCEES did not soften that definition when it addressed AI in August 2025. The drafting method does not change the obligation. NSPE Code section II.2.b says engineers must not sign documents outside their competence or documents that were not prepared under their direction and control. An AI-generated drawing cannot satisfy those conditions merely by appearing complete.

NSPE Case 24-2 establishes the relevant ethical distinction. The Board of Ethical Review found that using AI tools is not itself unethical, but sealing plans without meaningful supervision violates Code section III.8.a. ASCE Policy Statement 573, issued in July 2024, and the NSPE position revised February 2025 also say that AI cannot replace professional engineering judgment or change the licensure standard for engineers who oversee systems with public-safety impact. Gradon completes engineering workflows inside existing AEC software on the same basis, with licensed engineers responsible for reviewing and stamping the output. Whether geometry comes from a human drafter or a system trained on the firm’s standards and past projects, that responsibility remains with the PE.

Why AI drawing tools introduce failure modes that hand-drafted work does not

Human drafting errors commonly arise from inattention or misunderstanding. AI adds a different source of error. Statistical pattern-matching can create geometry that looks right even when its dimensions, element counts, or structural implications are wrong. These errors can occur on a large scale without the system giving any sign that the output is wrong.

A 2025 benchmark that tested seven multimodal LLMs on residential plans found that even the best model, Gemini 2.5 Pro, reached only 41% mean accuracy across five drawing elements. Some window-count errors were more than double the correct figure. Those results concern drawing understanding, rather than every possible engineering application, but they expose the danger of accepting polished output without verification. The benchmark’s researchers found that the models were not reliably production-ready for general drawing understanding without human oversight. High-level spatial interpretation, including room types and rough layouts, was stronger than fine-grained element counting.

For the PE, the lesson is to structure review around the details that a convincing layout can conceal. Quantity takeoffs, code compliance checking, and construction documentation need near-perfect accuracy, which takes more than a drawing in which the project is broadly recognizable. Every element still requires human review. Dimensional strings and cross-discipline interfaces need particular attention because plausibility within one view does not establish agreement across a drawing set.

Dimensional accuracy — verifying that what the drawing shows is what the numbers say

The dimensional review should begin with an independent calculation of at least one critical dimension path per sheet. The reviewer traces the geometry and compares the result with the annotation. This makes responsible charge concrete and measurable. It is a required review step whether the output comes from a point tool or a workflow system such as Gradon, because plans, elevations, and sections can contain conflicting dimensions without the AI flagging them. The critical path is a deliberate independent check that must be supplemented by review of the remaining dimensions.

From that path, the PE must verify that every dimension string closes and that dimensions agree between plan views and details. Grid lines must remain consistent across sheets. Section cuts and elevations must show the same geometry, and keynotes must point to elements that are there and have the stated size. Where a view and its annotation disagree, the discrepancy requires resolution before downstream review can rely on either. Otherwise, code and coordination checks may be evaluating different versions of the same element.

Schedules must also be checked against the drawn geometry. The same benchmark reported window extraction accuracy of only 14% in the best-performing model, making AI-generated window and door schedules unverified until their counts are reconciled with the drawings. Visual proportion is not an adequate substitute. Training drawings at inconsistent scales can also lead a model to reproduce proportions that look correct on screen but conflict with dimensions in the title block. The PE must therefore establish agreement between geometry, scale, dimensions, and schedules rather than treating any one of them as proof of the others.

Code compliance — confirming that the design meets the jurisdiction’s actual current requirements

A dimensionally correct drawing can still be unpermittable. AI tools trained on aggregated drawing data may combine requirements from different jurisdictions and code editions into a design that complies with no single applicable code. The PE must start the review with the edition that the project’s jurisdiction enforces. The governing requirements must be tied to this project rather than inferred from the apparent completeness of the drawing.

The next step is to check the site-specific design parameters. Wind, seismic, snow, and flood zone parameters must be based on the project site. Prescriptive path exemptions also require verification against the scope as designed. An exemption that appears in the generated documentation is still an engineering claim to be tested before further analysis can be omitted. This separates identifying a potentially relevant provision from determining that the project qualifies to use it.

