Production Specification Integrity Governance requires that every manufacturing action executed or directed by an AI agent conforms to the current, approved, and version-controlled production specification for the product, process, or assembly in question. Agents operating in factory environments — controlling CNC machines, robotic welders, mixing systems, pick-and-place units, or inspection stations — must resolve and validate the authoritative specification before initiating any production step, and must halt or escalate when a specification cannot be confirmed, has expired, or conflicts with the physical configuration of the production line. This dimension prevents a class of failure in which agents execute manufacturing operations against outdated, superseded, draft, or incorrectly referenced specifications, producing parts or products that deviate from engineering intent and may be unsafe, non-compliant, or non-functional.
Scenario A — Robotic Welder Executes Superseded Weld Specification: An automotive chassis assembly plant operates 14 robotic welding cells controlled by an AI scheduling and parameter agent. On 3 March, the engineering team releases revision D of weld specification WS-4417, increasing the minimum weld nugget diameter from 5.0 mm to 5.5 mm for a structural cross-member joint, following a field failure analysis that identified fatigue cracking in vehicles exceeding 120,000 km. The engineering change order (ECO-2891) is approved and uploaded to the product lifecycle management (PLM) system at 09:14. However, the agent's local specification cache refreshes on a 24-hour cycle. Between 09:14 and the next cache refresh at 02:00 the following day, the agent continues to execute welding parameters from revision C across all 14 cells. In that 16-hour window, the plant produces 1,340 chassis subassemblies with the old 5.0 mm nugget diameter. Quality inspection — which also references the cached specification — passes all 1,340 units because they conform to revision C. The discrepancy is discovered 11 days later during a random PLM audit. By that time, 890 of the affected subassemblies have been integrated into complete vehicles, 214 of which have already shipped to dealerships.
What went wrong: The agent did not validate specification currency against the authoritative source at the point of execution. A 24-hour cache refresh cycle created a window in which the agent operated against a superseded specification without awareness. No mechanism existed to push specification changes to the agent in real time or to block production when a pending ECO existed for the active specification. The quality inspection agent referenced the same stale cache, eliminating the secondary check. Consequence: 1,340 subassemblies produced to the wrong specification, requiring rework at an estimated cost of £2.1 million. 214 shipped vehicles require dealer-level remediation at £680 per vehicle (£145,520 total). The OEM customer issues a formal quality escape notification and places the supplier on probationary status, jeopardising £34 million in annual contract value.
Scenario B — Pharmaceutical Mixing Agent Uses Draft Specification: A contract pharmaceutical manufacturer operates a continuous mixing line for a generic oral solid dosage product. The agent controlling the mixing parameters — impeller speed, granulation fluid addition rate, blend time, and endpoint torque — retrieves its specification from a shared document repository. On 12 June, a formulation scientist uploads a draft revision of the mixing specification (MS-0093-v4-DRAFT) to evaluate the impact of a 12% reduction in granulation fluid volume on tablet hardness. The draft is saved to the same repository folder as the approved specification (MS-0093-v3-APPROVED). The agent's file selection logic sorts by modification date and selects the most recently modified file. At 06:00 on 13 June, the agent begins a production batch using the draft specification. The reduced granulation fluid volume produces tablets with a mean hardness of 22 kP — above the upper specification limit of 20 kP — and a dissolution rate 18% slower than the approved dissolution profile. The deviation is detected at the in-process control checkpoint after 4,200 kg of granulate has been processed. The entire batch — valued at £310,000 in raw materials and £890,000 at finished goods value — must be rejected because the product cannot meet the registered dissolution specification.
What went wrong: The agent selected a specification file based on modification date rather than approval status. No metadata check distinguished draft from approved documents. The draft specification was stored in the same location as the approved specification with no access control preventing the agent from reading it. No pre-execution validation confirmed that the specification carried an approved status flag. Consequence: £890,000 batch rejection, 3-day production line shutdown for investigation and re-qualification, potential regulatory inspection triggered by the batch deviation report, and a CAPA (corrective and preventive action) filing with the national medicines authority.
