Before selecting an AI drawing review product, a manufacturer must define the review boundary: which questions arise before formal release, which findings go back to engineering, and who approves the drawing. AI can extract title-block fields, surface possible conflicts and display differences between revisions. It does not automatically certify design intent, manufacturability or customer compliance. This guide turns that boundary into an RFP and FAT/SAT plan for a Thai factory.
Define pre-release engineering review
The input is a controlled candidate package: native CAD model, exported 2D drawing, title-block data, bill of materials and customer requirements. The output is a finding list, correction and recheck history, human approval and the released master. An AI finding is a question, not an approved drawing. Release is permitted only when drawing number and revision agree, significant findings have a documented disposition, the model/drawing relation has been checked, and an authorized person approves.
This activity differs from AI drawing search, which finds earlier drawings for reuse, and AI design-change management, which propagates approved changes into BOMs and departments. RFQ drawing analysis supports quotation on the supplier side. The point of control, owner and authoritative file differ.
Do not make “zero alerts” a cosmetic target. Each finding needs a status: correct, accept with a recorded engineering rationale, outside scope or hold. Product and customer rules determine whether a hold blocks release.
Fix the authoritative drawing and revision
Accuracy is meaningless if several PDFs and models carry the same drawing number without a known master. Identify every review package by part number, drawing number, drawing revision, model revision, export time and approval state. Store an immutable file ID or hash with the report.
Title-block candidates include number, part name, material, units, scale, projection method, revision, revision date and signatories. A letter “B” could mean revision B or sheet size B. Customer layouts, rotated scans and multiple languages complicate extraction. Require the source page and coordinates, confidence and master-data key for every extracted field.
Check title block against PDM/PLM and CAD attributes; check the associated model against the candidate revision; and check that the approval workflow points to the exact exported file. A new PDF export can invalidate a prior review. NIST’s digital-thread work identifies interoperability and trustworthy product data as continuing challenges across engineering, manufacturing and inspection.
Review dimensions and tolerances in three layers
Layer one consists of deterministic rules: mandatory notes, units, decimals, empty fields, frame conventions and native CAD checks. Creo ModelCHECK already analyzes drawings and reports rule and geometry checks. These established functions should not be described as novel generative AI.
Layer two extracts and compares information. A system can attempt to associate a dimension with a feature, flag conflicting callouts across views, compare a drawing value with a model feature and pass a tolerance to inspection. A scanned PDF provides a weaker basis than native CAD and semantic PMI. Misreading a diameter sign, plus/minus symbol or decimal point can change the conclusion; every finding needs a link to its source region. Identical text does not prove two dimensions refer to the same feature.
Layer three requires engineering judgment: functional dimensions, datum selection, tolerance stack-up, assembly clearance, measurability and customer-specific conditions. ASME Y14.5 defines the language and rules of GD&T, but it does not authorize an AI to select a suitable tolerance for a particular part. Specify the applicable edition and customer rules. ISO 16792:2021 addresses preparation, revision and presentation of digital product definition for both model-only and model-plus-drawing use. Its revision is under development; do not specify an unpublished replacement as a current standard.
| Suspected issue | Input | Candidate method | Required decision |
|---|---|---|---|
| Title-block number differs from PDM | PDF and PDM | OCR and keyed comparison | Engineer establishes the master |
| Drawing rev B points to model rev A | CAD metadata | Version-link rule | Correct link and recheck |
| Hole sizes appear inconsistent | Drawing and geometry | Extraction plus feature mapping | Engineer confirms same feature |
| Datum reference may be missing | Drawing/PMI and rules | Symbol extraction and rule check | Engineer assesses intent |
| Tolerance may exceed process capability | Drawing and process data | Comparison | Engineering and production decide |
An RFP that only asks a vendor to “detect defects” leaves the actual release decision unspecified. Require traceability to the source, disposition history and consistent results on repeat runs.
Compare the model and 2D drawing through features
A pixel overlay is insufficient when a drawing contains front, side, section and detail views. The system must show which view corresponds to which model feature. A hole visible only in section does not map to a simple image difference. Material and surface-treatment notes may not be encoded in geometry at all.
