Addressing Engineering Drawing Compliance with AI

What machines are genuinely good at here

The productive split is mechanical checks to the machine, engineering judgment to the engineer. Symbol recognition, tag-format validation, line-number continuity, legend conformance, title-block completeness and cross-drawing consistency are deterministic checks with objectively correct answers.

Modern vision models combined with rule engines handle these reliably. What they should not do is decide whether a design intent is sound — that stays with the reviewer, now working from a pre-annotated drawing instead of a blank one.

• Symbol and equipment detection against the project legend
• Tag numbering and format validation against the project specification
• Line continuity and off-page connector reconciliation across sheets
• Revision-to-revision difference detection to focus review on what changed

Designing the workflow, not just the model

The model is the easy part. The workflow decides whether the capability gets adopted. Every automated finding needs a location on the drawing, a rule reference, a severity, and a one-click accept or reject from the reviewing engineer.

Rejections are not failures — they are the training signal. A compliance workflow that captures why an engineer overrode a finding improves faster than one that only counts detections.

Measuring it honestly

Track three numbers: review cycle time per sheet, findings caught before issue versus after, and reviewer override rate. The first proves efficiency, the second proves quality, and the third tells you whether the rule set is actually calibrated to the project.

A falling override rate over the first few hundred sheets is the clearest sign the system is converging on how this particular project defines compliance.

Anchor rules to published standards, not tribal memory

A compliance engine is only as defensible as the rule references behind it. Symbol and identification conventions on P&IDs are governed by ISA-5.1, 'Instrumentation Symbols and Identification'; reference designation and structuring principles sit in the IEC/ISO 81346 series; and safety instrumented functions in the process sector carry separate obligations under IEC 61511.

Writing each automated check against a citable clause changes the conversation with reviewers. A finding that says 'tag format deviates from project specification §4.2, consistent with ISA-5.1 identification rules' is discussed on its merits. A finding that says 'anomaly detected' is argued about.

Where a project deliberately departs from a standard — and capital projects frequently do — that departure belongs in the rule set as an explicit, versioned exception, not as an undocumented tolerance in someone's head.

• ISA-5.1 for instrumentation symbols and tag identification
• IEC/ISO 81346 for reference designation and system structuring
• IEC 61511 for safety instrumented systems in process industries
• ISO 19650 for the information-management envelope around issue and review

What the research says about machine vision on drawings

Automated interpretation of P&IDs is an active and reasonably mature research area, not a vendor claim. Published work in engineering informatics — including digitisation pipelines that combine symbol detection, text recognition and line-tracing to reconstruct connectivity graphs from scanned P&IDs — consistently reports high accuracy on symbol and text recognition, with line tracing and connectivity reconstruction remaining the harder problem.

That asymmetry should shape expectations. Detecting that a control valve exists is close to solved; correctly asserting that it sits on line 6"-P-1204-A1 across a sheet break is where residual error concentrates. Deployments that treat connectivity findings as advisory and symbol/tag findings as near-authoritative match the actual error profile of the technology.

The Stanford AI Index has tracked the steady fall in error rates and cost of vision and multimodal models over successive editions; the practical consequence for engineering teams is that re-running a full drawing set after a rule change has become economically routine rather than a special project.

Human oversight as a design requirement

Compliance review is exactly the class of use where regulators expect a human in the loop. The EU AI Act sets out human-oversight and record-keeping duties for higher-risk deployments, and the NIST AI RMF frames the same expectation as an organisational practice: define who reviews, what they can override, and how those overrides are recorded.

Practically, this means the reviewing engineer's signature — not the model's output — remains the compliance artefact. The system's contribution is that the engineer signs having seen every mechanical check performed consistently across every sheet, which is more than a manual process can honestly claim at volume.