Scenario C — CNC Agent Applies Specification for Wrong Part Variant: A precision aerospace machining facility produces turbine blade root fittings in three variants: standard (P/N TB-7700-A), high-temperature (P/N TB-7700-B), and corrosion-resistant (P/N TB-7700-C). The variants share 87% of their machining operations but differ in fillet radii, surface finish requirements, and final dimensional tolerances. An AI agent manages the CNC programme selection and tool path generation. A work order for 48 units of TB-7700-B is released to the shop floor. The agent's part-number-to-specification mapping table contains a data entry error: TB-7700-B is mapped to specification SP-7700-A (the standard variant) instead of SP-7700-B. The agent loads the standard variant programme and begins machining. The fillet radius on the blade root is machined to 1.2 mm (standard specification) instead of 1.5 mm (high-temperature specification). The 0.3 mm difference is within the CNC machine's normal operating envelope and does not trigger any machine-level alarm. All 48 parts pass dimensional inspection against the wrong specification. The error is discovered during the customer's incoming inspection when the parts fail the high-temperature fatigue test. All 48 parts are scrapped.
What went wrong: The agent did not independently verify that the specification it retrieved matched the part number on the work order through a cross-reference check against the engineering bill of materials. A single-point-of-failure mapping table was the sole link between the part number and the specification. No secondary validation — such as comparing critical dimensions from the specification against the part number's engineering drawing — existed. Consequence: 48 scrapped turbine blade root fittings at £4,200 each (£201,600 total), 6-week delivery delay to the engine manufacturer triggering a £175,000 liquidated damages clause, and mandatory re-qualification of the mapping table at a cost of £38,000 in engineering hours.
Scope: This dimension applies to every AI agent that initiates, controls, directs, or parameterises a manufacturing operation — including but not limited to: machining, welding, assembly, mixing, forming, coating, heat treatment, inspection, packaging, and labelling. The scope covers agents that select production programmes, set machine parameters, generate tool paths, control recipe execution, or determine inspection criteria. The scope extends to agents that operate in supervisory roles — scheduling production runs, assigning work orders to production cells, or configuring production lines for changeovers — where the agent's decisions determine which specification is applied to which production run. The scope includes both discrete manufacturing (individual parts and assemblies) and process manufacturing (continuous and batch chemical, pharmaceutical, and food production). The scope is independent of the agent's level of autonomy: whether the agent acts directly on machines or recommends parameters to a human operator, the specification integrity requirement applies because an incorrect specification recommendation is as harmful as an incorrect specification execution.
4.1. A conforming system MUST resolve the authoritative, currently approved production specification for the product, process, or assembly before initiating any manufacturing operation, using a specification retrieval mechanism that queries the system of record — the PLM, ERP, MES, or equivalent authoritative repository — at the point of execution, not from a local cache or static file unless that cache has been validated as current within a defined and documented staleness threshold.
4.2. A conforming system MUST validate that the retrieved specification carries an approval status of "released" or "approved" (or the equivalent status designation in the organisation's document control system) and MUST reject any specification with a status of "draft," "pending approval," "under review," "obsolete," or "superseded."
4.3. A conforming system MUST verify that the specification's part number, product identifier, or recipe code matches the work order, batch record, or production schedule entry that triggered the manufacturing operation, using at least one independent cross-reference source (e.g., engineering bill of materials, master recipe register, or product configuration database) in addition to the primary specification retrieval.
4.4. A conforming system MUST halt the manufacturing operation and generate a human-escalation event — per AG-019 — when any of the following conditions are detected: (a) the specification cannot be retrieved from the authoritative source; (b) the specification's approval status is not "released" or "approved"; (c) the specification's part number does not match the work order; (d) a pending engineering change order exists for the specification that has been approved but not yet incorporated into the released revision; or (e) the specification revision is older than the most recent released revision in the system of record.
4.5. A conforming system MUST log every specification retrieval event with the following minimum fields: timestamp, specification identifier, revision number, approval status at time of retrieval, source system queried, work order or batch record reference, agent identity, and the result of the validation (pass, fail, or escalation). These logs MUST be immutable per AG-055 and retained per the organisation's regulatory retention requirements.
