In heavy industry, AI does not win on its ability to generate drawings at volume. It wins on its ability to read a change request precisely, produce a revision inside an authorized scope, pass the calculations and the rules, and then commit the change safely once a person has approved it. For P&ID and fabrication drawings in particular, controlled change comes before unconstrained automatic editing.

Why Change Control Comes Before Automatic Generation

A demo that builds a new P&ID from scratch is impressive. But most of the work on a real project is applying changes to drawings and data that already exist: renaming tags, editing equipment properties, swapping symbols, adjusting a limited set of piping connections. What matters here is not whether the new picture looks plausible, but whether the existing baseline, the rules, and the related documents stay consistent.

Unconstrained generative editing creates two risks. The first is that AI touches elements it was never asked to touch. The second is output that looks visually natural while violating engineering rules. The right shape for a first product, then, is one that restricts which objects and commands are editable, makes the before-and-after comparison explicit, and never modifies the official drawing before approval.

Free-form generative drawing editing

Edits the whole drawing directly from a natural-language instruction
Objects outside the intended scope can move along with it
Symbols may look correct while breaking connectivity and property rules
The diff against the original, the rollback path, and the approval route stay unclear

Controlled Change Copilot

Generates patch candidates only for authorized objects and properties
Shows the scope, the evidence, and the impact first
Checks P&ID, data, and calculation rules deterministically
Commits to an official revision only after preview and human approval
Figure 1. A heavy-industry drawing AI is safer and more practical when it starts as a tool that “proposes verified patches” rather than a tool that “draws for you.”

Eight Stages from Change Request to Official Revision

Suppose a user asks: “Uprate the control valve on the P-2101 discharge line by one step and update the related tags.” AI does not turn that sentence straight into drawing commands. It first interprets the target objects and the scope of the request, then confirms the current baseline and the user's edit rights.

It then turns the candidate revision into a structured patch. Changes to tags, symbol properties, line relationships, and related schedules are separated out, and the calculations and rules are run. Results that pass are presented as a drawing overlay and a data diff; once the responsible engineer approves, they are written into the official file and the revision history. If anything fails, the original is left untouched.

Request ParseStructures the target, intent, and constraints from natural language, RFIs, and change notices.
Scope LockFixes the current baseline, the edit rights, and the objects and properties in scope.
Impact QueryExplores the impact on linked tags, lines, datasheets, schedules, and calculations.
Patch DraftGenerates drawing, property, and schedule changes as a machine-readable patch.
Rule ValidationChecks connectivity, tags, specifications, process, drawing rules, and calculations.
Visual PreviewShows the before/after overlay, plus deletions, additions, edits, and open items.
Human ApprovalDiscipline leads approve, amend, or reject, and the reason is recorded.
Commit & AuditApplies the official revision, raises linked tasks, and stores audit and rollback data.
Figure 2. A change should be handled not as a single prompt but as a controlled transaction: Scope–Patch–Validation–Approval–Commit.

P&ID and DXF Editing Starts Inside a Narrow Scope

The initial scope for automatic editing has to be explicit. Tag strings and properties, replacing, adding, and deleting symbols from an approved library, updating attribute blocks, a limited set of orthogonal piping connections, and references to schedules and drawing numbers — that is a realistic starting set. Free-form curves, complex layout rearrangement, and multi-discipline detail design are better left for later.

In a vector format such as DXF, each entity's layer, block, attribute, handle, and relationships have to be preserved. The editing engine should work through rule-based geometry operations and approved symbol templates, not by having AI emit arbitrary coordinates. AI proposes “what to change”; a deterministic engine performs the actual geometry edit.

Geometry·Rule Engine

Produces the same result for the same input, with standards and units controlled explicitly.

Reads DXF blocks, attributes, and layers reliably and edits only permitted entitiesApplies orthogonal routing, minimum clearance, connection point, and symbol anchor rulesChecks tag duplication, line class, direction, connectivity, and mandatory propertiesRegression-compares file integrity and rendered output before and after the patchPreserves original handles, revisions, and checksums so rollback is guaranteed

AI Change Planner

Reads the context in documents and drawings, and connects evidence and change impact in terms a person can follow.

Interprets the target objects, intent, and constraints in a change requestExplores candidate impacts on linked equipment, lines, instruments, and documentsDrafts permitted patch commands and review commentsExplains rule failures and the options for resolving them in plain languageSuggests schedule and datasheet updates the approver may have missed
Figure 3. File integrity and working flexibility hold together only when a deterministic engine owns the geometry edit and AI owns the explanation of intent and impact.

Make the Patch the Official Language

The output of AI should never be the edited file itself. What comes first is a change patch that a person can read and a system can validate. The patch carries the target object, the current value, the proposed value, the reason for the change, the related objects, the rules that apply, the expected document impact, and the approval roles.

You might define a bounded command set such as `ReplaceSymbol`, `UpdateAttribute`, `ReconnectLine`, `CreateTag`, `RetireTag`, and `UpdateListRow`. Every command has preconditions and postconditions. If the current revision and the object handle do not match, the command is not applied; once it is applied, connectivity and tag uniqueness are checked.