Split the requirement into file identity, geometry and manufacturing information. File identity includes number, revision, model link and export time. Geometry includes holes, outlines, sections and assembly components. Manufacturing information includes dimensions, tolerances, material, notes and PMI. Contract for the specific coverage and input formats. Without PMI in the model, not every drawing tolerance has a semantic model counterpart.
PTC’s June 2026 Creo 13 announcement describes an available AI Assistant, but calls direct 3D-model reading for issue detection a beta capability. Separate beta, released and conventional rule functions in the comparison matrix. Verify CAD version, licensing and results on your own files.
Separate visual difference from engineering meaning
Autodesk documents that DWG Compare highlights added, modified and removed objects between drawings or revisions. That is a visual comparison, not approval of the effect of a change. A hole diameter changing from 10 to 11 is detectable; whether it solves interference, introduces an error or follows a customer request depends on requirements. A frame template update can move many lines without changing the part. Test both cases.
Connect accepted pre-release changes to the downstream design-change process. Passing a drawing check does not prove that an ECN was delivered to purchasing and production.
Preserve human approval and evidence
Use distinct workflow states for machine check, AI-raised question, engineer disposition and authorized approval. Display the original page and coordinate, comparison source, rule, model/version, reviewer and timestamp. “No finding” means only that nothing was detected within the defined test scope.
Separate author and approver roles. Add quality and manufacturing engineering for critical characteristics. Accepted exceptions require reasons and references, but an old exception must not automatically approve a new material, process or customer condition. Only an authorized person releases the drawing. Keep AI output as auditable supporting evidence.
Where Japanese engineers and a Thai factory collaborate, translate explanatory comments while preserving drawing numbers, symbols and original callouts. Do not approve a dimension based solely on translated prose. Specify UI, notification, log and export languages separately.
Prepare representative evaluation data
A clean vendor demo does not establish performance at your site. Stratify by machining, sheet metal, assembly and tooling; native CAD, vector PDF and scans; customer and internal frames; current and superseded revisions; languages and scan quality. Declare excluded formats.
Include clean drawings, accepted exceptions, misleading near-matches and difficult scans, not only planted defects. For each expected finding, record type, severity, evidence and required workflow disposition. Agree ground truth with engineering, quality and production where needed. If experts disagree, document acceptable interpretations or remove the case from a binary score.
An illustrative planning example, not an industry benchmark, could use 200 drawing packages, 100 known finding events and 200 non-finding opportunities. Adjust to error rarity and consequence. Keep a tuning set separate from a held-out scoring set. Protect customer files through contracts and approved data handling.
Measure misses, false alerts and total review time
For each finding type, record true positives, false negatives and false positives. Recall equals true positives divided by true positives plus false negatives; precision equals true positives divided by true positives plus false positives. “95% accurate” is incomparable without denominators, file mix and exclusions. Report scans separately from native CAD.
Set thresholds by risk. As an illustrative contract example, not a proven product capability, require all revision mismatches in a held-out set to be detected, at least 90% recall for defined dimension-conflict candidates and no more than two false alerts per drawing on average. Buyers must choose their own values. Do not calculate a percentage if a category has no samples. Specify an “unable to assess” result and failure handling.
Compare full manual review time with full AI-assisted review time for similar drawings. Include investigating false alerts, re-export and recheck. A 30-second AI run provides no benefit if it adds 20 minutes of human work. Disclose volume, labor rates and recoverable time in any ROI model.
Write specific RFP requirements
List CAD versions, native formats, STEP, PDF and scan quality; frame variations; PDM/PLM and BOM connections; and read-only or write access. Define finding types, severity, source location, evidence, approval UI, audit log, export and exception handling. Mark each function standard, custom, beta or roadmap. Ask for proof on real samples.
Separate title-block OCR, fixed rules, native CAD checks, revision comparison, semantic model/drawing comparison and AI-generated explanation. Specify storage region, retention, deletion, access control and whether customer files may be used for training. Do not assume the supplier has permission to repurpose customer drawings.