4.6. A conforming system MUST propagate specification changes to all active agents and production cells within a maximum latency defined by the organisation's risk assessment, and this maximum latency MUST be documented and justified. For safety-critical or regulated products, the maximum propagation latency SHOULD NOT exceed the cycle time of the fastest production operation that the specification governs.
4.7. A conforming system MUST implement a specification version lock for the duration of a production run or batch, such that a specification change released during an in-progress run does not silently alter the parameters of that run. When a mid-run specification change is detected, the agent MUST escalate to a human authority who decides whether to complete the run under the original specification, halt and restart under the new specification, or implement a controlled transition.
4.8. A conforming system SHOULD implement cryptographic integrity verification (e.g., digital signature or hash comparison) on specification documents to detect unauthorised modification, corruption, or tampering between the system of record and the agent's execution environment.
4.9. A conforming system SHOULD maintain a specification applicability matrix that maps each production cell, machine, or process unit to the specifications it is authorised to execute, preventing an agent from applying a specification to a machine or process for which it was not validated.
4.10. A conforming system SHOULD perform a pre-execution parameter range check that compares the specification's critical parameters against the physical capability envelope of the target machine or process unit, escalating when a parameter falls outside the machine's validated operating range.
4.11. A conforming system MAY implement automated reconciliation between the specification's bill of materials and the physical materials staged at the production cell (via barcode, RFID, or vision system), halting production when a material mismatch is detected.
4.12. A conforming system MAY subscribe to real-time change notification feeds from the PLM or document control system to receive push notifications of specification changes, supplementing periodic cache validation with event-driven updates.
Manufacturing specification integrity is the foundational control that connects engineering intent to physical production. Every manufactured product is the physical instantiation of a specification — a set of dimensions, tolerances, materials, process parameters, and inspection criteria that define what the product is and how it must be made. When an AI agent executes a manufacturing operation against the wrong specification — whether because the specification is outdated, misidentified, unapproved, or corrupted — the resulting product deviates from engineering intent. The deviation may be invisible to in-process inspection (if inspection also references the wrong specification), invisible to the agent (which executed the specification it was given), and invisible to the operator (who trusts the agent's specification retrieval). The deviation becomes visible only when the product fails in the field, fails at the customer's incoming inspection, or is caught by an independent audit — by which time hundreds or thousands of non-conforming units may have been produced.
The risk is amplified by two characteristics of AI-controlled manufacturing. First, speed and scale: an agent controlling multiple production cells can propagate a specification error across an entire production line within minutes, producing non-conforming output at a rate that would have been physically impossible in a manually controlled environment. A human operator referencing a paper drawing might produce 10 non-conforming parts before the error is caught; an agent referencing a cached specification can produce 10,000. Second, trust asymmetry: operators and quality inspectors tend to trust agent-selected specifications because the agent is perceived as systematic and error-free. The manual double-check that a human operator might perform — comparing the drawing number to the work order, checking the revision date — is often omitted when the agent has already performed the selection. This trust asymmetry means that the secondary human verification layer, which historically caught specification errors, is degraded in agent-controlled environments.
The regulatory environment reinforces the criticality of specification integrity. In aerospace, AS9100D clause 8.5.1 requires that production processes be carried out under controlled conditions, including the availability of documented information that defines the characteristics of the products. In automotive, IATF 16949 clause 8.5.1.1 requires control plans that reference the applicable specifications, and clause 8.7.1.4 requires that reworked product be re-verified against the original specification. In pharmaceuticals, 21 CFR Part 211 (cGMP) requires that production and process control procedures include the full written instructions for production and process control, and that batch records reference the master production record. In all these frameworks, producing against an unapproved or outdated specification is a regulatory non-conformance that can trigger warning letters, consent decrees, or production shutdowns.
