AI welding robot implementation is moving the project question from “Who will teach every path?” to “Who will validate and approve the welding parameters and motion generated by AI?” A drawing-to-program workflow can shorten the route to a first proposal, but it does not automatically establish weld quality, cell safety, or customer acceptance. This guide shows Thailand and ASEAN factories how to structure a five-stage PoC, supplier RFP, FAT/SAT and rollback around controlled approval gates.
Why the September 2026 AI Welding Agent announcement matters
On 11 September 2026, FANUC announced an AI Welding Agent developed with Google. According to the company, the system reads a drawing for an arc-welded component, understands the material and final shape, and generates welding parameters such as current and voltage together with robot motion programs. Operators can execute the generated result as-is or fine-tune it before welding. Shipments are scheduled to begin by the end of December 2026.
This changes the entry point for automation. Traditional deployment spreads engineering time across workpiece orientation, coordinate registration, path teaching, parameter adjustment, trial welding and correction. A drawing-to-program workflow may provide a much faster first proposal, especially in high-mix, low-volume production.
However, “zero setup” and “zero teaching” are product-announcement terms. They must not be generalized into a promise that every joint, material, thickness, fixture and workpiece variation can run without validation or approval. The sources reviewed for this article do not confirm price, Thailand availability, supported drawing formats, material and thickness limits, certification, accuracy, or a cycle-time improvement percentage.
Generation does not transfer manufacturing accountability
The output is not merely text. Current, voltage, speed, torch attitude, aim point, approach, retreat and recovery instructions directly affect the physical process. An incorrect result can contribute to lack of fusion, undercut or distortion, but also collision, spatter, fire, fume, arc radiation, contact with a hot workpiece, and unexpected restart.
The control objective is therefore not to label the AI “accurate” in the abstract. It is to establish who owns the input, who reviews each class of output, what must be tested, who releases the approved version, and how evidence is retained. Production release should be based on an approved drawing, applicable WPS or other welding-procedure controls, equipment configuration, test results, safety assessment and customer requirements.
Govern the full drawing-to-welding-program chain
A drawing rarely contains every manufacturing condition
A weld symbol does not necessarily provide the base-metal specification, tolerance, joint preparation, root gap, surface condition, tack sequence, shielding gas, wire, position, preheat, interpass temperature, allowed distortion, inspection method, applicable WPS or customer-specific requirement. Many factories hold these in separate controlled documents.
The PoC should begin with a defined input package rather than the instruction “show the AI a drawing.” Control at least the following:
- drawing number, revision, approval status and language;
- relationship to a 3D model and bill of materials, when used;
- material, thickness, joint type, welding position and weld length;
- applicable WPS or preliminary pWPS and its qualification range;
- power source, torch, wire feed, gas, fixture and positioner configuration;
- expected workpiece tolerance, positional variation and thermal distortion;
- quality level, inspection method, sampling and acceptance criteria;
- prohibited robot-cell zones, obstacles, access and recovery conditions.
When an input is missing, the controlled response should be “generation blocked,” “human decision required,” or “additional data requested.” Letting the system silently infer a production-critical value weakens traceability. Any human-supplied value should record its owner, rationale and time.
Review output in three layers
| Layer | Typical content | Primary reviewer | Approval question |
|---|---|---|---|
| Welding process | Current, voltage, speed, heat-related settings, sequence | Welding coordinator and quality | Is it within the WPS or pWPS basis and suitable for the joint? |
| Robot motion | TCP, attitude, path, approach, retreat, standby point | Robot engineer and integrator | Are collision, singularity, reach and cable posture controlled? |
| Cell control | Start conditions, interlocks, fault response, restart and version | Controls, safety and production engineering | Can it bypass a safety function or permit an unintended restart? |
This separation matters because one edit can cross boundaries. A welding-speed change can affect process quality and interaction timing. A retreat-point edit may not change the bead but can change collision or cable risk.
Make human review substantive
The review package should show more than an APPROVE button. Present the difference from the approved baseline, drawing revision, WPS or pWPS reference, robot and tool configuration, generation time, service or model version, warnings, simulation evidence and reason for change.
Choose approvers by competence and authority, not just job title. A welding coordinator, robot engineer and machinery-safety owner may need separate approvals. Define deputies and after-hours escalation without allowing an unauthorized production release.
