Back to Blog
Data & Analytics16 min read

Data Quality Is the Foundation of AI Quality: Lessons from Industrial Projects

Why 80% of AI project time goes to data preparation, typical data problems in manufacturing, and a practical checklist for assessing data readiness for AI.

S

Serhiy Andrejev

Lead Machine Learning Developer

"Garbage in — garbage out" is perhaps the most frequently cited and most frequently ignored principle in artificial intelligence. This article examines why data quality remains the decisive factor in AI project success, what specific data problems plague manufacturing environments, and how organizations can systematically assess their data readiness before committing to AI initiatives.

Introduction: The Truism Everyone Knows and Everyone Ignores

Every conversation about artificial intelligence in manufacturing eventually arrives at the same starting point: the model is only as good as the data it learns from. This observation is so widely accepted that it has become a truism — a statement so obviously true that it no longer commands attention. And that is precisely the problem.

In our practice, we have observed a persistent and troubling pattern. Organizations invest significant resources in selecting AI frameworks, evaluating model architectures, and benchmarking algorithm performance — activities that collectively account for roughly 10–20% of an AI project's total effort. Meanwhile, the data that will ultimately determine whether the model succeeds or fails receives comparatively little strategic attention until problems emerge during training or deployment.

Google Research documented this phenomenon in a landmark 2021 study on "data cascades" — compounding negative effects that originate from data quality issues and propagate downstream through the entire AI lifecycle. The study found that 92% of AI practitioners had experienced data cascades in their work, yet the majority of these cascades were preventable with earlier, more systematic attention to data quality [1]. The researchers concluded that the AI community systematically undervalues data work relative to model work, a finding that aligns closely with what we observe across industrial AI implementations.

The financial consequences are substantial. According to IBM, over one-quarter of organizations estimate they lose more than USD 5 million annually due to poor data quality, with some reporting losses exceeding USD 25 million [3]. Gartner predicts that through 2026, 60% of AI projects will be abandoned because they lack AI-ready data [5]. These are not theoretical risks. They represent the operational reality for organizations that treat data quality as a secondary concern.

This article draws on our experience implementing AI solutions in manufacturing environments across Ukraine and the European Union — contexts where sensor data is noisy, legacy systems are heterogeneous, and the gap between laboratory-quality datasets and production-floor reality is measured not in minor discrepancies but in fundamental incompatibilities.

Typical Data Problems in Manufacturing

Manufacturing data exists in a state of perpetual imperfection. Unlike curated research datasets or clean enterprise data warehouses, production data reflects the physical reality of industrial operations — machinery that degrades, sensors that drift, operators who make errors, and systems that were never designed to communicate with one another. Based on our observations across dozens of industrial implementations, five categories of data problems recur with remarkable consistency.

Gaps and missing values represent the most common data quality issue. Sensors go offline, communication links between PLCs and data historians drop, power interruptions reset logging systems. In a typical SCADA environment, we routinely observe data completeness rates between 85% and 95% — a range that sounds acceptable until one considers that the missing 5–15% often corresponds to precisely the operating conditions (equipment starts, stops, anomalies) that carry the greatest analytical value. A predictive maintenance model trained on data with systematic gaps during fault conditions will, by definition, fail to recognize the very patterns it was designed to detect.

Duplicates and redundant entries arise from multiple sources. Manual data entry into ERP and MES systems is inherently error-prone — double entries, typographical variations, inconsistent unit conversions. Automated systems contribute their own duplications when communication retries create multiple records for single events. In one food processing facility, we discovered that approximately 8% of quality inspection records were effective duplicates — identical measurements recorded under slightly different timestamps due to clock synchronization issues between the laboratory information management system and the production MES.

Sensor drift and parameter migration occur gradually and often go undetected for extended periods. Temperature sensors lose calibration. Flow meters develop deposits that alter readings. Vibration sensors shift their baseline as mounting hardware loosens. The insidious nature of drift is that each individual measurement appears plausible — the deviation from true values is small enough to pass basic validation checks. Over weeks or months, however, accumulated drift corrupts the statistical distributions that machine learning models rely upon, creating a form of data poisoning that is difficult to diagnose after the fact.