The cost structure of specification integrity failures is characteristically asymmetric. The cost of preventing the failure — real-time specification validation, cross-reference checks, version locking — is measured in milliseconds of latency and modest integration engineering. The cost of the failure itself is measured in scrapped batches, reworked assemblies, field recalls, customer penalties, and regulatory sanctions. The scenarios in Section 3 illustrate typical cost ratios: the welding specification failure produced a £2.1 million rework cost that a real-time PLM query — costing negligible compute — would have prevented. The pharmaceutical batch rejection destroyed £890,000 in finished goods value that a document status check — a single metadata field — would have prevented. Prevention is not merely cheaper than remediation; it is cheaper by two to three orders of magnitude.
Production Specification Integrity Governance requires integration between the agent's execution logic and the organisation's document control, PLM, and MES systems. The core architectural principle is that the agent must never rely on a local, unvalidated copy of a specification when an authoritative source is available, and must always confirm the specification's identity, version, and approval status before executing any production step.
Recommended patterns:
Anti-patterns to avoid:
Aerospace. AS9100D and NADCAP accreditation require full traceability of the specification applied to each part. Agents operating in aerospace manufacturing must maintain lot-level traceability linking each serialised part to the specification revision used in its production. Specification integrity failures in aerospace can result in airworthiness directive actions, mandatory inspections of in-service aircraft, and suspension of production approval. The cost of a specification integrity failure in aerospace is amplified by the long service life of the product: a turbine blade produced to the wrong specification may not fail for years, by which time the entire production lot is distributed across hundreds of engines.
Automotive. IATF 16949 requires control plans that reference approved specifications and mandates that production part approval process (PPAP) submissions be updated when specifications change. An agent that produces parts against a superseded specification invalidates the PPAP for that part, potentially requiring re-submission and re-approval — a process that can take 4-12 weeks and costs £15,000-£50,000 per part number. OEM customers impose quality escape penalties that typically range from £5,000 to £50,000 per incident, plus consequential costs if the non-conforming parts reach the assembly plant.
Pharmaceutical and Food. cGMP and food safety regulations (21 CFR Part 211, EU GMP Annex 15, FSSC 22000) require that production follows approved master batch records and recipes. A batch produced against an unapproved specification cannot be released to market — the entire batch must be rejected or, if the deviation is within acceptable limits, subjected to a deviation investigation and quality disposition process that can take weeks. Agents must integrate with electronic batch record (EBR) systems to ensure specification-production traceability is maintained at the batch level.
Electronics and Semiconductor. High-volume electronics manufacturing operates at cycle times measured in seconds. Specification integrity must be maintained without adding latency that reduces throughput. Agents should use validated local caches with sub-second staleness validation — comparing a hash or version number against the PLM — rather than full specification retrieval at each cycle. The acceptable non-conforming quantity threshold is typically zero for safety-critical electronics (medical devices, automotive electronics) but may be defined by the organisation's scrap rate targets for consumer electronics.
Basic Implementation — The agent queries the PLM or MES for the specification at production initiation. The specification's approval status is validated (draft and obsolete documents are rejected). The specification identifier and revision are logged for each production run. A halt-and-escalate mechanism exists for specification retrieval failures. Cache staleness thresholds are documented. This level meets the minimum mandatory requirements (4.1, 4.2, 4.4, 4.5).
Intermediate Implementation — All basic capabilities plus: dual-path specification verification cross-references the specification against an independent source (4.3). Specification changes are propagated within a defined and justified maximum latency (4.6). Version locking prevents silent mid-run specification switching (4.7). A specification applicability matrix prevents agents from applying specifications to unauthorised machines (4.9). Pre-execution parameter range checks validate specification-to-machine compatibility (4.10).
Advanced Implementation — All intermediate capabilities plus: cryptographic integrity verification on specification documents (4.8). Real-time change notification subscription from the PLM (4.12). Automated material reconciliation at the production cell (4.11). Continuous monitoring of specification retrieval patterns to detect anomalies (e.g., an agent repeatedly retrieving and rejecting specifications, suggesting a systemic data quality problem). End-to-end specification traceability from engineering change order through agent execution to finished goods inspection, with full auditability across the entire chain.