A five-stage welding automation PoC
The goal is not to prove that one attractive weld can be produced once. The PoC must show that complete inputs lead to a reviewable output, the result can be qualified on coupons, quality and safety can be evidenced separately, and a production version can be changed and restored under control.
Stage 1: Freeze the joint, drawing and WPS data envelope
Start with one or two representative joint families whose boundaries are clear. Fix material, thickness, position, joint preparation, weld length, fixture, power source, torch, wire and gas, then state which variables the trial will explore. A very easy sample provides little evidence; the whole product range prevents useful root-cause analysis.
Create a matrix of what is read from the drawing and what must arrive from another source. For every manually entered value, define the screen, unit, allowed range and validation. Because the reviewed announcement does not confirm supported file formats, the RFP should request a vendor answer instead of assuming support for PDF, image or CAD data.
The exit criterion is an approved input-package revision, out-of-scope list, stop conditions for missing information, and named owners.
Stage 2: Contain and review the generated output
Do not send a newly generated program directly to the production controller. Place it in a segregated validation state. Perform a static review of welding parameters against the procedure basis; verify coordinate frames and TCP; inspect approach, retreat, air moves, arc start and finish, fault response and restart position.
Then use offline simulation or a controlled dry run to check collision, joint limits, singularities, cable twist, fixture clearance and speed transitions. A model is not the physical cell, so retain the simulation model revision, fixture revision and TCP calibration date.
Assign a unique ID to the generated result and bind it to the exact input revision. If a reviewer edits it, preserve the original output and record the difference and reason. This enables later investigation of whether the AI output, stale input, or human correction caused a problem.
Stage 3: Qualify on a traceable coupon
Before production workpieces, weld a representative coupon. Control the material certificate, dimensions, preparation, tack condition, position, fixture, program version, equipment version and operator. Define the visual, dimensional, non-destructive or destructive examinations required by the product, contract and applicable code.
ISO 15614-1:2017 provides a procedure-test framework for qualifying welding procedures for the metals and processes within its scope. It is not a standard for automatic approval of AI output. Where the project applies it, manage the AI proposal through a pWPS, coupon, test result and qualified range, along with customer-specific requirements. The ISO page indicates that the edition was reviewed and confirmed in 2022 and that replacement work exists; do not predict a publication date. Check the current edition at RFP and approval time.
Stage 4: Validate quality and safety with separate evidence
A conforming bead does not prove that the cell is safe. A robot meeting its own safety requirements does not prove that the weld meets product requirements. Keep the evidence streams separate and join them only at the release decision.
ISO 5817:2023 defines imperfection quality levels B, C and D for fusion-welded joints within its stated materials and processes, with B representing the highest requirement on the finished weld. These levels do not by themselves establish fitness for service or create an automatic accept/reject algorithm. The designer must consider load, fatigue, corrosion, product standards, customer requirements and inspection method. Beam welding is excluded from its scope, so the standard should not be casually applied to the separate laser-welding system announced by FANUC.
ISO 3834-1:2021 provides criteria for selecting the appropriate level of quality requirements in the ISO 3834 series. It can apply in workshops and at field installation sites, but it is not a complete quality management system. The official page currently shows it under systematic review, so confirm the applicable edition and contract requirements.
For robot safety, ISO 10218-1:2025 addresses the industrial robot itself before system integration. ISO 10218-2:2025 addresses robot applications and cells, including integration, commissioning, operation and maintenance. The ISO 10218-1 description explicitly notes that application hazards created by welding, laser cutting or machining are addressed during application design. ISO 10218 alone should not be treated as a complete control for fume, arc radiation, sparks, hot material, shielding gas, electricity and fire. Combine the cell risk assessment with local law, ventilation, screening, PPE, hot-work controls, energy isolation and training.
| Gate | Example evidence | Decision owner | If failed |
|---|---|---|---|
| Weld quality | Program version, coupon ID, inspection record, WPS reference | Welding coordinator and quality | Correct settings, retest, revise the applicable range |
| Robot motion | Simulation, dry run, collision and axis checks | Robot engineer and integrator | Revise path, speed or fixture |
| Cell safety | Risk assessment, safety-function tests, stop and recovery | Safety owner and production engineering | Block release, redesign and revalidate |
| Production readiness | Work standard, training, maintenance, spare parts, backup | Plant manager, production and maintenance | Conditional acceptance or delayed release |
Stage 5: Exercise release, change control and rollback
Run the actual promotion process from validated candidate to production version. Verify who electronically signs, which device and account can transfer a program, whether an unapproved file is blocked, how the running version is identified, and whether the approved backup can be restored.