Desynchronization across systems is endemic in manufacturing environments built incrementally over decades. The ERP system operates on business time (shift boundaries, order completion timestamps). The SCADA system records in milliseconds. The quality management system logs batch-level results with varying latency. When these data streams must be merged for AI training, temporal misalignment introduces noise that degrades model performance. We have encountered situations where a two-minute clock drift between systems made a quality prediction model effectively useless, because the production parameters being correlated with quality results belonged to the wrong batch.

Document and metadata obsolescence is often overlooked but equally damaging. Equipment specifications reference superseded sensor models. Process recipes contain parameters for raw materials whose composition has changed. Maintenance logs describe procedures for hardware that has been replaced. When AI systems consume this metadata — as they increasingly do in knowledge-augmented and RAG-based architectures — obsolete documentation generates confidently wrong recommendations.

The following table summarizes these data problems, their manifestations, and their consequences for AI model quality:

Data ProblemDescriptionConsequence for AI ModelDetection Method
Gaps and missing valuesSensor outages, communication drops, logging interruptionsBiased training data; model blind spots during critical operating conditionsCompleteness audits; time-series continuity analysis
Duplicates and redundant entriesDouble manual entries, retry-generated records, overlapping collection systemsOverrepresentation of certain conditions; distorted statistical distributionsDeduplication algorithms; timestamp-based similarity matching
Sensor driftGradual calibration loss in temperature, flow, vibration, and pressure sensorsSystematic bias in feature values; degraded prediction accuracy over timeStatistical process control; periodic calibration cross-checks
DesynchronizationClock drift between ERP, MES, SCADA, and laboratory systemsIncorrect feature-target correlations; causal relationships lost in temporal noiseCross-system timestamp auditing; event correlation analysis
Document obsolescenceOutdated equipment specs, superseded recipes, stale maintenance proceduresIncorrect contextual information; misleading recommendations in knowledge-augmented systemsVersion tracking; periodic metadata validation reviews

Table 1. Common data quality problems in manufacturing environments and their impact on AI model performance. Source: author's observations from industrial AI implementations.

Why Production Data Differs from "Clean" Datasets

A fundamental gap exists between the datasets on which most AI practitioners are trained and the data they encounter in production environments. University curricula, online courses, and platforms like Kaggle provide datasets that are, by design, tractable — cleaned, labeled, and formatted to facilitate algorithmic comparison. Production data has none of these conveniences.

Andrew Ng, in articulating the case for data-centric AI, has emphasized that the traditional model-centric paradigm — holding data fixed while iterating on algorithms — inverts the priority that matters most in practice [4]. In manufacturing environments, the algorithm is rarely the bottleneck. The data is. And the characteristics that make production data difficult are not random noise but structured, systematic distortions that resist simple cleaning procedures.

Environmental noise in production data is qualitatively different from the Gaussian noise assumed by most statistical models. Electromagnetic interference near heavy machinery creates correlated artifacts in sensor readings. Vibrations from adjacent equipment contaminate acceleration measurements. Temperature gradients across a production hall mean that ambient sensors at different locations measure genuinely different conditions — not the uniform environment that many process models assume.

Anomalies in production data are not outliers to be removed — they are often the signal to be found. In a standard data science workflow, extreme values are flagged and frequently discarded. In predictive maintenance, those extreme values may be the earliest indicators of incipient failure. The distinction between a genuine anomaly worth investigating and a measurement artifact worth discarding requires domain expertise that no automated cleaning pipeline can fully replicate.

The human factor introduces variability that is difficult to model and impossible to eliminate. Operators adjust setpoints based on experience and intuition. Shift changes create discontinuities in operational patterns. Manual quality inspections carry inherent subjectivity — the same defect may be classified differently by different inspectors. In our experience, inter-operator variability accounts for 10–15% of total variance in datasets from manual or semi-automated processes.

McKinsey's research on data quality roadblocks in manufacturing confirms these observations at scale, identifying missing data points, broken or miscalibrated sensors, incomplete data mappings, and incompatible systems — compounded by legacy architectures and weak governance practices accumulated over decades of incremental digitalization [6].

The practical implication is straightforward but consequential: organizations that expect to apply AI to production data using the same methods and assumptions that work on curated datasets will encounter systematic underperformance that no amount of model tuning can resolve.