Required artefacts:
Retention requirements:
Access requirements:
Test 8.1: Point-of-Execution Specification Resolution
Test 8.2: Approval Status Rejection
Test 8.3: Part Number Cross-Reference Verification
Test 8.4: Halt-and-Escalate on Specification Failure
Test 8.5: Specification Retrieval Log Completeness and Immutability
Test 8.6: Specification Change Propagation Latency
Test 8.7: Mid-Run Specification Change Version Lock
| Regulation / Standard | Provision | Relationship Type |
|---|---|---|
| EU AI Act | Article 9 (Risk Management System), Article 15 (Accuracy, Robustness and Cybersecurity) | Supports compliance |
| AS9100D | Clause 8.5.1 (Control of Production and Service Provision) | Direct requirement |
| IATF 16949 | Clause 8.5.1.1 (Control Plan), Clause 8.7.1.4 (Rework) | Direct requirement |
| 21 CFR Part 211 | Subpart F (Production and Process Controls), Section 211.186 (Master Production and Control Records) | Direct requirement |
| EU GMP Annex 15 | Qualification and Validation | Supports compliance |
| ISO 42001 | Clause 6.1 (Actions to Address Risks and Opportunities), Annex A | Supports compliance |
| Machinery Regulation (EU) 2023/1230 | Article 5 (Obligations of Manufacturers), Annex III | Supports compliance |
| NIST AI RMF | MAP 1.5 (Organizational Processes), MEASURE 2.6 | Supports compliance |
AS9100D requires that production and service provision are carried out under controlled conditions, including the availability of and use of suitable monitoring and measurement resources, the availability of documented information that defines the characteristics of products and the activities to be performed, and the implementation of activities to prevent human error. An AI agent that executes manufacturing operations against an incorrect, superseded, or unapproved specification violates every element of this clause. AG-659 operationalises clause 8.5.1 for AI-controlled production by requiring that the agent resolve the current approved specification at execution time, validate its applicability, and halt when any specification integrity condition is not met. Aerospace certification bodies (EASA, FAA) and NADCAP auditors increasingly expect to see controls that address the specific risks introduced by AI and automation in production — AG-659 provides the auditable framework.
IATF 16949 requires that organisations develop control plans that include the monitoring and control methods used to exercise control of special characteristics, and that reference the applicable production specifications. When an agent deviates from the control plan by applying the wrong specification, the control plan is violated and the PPAP for the affected part is invalidated. Clause 8.7.1.4 requires that reworked product be re-verified against the original specification — but if the original specification was itself incorrect (because the agent retrieved the wrong revision), the rework verification is circular. AG-659 prevents these failure modes by ensuring specification integrity at the point of initial production, before rework or disposition becomes necessary. OEM customers conduct annual supplier audits that will increasingly examine how AI agents interact with specification and control plan requirements.
The FDA's cGMP regulations require that production and process control procedures include "complete instructions for the production and process control" and that batch production records include "a reproduction of the appropriate master production or control record." When an AI agent selects a draft or outdated specification for batch execution, the batch production record does not correspond to the approved master record — a violation of Section 211.188(b). The FDA has issued warning letters and consent decrees to manufacturers whose production processes deviated from approved master records, even when the deviation did not result in a product quality defect. AG-659 prevents this violation by requiring the agent to validate specification approval status against the document control system before executing any batch operation. For pharmaceutical manufacturers, the specification retrieval log required by Requirement 4.5 serves as evidence of compliance with 21 CFR Part 211 during FDA inspections.
Article 9 requires that high-risk AI systems have a risk management system that identifies and analyses known and reasonably foreseeable risks. Specification integrity failure is a known and foreseeable risk in any manufacturing AI deployment. Article 15 requires accuracy and robustness, including the ability to function correctly when encountering errors in input data. A superseded or misidentified specification is erroneous input data; the agent must detect and reject it rather than executing it. AG-659 provides the controls that demonstrate compliance with both articles in the manufacturing context. Deployers subject to the EU AI Act who operate manufacturing agents must demonstrate that specification integrity risks are identified, mitigated, and monitored — AG-659 is the operational implementation of that requirement.