Introduce one simulated change: a drawing revision, thickness change, fixture adjustment, TCP recalibration, power-source replacement or software update. Run impact assessment, regeneration, difference review, required retest, approval, release and rollback. A PoC without this exercise demonstrates only the first success and ignores the changes that dominate the production lifecycle.
Turn PoC acceptance into measurable criteria
There is no need to invent an accuracy or ROI percentage. Define numbers as acceptance criteria measured on the factory’s own workpieces.
Define the denominator before discussing a rate
For a “generation success rate,” state whether the denominator is drawings, joints, weld seams or programs. A drawing with ten seams and nine generated correctly can look like zero or 100 percent at drawing level, but 90 percent at seam level. The key is not the attractive rate; it is a stable decision unit.
A conservative engineering-time comparison can use an explicit expression:
Preparation-hour difference = conventional teaching, setup and review hours minus AI-input preparation, review and correction hours
Coupon welding, inspection, approval and safety validation belong in both cases. Removing validation only from the AI scenario would overstate the benefit. Use the same workpiece, acceptance criteria and skill level, and separate first-attempt data from repeat trials.
Recommended PoC scorecard
| Dimension | Example measure | Boundary to define first |
|---|---|---|
| Input completeness | Required fields present, missing-input detection | No silent inference of critical values |
| Generated review | Corrections, reasons and approval time | Unit is joint, seam or program |
| Quality | Specified inspection, variation, repeat trial | WPS, product code and customer criteria |
| Motion | Collision, limits, TCP, air move and recovery | Separate simulation from physical test |
| Safety | Functions, faults, stop, restart and access | Separate robot, cell and process hazards |
| Governance | Version trace, privilege, audit log and restore | Rollback test is mandatory |
| Operations | Multilingual instruction, training and support | Validate on a Thai production shift |
Useful release criteria include zero unresolved items involving a significant hazard, a matching version ID or hash between approved and running programs, and completion of every specified qualification test. Product-specific quality limits must come from the applicable engineering and contractual basis, not from a generic blog figure.
Put AI welding deliverables into the RFP
“Automatic teaching from drawings” is not an auditable purchase requirement. Separate functional scope, data, safety, quality, operations and commercial obligations so the vendor cannot answer every line with a single “supported.”
Functional and scope questions
- Supported drawing input, image quality, language, units, weld symbols and exception handling.
- Supported materials, thicknesses, joints, positions, processes, power sources, robots and external axes.
- Information generated beyond current, voltage, speed and motion; information that a person must enter.
- Conditions that block, warn or request confirmation for missing, inconsistent or out-of-range inputs.
- Fine-tuning scope, privileges, difference display and reapproval triggers.
- Offline validation, collision checking, dry-run and controller-transfer methods.
FANUC says the announced system uses the built-in camera on the CRX tablet teach pendant, needs no additional camera or dedicated site device, and is independent of a particular welding power source connected to a FANUC robot. Treat these as project verification points. Confirm the actual interface, signals, allocation of responsibility and retrofit conditions.
Data, AI and security questions
FANUC states that Gemini Enterprise security protects drawings and production data and that the information is not used to train other users’ AI models. Procurement should turn that statement into contract questions: storage location, retention, transfer, encryption, administrator access, support access, logs, deletion, backup, outage behavior and contract-exit handling.
Ask what happens when the AI service or model changes. Can a version be pinned? Is an update announced? Will identical input produce a reproducible output? What revalidation is required? Also define operation during cloud outages, local execution of approved programs and post-recovery synchronization.
FANUC’s May 2026 Physical AI announcement describes open-platform support including ROS drivers, Python, high-speed external control and PLC interfaces. This background does not prove the AI Welding Agent’s complete API scope. MES, document control, quality database and equipment-history integrations must be verified against the purchased configuration.
