AI Agents in Quality Management - Guest Article on QZ-Online

AI Agents in Quality Management: How Semantic Matching Reduces Processing Effort by 70%

Together with the prestigious trade journal QZ, we discussed the topic of "AI Agents in Quality Management". In this article, we demonstrate why the primary bottleneck in product compliance requests is not a lack of data, but inconsistent formats—and how AI agents resolve this challenge.

The Challenge: Identical Questions, Hundreds of Formats

Quality managers are well acquainted with this scenario: a customer submits a 150-row spreadsheet requesting compliance data for REACH, RoHS, or industry-specific standards. The next customer asks the exact same questions, but in an entirely different format. What follows is manual copy-and-paste work: questionnaire by questionnaire, column by column.

One customer requests "SVHC > 0.1% w/w", the next asks for "Candidate List status", and a third specifies "Art. 33 REACH compliant". The same enquiry regarding RoHS compliance may arrive informally via email, as a sub-question in a PDF form, or within a comprehensive Excel questionnaire. The administrative burden does not lie in providing the answer itself, but in manually populating the templates.

The Solution: Semantic Matching with AI Agents

Large language models facilitate a novel approach: they interpret the layout and format of questionnaires, identifying queries both contextually and structurally. The AI analyses column headers, response fields, and cross-sheet dependencies. Simultaneously, it comprehends the semantic intent behind differing formulations of the same query.

This enables the automated transfer of knowledge between internal corporate data and customer enquiries, as well as across disparate customer requests.

Case Study: 70% Reduction in Effort within Intralogistics

A multinational intralogistics provider routinely receives extensive customer questionnaires concerning supplier declarations, material disclosures, and product compliance. Whilst the requisite data was largely available, previous processing workflows required several hours per request.

Following the deployment of the AI system, incoming requests are ingested, their semantic meaning is interpreted, and the centralised knowledge base is queried for pertinent information. Based on this, the system generates a draft response for review and sign-off by the specialist department. The result: internal coordination effort was reduced by approximately 70%.

Preventing Hallucinations: Security by Design

In a compliance environment, generating plausible yet inaccurate statements carries severe risks. To mitigate this, the systems exclusively access internally verified and approved documentation. General background data from the language model's pre-training is not incorporated into the responses.

Furthermore, the platform employs an LLM-as-a-judge architecture: a secondary language model rigorously audits the generated responses against the underlying source documentation. In the absence of an adequate data foundation, no response is generated; instead, the item is explicitly flagged as pending review.

The Role of the Human in the Loop

Even if only 70% of the information is structured, automating its transfer yields substantial time savings. The remaining 30% is systematically flagged for manual evaluation—focusing human expertise precisely where it is essential.

Key operational takeaways:

• The achievable degree of automation is directly dependent on the quality of the baseline data

• Previously completed questionnaires constitute the most valuable data source

• Team adoption increases when every generated response transparently references its source

• Domain experts retain full responsibility for final sign-off

• With each manual correction, the knowledge base expands and accuracy improves continuously

The full article including the video interview is available on QZ-Online: AI Agents in Quality Management