# AI in Quality Management: Agents for Complaints, Audits and Inspection

> AI agents help quality teams where the work is text-heavy and repetitive: complaints and 8D reports, FMEA upkeep, audit preparation, inspection planning and supplier evaluation. They draft, search and cross-check; qualified people still approve the result and remain accountable for it.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-in-quality-management  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-11-16

Quality teams in manufacturing spend much of their day reading and writing: customer complaints, 8D reports, FMEA rows, audit questions, work instructions, supplier responses. That is where AI agents can take on real work. This guide maps the main use cases, explains the building blocks and tackles the question that matters more in quality than anywhere else: who validates the tool, and who signs off on what it produces?

## What is an AI agent in quality management?

An AI agent in quality management is a language model that carries out defined steps rather than just chatting. It reads a complaint record, finds similar past cases, proposes a defect category and drafts a containment action. It works with your own systems and documents, and it hands over to a person at fixed points in the process.

Three things separate an agent from a general chat assistant:

- **Tools.** It can read specific systems – a quality management system (QMS) or CAQ tool, the ERP, document control – and, at most, create drafts. Access goes through approved interfaces, not open-ended permissions.
- **Process.** It follows a defined workflow, such as the disciplines of an 8D report, instead of answering freely.
- **Traceability.** Every suggestion links back to its sources: the earlier case, the control plan, the FMEA row. In quality this is not a nice extra; auditors will ask for it.

An agent is not a gauge, a measurement system or an automatic release step. Decisions about a part, a batch or a customer response stay with a qualified person.

## How does a quality agent work?

A quality agent has four building blocks: a language model, a knowledge base built from your controlled documents and case history, tools to reach your systems, and an approval step where a person reviews the output. Remove any one of them and you get either generic text or results nobody can stand behind.

**Language model.** It handles unstructured input – free-text complaints, inspection reports, supplier statements – and writes drafts. Whether it runs in the cloud or on your own hardware depends on data protection, confidentiality and customer contracts. Suppliers in particular often have contractual limits on where drawings and complaint data may be processed; our page on [sovereign AI](https://www.ruebstudio.de/en/sovereign-ai) covers the options.

**Knowledge base.** Controlled procedures, control plans, FMEAs, closed 8D cases and supplier ratings. Revision status is critical: the agent must only see current revisions, or it will quote an obsolete work instruction. The general approach is described on the page about [knowledge management agents](https://www.ruebstudio.de/en/ai-agents/knowledge-management).

**Tools.** Read access to the QMS, the ERP (batches, delivery notes, suppliers) and document control; write access limited to drafts. Every call is logged.

**Approval.** The agent produces drafts with references, a person reviews, edits and releases. Release is where accountability sits, and it is what you show an auditor. Designing approval levels is covered in [human-in-the-loop for AI agents](https://www.ruebstudio.de/en/guides/ai-agent-architecture/human-in-the-loop-ai-agents).

## Use cases: a map of quality work

Five areas lend themselves to agents today: complaints and 8D, FMEA, audits, inspection planning and supplier quality. They differ in how much free text they involve, how often the workflow repeats and how much harm a wrong suggestion can do. Those three factors tell you where to start.

| Area | What the agent does | Impact of an error | Good first project? |
|---|---|---|---|
| Complaints / 8D | Capture, classify, find similar cases, draft report | medium – the customer sees it | yes |
| FMEA | Propose failure modes from complaints, check consistency | medium – drives actions | after 8D |
| Audits | Collect evidence, prepare question sets | low to medium | yes |
| Inspection planning | Derive characteristics from drawings and specs | high – affects product release | later |
| Supplier quality | Review supplier responses, prepare ratings | medium | after 8D |

### Complaints and 8D reports

Complaints arrive by email, through customer portals or as PDF forms. The agent extracts part number, batch, defect description and quantity, assigns a defect category and pulls up closed cases with a similar pattern. For the 8D report it drafts the problem description and suggests containment actions that worked before. Root cause analysis – the core of any 8D report – remains a team job at the part and on the line; the agent offers hypotheses and data, not the cause. Building an 8D agent and capturing and classifying complaints each get their own article in this section.

### FMEA

FMEAs depend on field experience that usually sits in complaints and 8D reports and rarely makes it back. An agent can turn new failure modes from closed cases into proposed FMEA entries and check whether effects, actions and ratings are consistent. For automotive, AIAG and VDA released a joint FMEA handbook in 2019, which [according to NCMEP](https://ncmep.org/new-aiag-vda-fmea-handbook-released/) introduced a seven-step approach and Action Priority (AP) tables in place of the risk priority number. The agent has to know which method your company uses, or it will rate risks on the wrong scheme.

### Audits

Before any internal, customer or certification audit, people spend hours hunting for evidence: records, minutes, open actions from the last round. An agent can match a question set to the relevant documents and records and flag what is missing. For process audits in the German automotive supply chain there is, for example, the VDA 6.3 questionnaire; [VDA QMC lists](https://www.vda-qmc.de/publikationen/) its fourth revised edition as published in January 2023. The agent prepares; the auditor judges.