Required RFP deliverables
| Deliverable | Content | Due point |
|---|---|---|
| Applicability matrix | Drawing, material, thickness, joint, position, source and robot | Proposal and design freeze |
| Input/output data dictionary | Required field, unit, range, missing and exception behavior | Basic design |
| Generated-review specification | Difference, warning, privilege, approval record and signature | Basic design and pre-FAT |
| Quality plan | WPS/pWPS, coupon, inspection, acceptance and retest | Before PoC |
| Safety file | Cell boundary, risk assessment, safety functions, welding controls | Design review and FAT |
| Software bill of versions | AI service, robot, PLC, source, settings and dependencies | Every release |
| FAT/SAT plan | Condition, expected result, measurement, evidence and witness | Before each test |
| Backup procedure | Approved version, restore, integrity and rollback | Pre-FAT |
| Change procedure | Impact, regeneration, retest, approval and notice | Pre-FAT |
| Training and maintenance | Multilingual instructions, privileges, faults and support | Pre-SAT |
For cost and production-type selection, see our 2026 welding robot implementation guide. For supplier capability and responsibility, use the robot SIer selection guide alongside this evidence-focused RFP.
FAT establishes the approved baseline
Freeze inputs and software before FAT
The FAT package should freeze the drawing, WPS or pWPS, fixture, TCP, power settings, robot software, PLC software, AI service or model identifier, reviewed program and inspection plan. If anything changes, issue a change record, assess impact and update the test scope.
Every FAT test row should contain a requirement ID, input revision, program revision, condition, expected result, measured result, evidence file, pass/fail, witness and nonconformity number. “Machine ran OK” and an unindexed video are not enough for a later drawing change or investigation.
Four FAT scenarios
- Normal: generate from approved inputs, review, weld the coupon, pass inspection and create a release candidate.
- Invalid input: use a mismatched drawing revision, missing critical value or out-of-range condition and confirm the workflow stops.
- Equipment fault: test communications loss, source fault, sensor fault, interrupted welding, safe stop and controlled recovery.
- Change and restore: edit one approved item, require reapproval, restore the baseline and verify integrity.
The ISO 10218 robot safety-fence guide is useful for access protection. In a welding cell, also include process hazards such as fume, radiation, spatter, hot material, gas and electrical energy.
SAT closes the Thailand-site delta
Transport, installation, utilities, grounding, gas, ventilation, adjacent equipment, network, floor, lighting, work flow and local fixtures can change the conditions after FAT. SAT should start with the delta from the FAT baseline and repeat all quality, motion and safety tests affected by that delta.
Operating conditions often missed in ASEAN projects
- Do Japanese drawings and English or Thai work instructions identify the same approved revision?
- Does an absent approver on night shift prevent an unauthorized release?
- Can the cell continue safely on an approved version or stop cleanly during a network outage?
- Are heat, humidity, dust and power quality within the equipment specification?
- Do locally sourced wire, gas and consumables match the approved condition?
- Are contractor accounts, remote support and removable media controlled?
- Can operators understand fault and recovery instructions in the working language?
SAT exit means every delta is assessed, affected tests are complete, open items have an owner and due date, and the running program matches the approved backup.
Production change control and rollback
Define reapproval triggers before production
Control the drawing, material, thickness, joint, WPS/pWPS, fixture, TCP, wire, gas, source, robot, PLC, safety setting, AI service, network, inspection method and work instruction. Classify changes by their quality, motion, safety and data impact instead of treating all edits alike.
| Change | Minimum impact review | Candidate retest |
|---|---|---|
| Drawing, thickness or joint | Procedure range, path, fixture and inspection | Regenerate, review, coupon and quality tests |
| TCP or fixture | Aim, collision, posture and repeatability | Calibrate, dry-run and representative weld |
| Power source, wire or gas | Parameter compatibility, quality and fault signal | Coupon and interlock test |
| Robot or PLC update | Motion, safety functions, communications and recovery | Regression and safety validation |
| AI service update | Output difference, warning, reproducibility and log | Golden-input comparison and reapproval |
Roll back the configuration, not just one file
The rollback package includes the program and its drawing, procedure, PLC version, source settings, fixture revision, TCP, inspection condition and work instruction. Restoring only the robot file can produce a dangerous mixed configuration.
After restoration, verify version ID or hash, parameters, communication, coordinates, interlocks and representative motion. Record privilege, dual approval, reason, affected lots and quarantine scope. A live restoration test during PoC and FAT proves that the backup is usable.
Decisions for management and procurement
An AI welding demonstration can be compelling enough that an investment discussion narrows too quickly to feature names and indicative price. Management and procurement should instead lock down the application envelope, required evidence, assignment of responsibility, cost and lead-time consequences of changes, and the operating method during a service outage.