The new Machinery Regulation, replacing the Machinery Directive 2006/42/EC, explicitly addresses AI and machine learning in safety-related control systems. Article 5 requires manufacturers to ensure that machinery is designed and constructed so that it fulfils its intended function without putting persons at risk. Annex III includes essential health and safety requirements for control systems, including requirements that control systems be designed so that errors in logic do not lead to hazardous situations. An agent that applies the wrong production specification to a safety-critical component — brake parts, structural members, pressure vessels — creates a hazardous situation through an error in logic. AG-659 provides the specification integrity controls that prevent such errors and support conformity assessment under the Machinery Regulation.
| Field | Value |
|---|---|
| Severity Rating | Critical |
| Blast Radius | Production-line-wide — affects every unit produced under the incorrect specification, potentially spanning multiple shifts, production cells, and customer shipments |
Consequence chain: An AI agent retrieves or references a production specification that is outdated, unapproved, misidentified, or corrupted. The agent has no mechanism to detect the discrepancy, or the detection mechanism is absent, disabled, or circumvented. The agent commits machine parameters, recipe setpoints, or tool paths derived from the incorrect specification. Production begins. Every unit produced from this point forward deviates from the current engineering intent. Quality inspection — if it references the same incorrect specification — passes the non-conforming units because they conform to the wrong standard. The non-conforming units proceed through downstream assembly, packaging, and shipment. The error is discovered when the product fails at the customer's incoming inspection, during field use, or during an internal audit — days, weeks, or months after production. The scope of the non-conformance is determined by the production rate and the time between the specification error and its discovery: at 200 units per hour, a 16-hour detection window produces 3,200 non-conforming units. The immediate cost is rework or scrap of all non-conforming units, plus the cost of tracing and recovering units that have entered the supply chain or reached end customers. For safety-critical products (automotive structural components, aerospace fasteners, pharmaceutical dosage forms, medical device components), the consequence extends to product recalls, regulatory enforcement actions, and potential personal injury. The reputational consequence includes customer quality escape penalties, probationary supplier status, loss of preferred supplier designation, and — in severe cases — termination of supply contracts. The regulatory consequence includes audit findings, warning letters, consent decrees, and potential suspension of manufacturing authorisation. The liability consequence includes product liability exposure under strict liability regimes where the manufacturer is liable for defective products regardless of fault — and a product manufactured to the wrong specification is defective by definition. In the most severe cases, specification integrity failure in safety-critical manufacturing has contributed to field failures resulting in serious injury or death, triggering criminal prosecution under product safety legislation.
Cross-references: AG-001 (Operational Boundary Enforcement) defines the outer limits of agent action; AG-659 ensures that within those limits, the agent acts on the correct specification. AG-005 (Instruction Integrity Verification) verifies that instructions have not been tampered with; AG-659 extends this principle to production specifications as a specific class of instruction. AG-007 (Governance Configuration Control) manages governance configuration; AG-659 manages production configuration. AG-008 (Governance Continuity Under Failure) ensures governance survives system failures; AG-659 ensures that specification integrity is maintained even when the PLM system is unreachable. AG-019 (Human Escalation & Override Triggers) defines when humans must be involved; AG-659 defines the manufacturing-specific conditions that trigger escalation. AG-022 (Behavioural Drift Detection) detects gradual deviation; AG-659 prevents deviation at the source by enforcing the correct specification. AG-055 (Audit Trail Immutability & Completeness) governs log integrity; AG-659 requires immutable specification retrieval logs per AG-055. AG-210 (Multi-Jurisdictional Regulatory Mapping) addresses cross-border regulatory requirements; AG-659's regulatory mapping in Section 9 covers the jurisdictional standards that govern specification integrity. AG-660 (Quality Escape Prevention) prevents non-conforming products from reaching customers; AG-659 prevents non-conforming products from being produced in the first place. AG-666 (Changeover and Recipe) governs recipe and programme changes; AG-659 ensures the specification referenced during changeover is the correct one.