Check existing CAD functions before buying duplicates. ModelCHECK and DWG Compare may already cover part of the work. AI must demonstrate extra value in difficult extraction, cross-document comparison, evidence-backed prioritization or workflow integration.
| RFP question | Evidence to request |
|---|---|
| Which formats and frames are supported? | Compatibility matrix and sample tests |
| How are model and drawing revisions linked? | Mapping and mismatch demo |
| Which alerts come from rules versus AI? | Method, version and limitations |
| Can users return to the source? | Page, coordinate and feature links |
| Who can release? | Role matrix and audit export |
| Where are files stored and deleted? | Architecture, contract and deletion method |
Prove performance at FAT and SAT
At FAT, run the contractual functions on a frozen held-out set in the vendor environment. Retain input IDs, configuration version, raw findings, scoring decisions, processing time and incident logs. Demand the underlying result list as well as a summary. A sample used for tuning is not an independent final test.
At FAT, define the scoring unit before calculating performance. A drawing-level “correct” result can conceal a missed critical tolerance if a minor formatting alert happens to fire on the same sheet. Track drawing, finding-event and feature-level results separately. Assign each predicted finding an ID and record which expected event it matches. For a revision mismatch, a useful result includes the drawing number, the source location of the displayed revision, the PDM revision, the join key and the recheck after correction, rather than a generic red warning.
At SAT, test failures that can block release: re-export the PDF after all alerts have been disposed of and verify that the old report becomes invalid; interrupt the PDM connection and verify that the screen says “unable to assess” rather than “pass”; and try the release API with a user who has no approval authority. Test customer-specific access in search, notification and export as well as in the review screen.
At SAT, run the entire workflow on the Thai site’s real CAD/PDM integration, network, permissions, frames, languages and approval roles. Create, check, resolve, approve, re-export, recheck and release. Confirm that a PDM revision change invalidates an old result, network retry avoids duplicate records and access controls protect customer drawings. SAT is operational, not merely a repeat of the FAT score.
The acceptance record should show scope, sample set, thresholds, observed results, exceptions, defects, corrective deadlines, approver and date. Keep human checking during initial operation, monitor misses and false alerts, and specify reassessment after a model or drawing-frame change.
Plan rollout and total cost
Separate spending on current-state survey, master/revision cleanup, existing CAD/PDM configuration, AI extraction, workflow/audit integration, FAT/SAT and ongoing support. Request setup and annual costs, users, volume, cloud usage and the price of each additional frame layout. A generic Thailand price cannot represent every drawing estate.
Start with master-file control and release accountability. Pilot one finding type with clear ground truth, often title-block/revision mismatch. Expand to dimension/tolerance questions, model correspondence and process capability later. Use actual drawings, approvers and the real release path even in a small pilot. The related RFQ drawing-analysis article addresses quotation, whose acceptance criteria differ from formal engineering release.
Frequently asked questions
Can AI find every missing dimension?
No complete-detection assumption is justified. Define the exact omission classes, test on held-out files, count misses and require “unable to assess” when appropriate.
Can AI approve GD&T?
It can flag symbols and some formal inconsistencies. Datum choice and tolerance values depend on function, standards and customer requirements. An authorized engineer decides.
Can we begin with PDF only?
Title-block extraction, notes and 2D revision comparison may be possible. Scans and absent native CAD limit feature mapping and semantic PMI comparison. Name the exclusions in the RFP.
How is this different from existing CAD checks?
Compare rule, geometry and visual-difference tools function by function against the proposed AI. Measure additional value on difficult site files.
Do we need both FAT and SAT?
FAT tests contracted functions and a scoring set. SAT tests integration, permissions, languages and release workflow at the actual factory. Keep evidence and unresolved items from both.
Conclusion
AI drawing review is useful when it brings likely problems to engineers early with source evidence. Define title-block and revision checks, dimensions and tolerances, model/drawing links and change comparison separately. Treat rules and AI inference honestly, and keep engineering approval with people. A defined evaluation set and FAT/SAT record support a buying decision based on your own drawings.
If you are still mapping pre-release checks at a Thai factory, contact TOMAS TECH to discuss the CAD formats, title blocks and approval path for an initial pilot.