Example Go, Hold and Stop decisions
- Go: The application envelope, input ownership, review authority, test plan, safety measures and version controls are defined, and the representative condition is repeatable.
- Hold: A single weld succeeded, but evidence for missing inputs, out-of-range conditions, change, failure recovery or restoration is incomplete.
- Stop: An unapproved output can reach production, the active version cannot be traced, a safety function can be bypassed, or the quality basis cannot be explained.
Before investment approval, also define the unit of production expansion after the PoC. The project must state whether one qualification and change-control package covers a product family, joint family, cell or factory. An unclear deployment unit creates two opposite risks: extending the PoC evidence too broadly or repeating the whole test for every workpiece without reason.
FAQ about AI welding agents and automatic teaching
What is an AI welding agent?
The current example is FANUC’s announced system that reads a drawing and generates welding current, voltage and robot motion. It is an example of Physical AI connecting generative AI to physical robot action. Do not treat one product announcement as proof of every AI welding system’s capability.
Does automatic teaching remove human approval?
No. A competent person still needs to verify the procedure range, physical path, qualification evidence, safety functions and release version. AI can support proposal generation; it does not automatically assume manufacturing accountability.
Can a drawing-to-welding-program workflow go straight to production?
Usually not. Critical information may reside outside the drawing. Confirm complete inputs, review the output, perform a dry run, qualify on a coupon, validate quality and safety, then release a controlled version.
Does ISO 5817 quality level B guarantee fitness for service?
No. B, C and D are imperfection quality levels within the standard’s scope, with B the highest requirement. Product use, loads, fatigue, corrosion, customer criteria and inspection methods still determine acceptance.
Does ISO 15614-1 automatically approve AI-generated settings?
No. It provides a procedure-test qualification framework within its scope. Manage an AI proposal through the applicable pWPS, coupon, testing and qualified range, and confirm the current edition and contract basis.
Does ISO 10218 cover every welding-cell hazard?
No. Part 1 concerns the robot and Part 2 concerns the application and cell. Welding-specific hazards such as fume, radiation, sparks, hot material, gas, electricity and fire require additional assessment and controls.
What should a welding automation PoC accept?
It should demonstrate missing-input detection, reviewable differences, coupon acceptance, quality and safety evidence, controlled release, version traceability and tested restore. A single good-looking bead is insufficient.
What should an AI welding robot RFP compare besides price?
Compare the application envelope, behavior with missing inputs, review and approval controls, data protection, version pinning, test deliverables, FAT/SAT, training, maintenance, change control and rollback. For specifications not confirmed in public information, require the supplier to state applicability, constraints and exclusions rather than relying on assumptions.
Will the planned December 2026 shipment mean immediate availability in Thailand?
The announcement states that shipments are scheduled to begin by the end of December 2026. It does not confirm Thailand availability, local price, contract terms or support timing. Ask the authorized local channel about the exact configuration.
Conclusion: AI generates; the approval gates carry accountability
Drawing-to-program generation can be highly valuable for high-mix work and scarce welding expertise. Production value comes from more than the “zero teaching” entry point. It depends on a controlled input envelope, output review, coupon qualification, separate quality and safety evidence, a FAT/SAT baseline, change control and rollback. These gates convert generated output into a defensible manufacturing asset.
TOMAS TECH can help at the exploration stage, before any product decision, by structuring workpiece data, a welding-automation PoC scorecard, RFP deliverables, FAT/SAT evidence and IT/OT version controls for a Thailand plant. Contact us to discuss an AI welding robot implementation with your current drawing, procedure, equipment and quality constraints.
Primary references
Checked 15 September 2026. Reconfirm product availability, specifications and standards status at the point of decision.
- FANUC: AI Welding Agent announcement, 11 September 2026
- FANUC: Physical AI collaboration with Google, 13 May 2026
- FANUC: CRX pipe laser welding system, 11 September 2026
- ISO 10218-1:2025 official page
- ISO 10218-2:2025 official page
- ISO 5817:2023 official page
- ISO 3834-1:2021 official page
- ISO 15614-1:2017 official page
*This article provides general technical and project-management information. It is not legal advice, certification advice or a guarantee of product conformity. Confirm applicable law, customer specifications, welding codes and safety requirements with competent authorities and qualified professionals.*