### Inspection planning

Here the agent derives inspection characteristics from drawings, specifications and customer requirements, then compares them with control plans for similar parts. This carries the most risk: a missed characteristic goes straight into product release. The agent should only highlight gaps and differences, never approve a plan.

### Supplier quality

Suppliers answer corrective action requests with their own 8D reports, of very uneven quality. An agent can check whether every discipline is covered, whether the stated root cause fits the defect and whether corrective actions are evidenced. The result is a draft follow-up question for the supplier and input for the supplier rating.

## Validation and accountability by regulatory setting

How rigorously you need to qualify a quality agent depends on your sector. Medical devices and pharmaceuticals have explicit expectations for software used in the quality system; in automotive and general manufacturing the duty follows indirectly from keeping processes under control. Everywhere, accountability stays with the company, not the tool.

The following is orientation, not legal advice. For your situation you need the actual texts and, where in doubt, specialist input.

**General manufacturing and automotive.** [According to the IATF](https://www.iatfglobaloversight.org/iatf-169492016/about/), IATF 16949:2016 replaced ISO/TS 16949:2009 and stays aligned with ISO 9001 through a liaison status. It contains no rules specific to AI agents. In practice, once an agent becomes part of a controlled process such as complaint handling, it belongs in the process description: purpose, limits, approval step and logging. Auditors ask about traceability, not technology. Supplier-specific topics such as PPAP and OEM requirements are covered in [AI agents for automotive suppliers](https://www.ruebstudio.de/en/guides/ai-by-industry/ai-agents-automotive-suppliers); this guide stays with the quality workflows themselves.

**Medical devices.** In the United States, the FDA's Quality Management System Regulation incorporates ISO 13485:2016 by reference and [took effect on 2 February 2026](https://www.fda.gov/medical-devices/postmarket-requirements-devices/quality-management-system-regulation-qmsr). For software used in production or the quality system, the FDA [announced its final guidance](https://public-inspection.federalregister.gov/2025-18468.pdf) "Computer Software Assurance for Production and Quality System Software" in September 2025. A quality agent at a device manufacturer falls into this scope, so its intended use and the assurance activities against that use must be documented. Wider sector topics, such as technical documentation, are covered in [AI agents for medical device makers](https://www.ruebstudio.de/en/guides/ai-by-industry/ai-agents-medical-device-manufacturers).

**Pharmaceuticals (GMP).** A new Annex 22 on artificial intelligence was drafted for the EU GMP guidelines. [PIC/S states](https://picscheme.org/en/news/131/joint-stakeholders-consultation-on-the-revision-of-chapter-4) that the public consultation ran from 7 July to 7 October 2025, and that the draft sets requirements for selecting, training and validating AI models, with continuous oversight including change control. A [Legal 500 summary](https://www.legal500.com/intelligence/european-union/technology/eu-consults-on-new-gmp-rules-for-ai-in-pharma-manufacturing) notes that the draft excludes generative AI and large language models from critical GMP functions while leaving room for non-critical uses with a human in the loop. Check whether, and in what form, the annex applies before you plan.

**EU AI Act.** Across all sectors, Article 4 of the AI Act in its [consolidated version](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4) requires providers and deployers to take measures supporting the AI literacy of their staff. For high-risk systems, [Article 14](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14) sets out effective human oversight and explicitly names automation bias – the tendency to over-rely on a system's output. Whether your use counts as high-risk needs a case-by-case check, but the principle applies to any quality agent. What the literacy duty means in practice is explained in [AI literacy under Article 4](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/ai-literacy-article-4-ai-act).

Outside the EU and US, the same pattern holds: sector rules on software in the quality system, plus general AI rules that differ by jurisdiction. Build your documentation so it can answer both.

## Prerequisites: data, systems, people

A quality agent needs three things: usable case history, controlled documents with a clear revision status, and a named person who owns the process. Without case history the agent cannot find similar cases; without document control it cites outdated material; without an owner it stays an experiment.

**Data.** Closed complaints and 8D reports, ideally with defect category, root cause and action. Free text is fine – language models cope with it – but part numbers and batch numbers must be clean, or the links break.

**Systems.** A QMS or CAQ tool, an ERP, or at least a structured file store the agent can read. Spreadsheets are no obstacle for a pilot, but they are a reason to tidy up first.

**People.** A quality engineer who owns the agent, sets release rules and reviews wrong suggestions. Where works councils or employee representatives exist, agree early on how far the agent makes individual work visible.

## Rolling it out in five steps

Start with a single, frequent workflow – usually complaint intake – measure suggestion quality against cases you have already closed, and only then widen the scope. That way you build the evidence you will later need for validation and audits as a by-product of the pilot.