Controlled Patch CP-2026-118Example change: P-2101 discharge control valve
Target BaselineP&ID P-101 Rev.12 · Checksum 8fa…
ScopeLine L-2101-D · Valve CV-2103 · Instrument Loop
Patch CommandsReplaceSymbol · UpdateAttribute · ReconnectLine
PreconditionsObject handle, tag, and revision match · Edit rights valid
Engineering RulesLine Class · Flow Direction · Fail Position · Tag Uniqueness
Linked UpdatesValve List · Instrument Index · Datasheet · Cause & Effect
Risk LevelHigh · Process Control / Procurement Impact
Approval RouteProcess → Instrument → Piping → Document Control
Figure 4. The patch turns a suggestion from AI into an auditable engineering change language. The file is modified only after the patch clears every precondition and validation.

Validation Must Cover Drawing Rules and Engineering Meaning Together

Checking that the lines connect is not enough for a drawing edit. Validation has to confirm that tags follow the project naming convention, that equipment and line specifications are compatible, that valve direction and fail position match the process intent, and that the change is reflected in the Instrument Index and the Cause & Effect matrix.

The rules can be layered: file integrity rules, graphic rules, object relationship rules, discipline rules, project standards, and contract and approval rules. Each rule returns Pass, Fail, or Review Required, and a single high-risk Fail blocks the commit.

01File IntegrityChecks DXF structure, blocks, layers, attributes, and referential integrity.
02Graphic RuleChecks symbol orientation, orthogonal connections, overlaps, minimum clearance, and legibility.
03Object RuleChecks tag uniqueness, equipment–line–instrument relationships, and mandatory properties.
04Engineering RuleChecks line class and the suitability of pressure, material, valve, and instrumentation logic.
05Project RuleApplies in-house standards, project naming, and issue and revision rules.
06Approval GateConfirms the reviewers required at that risk level and any unresolved issues.
Figure 5. Only when the validation layers are separated can you say clearly whether a failure is a graphic problem or an engineering specification problem.

A Preview Is Not Just Red Markup

A good preview shows visually what was added, deleted, moved, or modified, and supplies the data diff at the same time. The drawing overlay highlights the changed objects; the right-hand panel shows property before/after, the related documents, and the rule results. The user must be able to approve or amend each change item individually.

When automatic layout kicks in, the preview has to explain why the surrounding objects moved. The arrangement AI proposed and the coordinates the engine actually applied should be distinguished, and the deviation from the original recorded numerically. The preview itself becomes evidence in the official change package.

Can this patch be committed?

Drawing connectivity and tag uniqueness pass. However, the Body Material in the Valve List remains the previous CS, while the new Line Class requires SS316. Update drafts for the Instrument Index and Cause & Effect are ready, but the commit is blocked until the piping lead confirms the material.
P&ID Diff3 objects modified · 1 reconnect
Valve List RuleMaterial conflict · FAIL
Instrument IndexUpdate candidate ready
Approval GatePiping review pending
Figure 6. A preview must show the visual diff, the property conflict, the rule results, and the approvals still outstanding — all in one place.

On-Premise Deployment, Permissions, and Audit Are Product Features

Heavy-industry drawings and specifications are sensitive project assets. The deployment model has to account for on-premise, VPC, and air-gapped environments, and the customer must control how much is ever sent to an external model. Permissions are needed at the document, object, and property level. Everyone may be able to open the P&ID without being able to see purchase prices or vendor deviations.

Requests to AI and its responses, the model and prompt policy used, the evidence retrieved, the patch generated, the rule results, the approval, and the commit all have to persist in an audit log. Past changes must remain reproducible even after the model version changes. This control layer is not an add-on feature; it is the center of the product.

Automation Level Should Vary by Object and by Risk

Running every change through the same approval path yields little automation benefit; applying everything automatically is dangerous. Low-risk annotation fixes and document reference updates can take a light-touch approval once the rules pass. A tag rename reaches across many documents, so it needs multiple reviewers. Changes to line class, design pressure, and control logic should be classified as high risk and forced through human judgment.

0Review-onlyAI presents the diff and a proposed revision but does not modify the file.
1Low-risk PatchApplies tag notation, properties, and document references automatically after approval.
2Symbol OperationSupports adding, replacing, and deleting symbols within an approved library.
3Limited ConnectivityPerforms a limited set of orthogonal piping connections and local layout on a rule basis.
4Cross-document CommitSynchronizes the change work across related schedules, datasheets, and indexes.
5Lifecycle Controlled ChangeOperates change and rollback with procurement, fabrication, and site status taken into account.
Figure 7. A drawing AI should expand in stages: review-only first, then low-risk patches, bounded geometry editing, and finally lifecycle change.

What AI Will Change Is Less the Drawing Than the Way Change Happens

To earn trust in heavy industry, AI needs less the ability to draw anything than the ability to prove what it did not change. The core is the procedure: lock the scope, build a structured patch, pass the engineering rules, and write the official revision only after a person approves.

Once that structure is in place, AI moves beyond a simple drawing editor and becomes an Engineering Change Control Plane that links change requests, impact analysis, calculation and rule validation, document synchronization, and audit history. The end state of automation is not generation but control.

doAZ Point of View

Defensibility in heavy-industry AI comes less from the model than from the combination of a Patch DSL, a DXF and drawing engine, Rule-as-Code, Preview UX, Risk-based Approval, and Audit with Rollback. Starting from bounded editing is not the slow route; it is the fastest route to an actual product.

YT
Youngtae Kim · CEO · doAZ

Working across construction and engineering sites, R&D, and the industrial AI business, and thinking about how to connect the knowledge scattered through drawings, documents, and calculation sheets. This series is a public record of the product philosophy and field direction doAZ is pursuing.