The Miami-Dade load-path example makes that distinction concrete. A wall removed by a homeowner, or misclassified by AI as non-structural, can support roof loads and require retroactive engineering and stamping after the fact. The engineering role of the wall governs. A generated classification cannot replace evaluation of how the actual structure transfers loads under the jurisdiction’s requirements. Case 24-2 shows that using AI-assisted tools without checking the work fully fails the responsible-charge standard. The PE cannot delegate the compliance decision to AI and then treat its answer as sufficient grounds for a stamp.

Discipline coordination — checking that structural, civil, MEP, and architectural drawings agree with each other

Once the geometry and the requirements that apply have been checked, review must move from each sheet to the full drawing set. AI tools for each discipline may create each deliverable on its own and miss that a structural beam conflicts with an MEP duct run or that a civil slab penetration is missing from the architectural drawings. Each sheet can look coherent in isolation. The conflict becomes visible only when the sheets are treated as descriptions of one building.

The PE in responsible charge must verify that penetrations, openings, clearances, and elevations agree across the disciplines. A clash-check or manual overlay of structural, MEP, and architectural plans at key building sections is a practical way to review the work before stamping. AI-generated sets warrant the same coordination discipline that a construction manager would apply to drawings from multiple sub-consultants. Because Gradon works across disciplines within existing tools, it is better placed to help keep drawings consistent than point tools that create each drawing on its own, and coordination review is still needed.

Finding a clash is only part of the work. When structural depth conflicts with mechanical clearance, the engineer must decide how to reconcile the requirements rather than accept whichever solution each tool produced. NCEES’s August 2025 statement frames AI as a tool to assist professional judgment. Determining which discipline’s constraint governs a conflict is precisely the kind of decision that the licensed engineer must retain, within the competence required for the work.

Traceability of assumptions — documenting what the AI assumed so the stamp can be defended

Coordination decisions are defensible only when the underlying assumptions are also available for review. An AI drawing tool may embed a default live load from training data, a footing depth from a similar project, or a model-selected wind exposure category without making that choice explicit. The drawing can conceal the design’s starting point. A PE who stamps work with undocumented assumptions cannot demonstrate that those engineering decisions were personally made or made under supervisory direction.

The Board’s comparison with directing an engineering intern is useful here. The engineer must establish solution guidelines and constraints and review the work before it leaves the firm. An engineer can rely on an intern to explain their assumptions when asked, but cannot rely on AI to do so. A completed drawing is not an assumptions record. The PE must extract and evaluate the basis for the design rather than infer that familiar dimensions or specifications came from appropriate project inputs.

That record also depends on permissible handling of client information. The Board treats uploading client information to an AI open-source interface as tantamount to placing private information in the public domain, so the firm’s policy must identify permissible platforms before drawing generation begins. Before stamping, the PE should produce or require a written assumptions log covering soil parameters, load combinations, code edition, material specifications, and any AI-generated defaults. The log must be retained with the project file so that the basis for the design can be produced if it is questioned.

How to build pre-stamp review into the firm’s AI workflow rather than treating it as a final hurdle

For the firm, these checks belong within production rather than at the end of a fully generated drawing set. End-stage verification is slower and riskier because errors can progress from dimensions into discipline coordination before anyone resolves them. Earlier dimensional corrections are cheaper than coordination-stage corrections. NCEES’s view of responsible charge calls for quality assurance based on experience, with critical discussions and mentorship as part of a supervisory relationship that continues as the work develops. A final signature cannot replace that relationship.

The firm’s AI policy should turn that obligation into defined responsibilities. It should specify approved platforms for client work, the documentation required with each drawing, who verifies the work at each phase, and the sign-off required before the PE stamps. Review ownership must be explicit throughout production. Engineers with enough domain experience to judge the output need to take part as the work develops, so a PE is not left to find every unresolved issue in a final pass. Those responsibilities connect the dimensional check, jurisdictional review, coordination decision, and assumptions log to the people accountable for completing them.

A production model in which licensed engineers review and approve all AI output, as in Gradon’s approach inside the firm’s existing tools, illustrates one way to make review part of the work. The operational aim is checking at the stage where an engineer can evaluate and resolve the issue before it becomes the basis for further work. Faster drafting provides an advantage only when the resulting output receives the supervision and verification required for professional use. The PE’s stamp is the moment of accountability, and the firm’s documented review process is what makes that moment defensible.