1. **Choose a workflow.** Many cases, lots of free text, low harm from a wrong suggestion. Complaint intake and audit preparation fit; inspection planning does not.
2. **Define intended use.** What the agent does, what it does not do, where a person approves. This document underpins any later validation.
3. **Test on closed cases.** Run the agent on cases with known outcomes and analyse every deviation before a single live case goes through. Building such test sets is the subject of [evals for AI agents](https://www.ruebstudio.de/en/guides/ai-agent-architecture/ai-agent-evals).
4. **Run in parallel.** The agent works alongside the existing process. Only once its suggestions are reliable does it become part of the process.
5. **Monitor and control changes.** A new model, new documents or a changed prompt are process changes and go through change control.

For comparison, the page on [technical service agents](https://www.ruebstudio.de/en/ai-agents/technical-service) applies the same building blocks to customer requests and breakdowns. In quality the subject is deviations and evidence, and the bar for traceability is higher.

## Limits and risks

The biggest risks are root causes that sound convincing but are wrong, outdated documents in the knowledge base, and a gradual loss of scrutiny as people accept suggestions unchecked. All three can be reduced; none can be ruled out.

- **Plausible but false causes.** Language models write persuasively. A root cause that reads well but was never verified on the part is worse than none; this is a form of [AI hallucination](https://www.ruebstudio.de/en/glossary/ai-hallucination). Root causes stay with the team.
- **Stale documents.** If the agent reads an obsolete revision, every suggestion is wrong even when it looks correct. The knowledge base must follow document control.
- **Automation bias.** After a few weeks of good suggestions, attention drops. Periodic second-person reviews of released cases keep scrutiny up.
- **Confidentiality.** Complaint data often includes customer drawings and specifications. Where contracts restrict processing, only an on-premises setup will do.
- **When it does not fit.** With only a handful of cases a year, the effort for knowledge base and validation does not pay off. And an agent will not bring an uncontrolled process under control; it will only make it faster.

In our own operations, showing every suggestion with its sources and reviewing wrong suggestions in batches rather than one by one has worked well. Patterns then show whether the problem lies with the model, the data or the process description.

## Where should you start?

Start with an inventory: which quality tasks take the most time, how many cases arrive each month, where the data lives, and which regulatory setting applies. That almost always points to a first workflow that can be tested against closed cases within a few weeks.

Three questions help you choose:

- Are there at least a few dozen closed cases to measure the agent against?
- Is it clear who approves, and will that person challenge suggestions rather than wave them through?
- Which standards, customer contracts and data protection rules decide where the model may run?

For an overview of department-specific agents and the building blocks they share, see [AI agents for companies](https://www.ruebstudio.de/en/ai-agents). If you would rather work through workflows and data with your team first, the [workshop](https://www.ruebstudio.de/en/workshop) page describes a format for doing that.

## Frequently asked questions

### Should an AI agent send 8D reports to customers on its own?

It can, but it should not. An 8D report is a commitment your company makes to a customer, so root cause and release belong to a named quality engineer; the agent supplies the draft.

### Does a quality agent need to be validated?

In medical devices and pharmaceuticals there are explicit expectations for software used in the quality system. Elsewhere, a documented, risk-based check against the intended use is still worth doing, because auditors will ask how the tool is controlled.

### What data does a quality agent need?

Access to complaint or nonconformance records, the current procedures and control plans, and a history of closed cases. Without these, the agent produces generic text rather than case-specific suggestions.

### Can large language models be used in GMP manufacturing?

According to summaries of the draft EU GMP Annex 22, they are excluded from critical GMP uses but may be used in non-critical ones with a human in the loop. Check the current status of the annex before planning, and seek specialist advice if in doubt.

## Sources

- [IATF – About IATF 16949:2016](https://www.iatfglobaloversight.org/iatf-169492016/about/)
- [VDA QMC – Publications (incl. VDA 6.3 Process Audit)](https://www.vda-qmc.de/publikationen/)
- [NCMEP – New AIAG & VDA FMEA Handbook Released (2019)](https://ncmep.org/new-aiag-vda-fmea-handbook-released/)
- [FDA – Quality Management System Regulation (QMSR)](https://www.fda.gov/medical-devices/postmarket-requirements-devices/quality-management-system-regulation-qmsr)
- [Federal Register – Computer Software Assurance for Production and Quality System Software; Availability (2025)](https://public-inspection.federalregister.gov/2025-18468.pdf)
- [PIC/S – Joint stakeholders consultation on Chapter 4, Annex 11 and new Annex 22](https://picscheme.org/en/news/131/joint-stakeholders-consultation-on-the-revision-of-chapter-4)
- [Legal 500 – EU consults on new GMP rules for AI in pharma manufacturing](https://www.legal500.com/intelligence/european-union/technology/eu-consults-on-new-gmp-rules-for-ai-in-pharma-manufacturing)
- [EU AI Act Service Desk – Article 4 AI literacy](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-4)
- [EU AI Act Service Desk – Article 14 Human oversight](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14)
