The argument against AI in WPS authoring is straightforward: welding codes are too specific, too consequential, and too constrained for AI to do correctly. The argument is partly right. AI cannot replace a CWI's judgment on essential variables, joint suitability, or filler matching.

But that argument misses where AI actually fits — and what it eliminates from the WPS-authoring workflow.

What AI is good at

The parts of WPS authoring that already feel mechanical:

  1. Boilerplate narrative. "Manual SMAW with low-hydrogen E7018 electrode on ASTM A36 carbon steel..." — this paragraph is almost identical across hundreds of WPSs. AI generates it in seconds.
  2. Process descriptions. Every SMAW WPS describes shielded metal arc welding the same way. AI handles the descriptive blocks.
  3. Joint preparation notes. "Mill scale removed within 1 in of weld; rust, oil, paint removed within same band." — same on every WPS. AI fills it in.
  4. Storage rules for low-hydrogen filler. "Rod oven at 250°F minimum after seal break; maximum 4-hour exposure before redrying." — boilerplate.
  5. PWHT method descriptions when applicable.
  6. CVN supplementary note language when invoking Table 6.8.

Each of these is descriptive content that follows a pattern. AI generates the pattern correctly because the pattern is well-defined.

What AI is NOT good at

The parts that require code-table values or engineering judgment:

  1. Specific amperage / voltage / travel-speed ranges. These come from the supporting PQR or from prequalified Clause 5 limits. AI must not invent them.
  2. Filler classification matching. "What's the matched filler for A572-50?" requires Table 5.4. The rule engine answers this, not the AI.
  3. Preheat minimums. Table 5.8 by material × thickness × hydrogen. Rule engine.
  4. Essential variable cross-check vs PQR. Table 6.6 row-by-row. Rule engine.
  5. CVN supplementary requalification triggers. Table 6.8. Rule engine.
  6. Joint suitability decisions. Whether B-U2 with backing is right for this application. CWI judgment.

The architecture that works

A WPS tool that combines AI and a rule engine has a clean separation:

  • Rule engine owns every code-table value. Tables 5.4, 5.8, 6.6, 6.7, 6.8, 6.9, Annex B. Deterministic output.
  • AI owns narrative drafting. Descriptive paragraphs, technique notes, process explanations.
  • CWI owns final review and signature. Reviews the rule-engine-validated WPS and signs.

Code numbers come from the rule engine. Narrative comes from the AI. Approval comes from the CWI. No layer steps on the others.

What this changes in practice

A typical SMAW WPS for A36, hand-written:

  • Pull template (5 min)
  • Fill in process, joint, base metal blocks (15 min)
  • Cross-reference Tables 5.4, 5.8 for filler match and preheat (10 min)
  • Cross-reference Table 6.6 to confirm essentials (10 min)
  • Write narrative paragraphs (30 min)
  • Internal review (10 min)
  • CWI review (10 min)
  • Total: ~90 minutes

Same WPS with AI + rule engine:

  • Enter inputs in the form (3 min)
  • Rule engine fills in code-table values automatically
  • AI drafts narrative with inline code citations (15 seconds)
  • CWI reviews the validated WPS (5-10 min)
  • Total: ~10 minutes

The time savings are in the parts that were already mechanical — the table lookups and the boilerplate paragraphs. The engineering judgment time is unchanged.

What doesn't change

  • The CWI's responsibility for the document
  • The signature discipline (every revision re-signed)
  • The PQR archive and the audit trail
  • The code edition compliance burden
  • The need for a second pair of eyes on novel procedures

A good AI WPS tool feels less like "the AI wrote my WPS" and more like "the typing happened by itself."

The audit perspective

Auditors don't care how a WPS was drafted. They care whether it's compliant with the cited code. A WPS produced by AI + rule engine that passes code compliance is no different at audit from a hand-written one that passes. The compliance bar is the same.

In our experience, AI + rule-engine-produced WPSs are typically more compliant than hand-written ones because the rule engine catches the common deficiencies (single-value parameters, missing hydrogen designators, table-number drift) before the WPS ever reaches the CWI's desk.