Quality Management in Transition - Not error-free. But trustworthy

Interview by Simon Tischer (QZ) of Dr Ulrich Harmes-Liedtke originally published in QZ on 2 July 2026.

In this column, QM (Quality Management) practitioners set out their perspective on the changes underway in quality management. This time, Dr Ulrich Harmes-Liedtke, a freelance consultant for international cooperation and quality infrastructure, shares his ideas on how to deal with errors. He is a founding partner of the consultancy Mesopartner and the creator of the Global Quality Infrastructure Index (GQII).

Is the classic quality goal of “zero-defect tolerance” still appropriate in the age of AI?

No, and this has been true for longer than artificial intelligence (AI) has been around. In quality management (QM), there are two fundamentally different ways of dealing with errors.

The first treats an error as a deviation from the target that has to be eliminated. This makes sense wherever conformity, safety, and consumer protection count.

The second uses errors as a signal, as information about a system that is not yet fully understood. Errors of this kind are indispensable whenever something new is being tried out. In the software industry, this is known as “banana ripeness”. Products are shipped “green” and ripen at the customer’s premises. This is not a quality promise, but as long as the consequences of errors are limited and communicated transparently, it is a legitimate learning path.

Generative AI systems make this distinction impossible to ignore. They are stochastic, meaning they operate onprobabilities rather than certainties. Anyone who demands zero defects from them is placing the wrong requirement on the wrong technology. QM has to learn to choose between the two logics of error depending on the situation. That is not a weakness but a sign of maturity.

You call for errors not to be demonised but for their probability to be quantified. What does this handling of imprecision look like in practice?

The decisive question is not whether the result is error-free. We should be asking which errors occur, how often they occur and what that means.

Accredited testing and calibration laboratories have already grasped this. They distinguish very precisely between different kinds of error. An error is the deviation of the result from the “true value”. In metrology, the “true value” is regarded as a theoretical concept, since in practice it can never be fully known. Measurement uncertainty quantifies theuncertainty arising from all influencing factors, including environmental conditions, equipment limitations, and procedural variability. Both aspects feed into the risk assessment under DIN EN ISO/IEC 17025, which prescribes how identified deviations are to be handled. It is precisely this way of thinking that now has to enter our handling of AI systems as well.

Here is an example. Inspection technicians have to check thousands of fire extinguishers, machines, and switch cabinets for correct safety markings such as inspection stickers, hazardous-substance notices and warning signs in accordance with DIN EN ISO 7010:2020. AI-supported image recognition, trained on large datasets of correctly and incorrectly marked objects, takes over the visual inspection. If the system achieves, say, a 94 per cent hit rate, we should ask ourselves which marking types it fails on. Is an overlooked non-conformity or a false rejection the greater harm? Only this differentiation makes the figure usable. So it is less about making “no error” and much more about “which error occurs, where and with what consequence”.

In which areas may we allow ourselves “good enough”, and where does absolute freedom from error remain a non-negotiable obligation?

The answer depends on two questions. What is the consequence of an error? And how well do we understand the system?

In stable areas such as standardised testing procedures or safety-critical manufacturing processes, zero defects remain the goal. In dynamic, interconnected systems, by contrast, it is more important to detect errors early and to learn from them, as with the networking of automotive electronics.

A concrete example. The market surveillance authorities of the German federal states today face millions of product listings on e-commerce platforms. The planned AI Market Surveillance and Innovation Promotion Act (KI-MIG) provides for AI systems to automatically scan seller platforms for counterfeit CE markings or misleading claims. Such systems cannot capture every case, but they fundamentally change the ratio of effort to effect once their error structure is known and can be factored in.

“Good enough” is not an excuse but a well-considered decision. Absolute freedom from error does not exist. Even aviation works with failure probabilities per flight hour, and pharmacy with limit values and side-effect profiles. In safety-critical areas, the aim is not freedom from error but the highest reliability: residual risks that are demonstrably under control and reduced to a socially accepted minimum. This does not contradict the basic thesis that errors must be viewed in a differentiated way; it is its logical continuation.

Note the distinction between error and failure here. An error is a deviation from the true or intended value. A failure is the point at which a system no longer performs as required. Good quality management does not chase the impossible elimination of every error; it keeps failures under control and their residual risk acceptably low.

Does quality management not lose authority when it openly admits uncertainties to the customer instead of guaranteeing absolute freedom from error?

The opposite is true. QM loses credibility when it makes promises it cannot keep. There is nothing trustworthy about promising a hundred per cent freedom from error that cannot exist.

Testing and calibration laboratories have been showing for years how to do it better. They state the defined measurement uncertainties and clearly describe the limits of their statements. This is regarded as proof of competence, not as a weakness. Anyone who introduces an AI system and can explain where it is reliable and where it is not strengthens trust. Anyone who promises blanket freedom from error but cannot deliver it loses that trust for good.

Quality culture and error culture are not opposites but two sides of the same organisational maturity. Anyone who only avoids errors is optimising for a world of yesterday. The question is not about achieving zero defects. The question is to identify which errors occur, their context, frequency, consequences, and who has the competence to decide on them.

Questions posed by the QZ editorial team

AUTHOR

Dr Ulrich Harmes-Liedtke

works as a freelance consultant in international development cooperation. He focuses, above all, on promoting quality infrastructure. Dr Harmes-Liedtke is a co-founder and partner of the consultancy Mesopartner and is responsible for the Global Quality Infrastructure Index, available at www.gqii.org. His clients include the Physikalisch-Technische Bundesanstalt, the Gesellschaft für Internationale Zusammenarbeit, the World Bank and various United Nations organisations.

CONTACT

www.mesopartner.com | www.gqii.org

This is a translation of the original article written in German, published by the leading German QM journal QZ in its July 2026 edition. The original is available at https://www.qz-online.de/a/interview/nicht-fehlerfrei-aber-verlaesslich-7842062