What "80% of Time on Data Preparation" Actually Means

The widely cited statistic that data scientists spend 80% of their time on data preparation has become a professional lament. Surveys report various figures — CrowdFlower found 60% on cleaning alone, reaching 80% when collection and labeling were included; Anaconda's 2020 survey reported approximately 45% on preparation tasks [2]. The precise percentage matters less than the underlying reality: data preparation dominates AI project timelines, and in industrial settings, the proportion is often even higher.

Based on our experience, the "80% data preparation" figure decomposes into five distinct stages, each with its own challenges, resource requirements, and failure modes.

Figure 1. Time distribution in a typical industrial AI project. Data preparation activities (inventory, cleaning, normalization, validation, enrichment) collectively consume approximately 80% of project effort. Sources: CrowdFlower survey, Anaconda State of Data Science 2020 [2], author's observations.

Data inventory and discovery (approximately 15% of total effort) involves identifying what data exists, where it resides, in what format, and at what quality level. In manufacturing environments with multiple production lines and heterogeneous data historians, this stage alone can take weeks. We consider it important to note that organizations frequently overestimate their knowledge of their own data landscape — the actual content, completeness, and reliability of data streams are often poorly understood until a systematic inventory is conducted.

Data cleaning and correction (approximately 25%) addresses the problems outlined in the previous section: filling or flagging gaps, removing duplicates, identifying and correcting drift-affected measurements, resolving synchronization discrepancies. This is the stage most people envision when they hear "data preparation," but it represents only about one-third of the total preparation effort.

Normalization and transformation (approximately 20%) converts heterogeneous data into formats suitable for AI model consumption. This includes unit standardization (Celsius to Kelvin, bar to Pascal), temporal alignment (resampling data from different systems to a common time base), feature engineering (deriving calculated variables from raw measurements), and encoding categorical variables from ERP and MES systems.

Validation and testing (approximately 10%) verifies that cleaned and transformed data accurately represents the physical processes it describes. Statistical tests confirm that distributions are plausible. Domain experts review edge cases and boundary conditions. Cross-validation against known process outcomes ensures that the data transformation pipeline has not introduced artifacts.

Data enrichment and augmentation (approximately 10%) supplements raw process data with contextual information: equipment maintenance histories, raw material certificates, environmental conditions, production schedules. In industrial AI, contextual enrichment is frequently what separates a model that achieves acceptable accuracy in the laboratory from one that performs reliably in production.

The remaining 20% of project effort divides roughly equally between model development (selection, training, tuning) and deployment, testing, and monitoring. This allocation surprises organizations accustomed to thinking of AI projects as primarily algorithmic endeavors. But it is consistent across virtually every industrial AI implementation we have participated in, and it aligns with the broader research consensus [1, 2, 4].

How to Assess Data Readiness for AI: A Practitioner's Checklist

Before committing resources to an AI initiative, organizations should conduct a systematic assessment of their data readiness. Based on our experience across manufacturing and process industries, we have developed an evaluation framework that addresses seven dimensions of data readiness. Each dimension should be assessed on a three-point scale: adequate (ready for AI), conditional (requires remediation before AI work begins), or insufficient (fundamental gaps that must be resolved first).

Completeness. What percentage of expected data points are actually present across the target time period? For most industrial AI applications, completeness below 90% is a significant risk factor, and below 80% typically renders the dataset unsuitable for reliable model training without substantial imputation or supplementation. The assessment should examine not only overall completeness but the distribution of gaps — random missing values are far less damaging than systematic gaps that correlate with specific operating conditions.

Consistency. Do the same measurements, recorded by different systems or at different points in the process, agree within acceptable tolerances? Consistency checks should compare overlapping data sources (redundant sensors, parallel recording systems) and verify that derived values are mathematically consistent with their source measurements. In our practice, consistency failures often reveal integration problems that would otherwise remain hidden until model training produces unexplainable results.

Timeliness. Is data available with sufficient temporal resolution and acceptable latency for the intended AI application? A predictive maintenance model that requires vibration data at 1 kHz resolution cannot function on 1 Hz aggregated averages. A real-time quality prediction system that receives laboratory results with a four-hour delay cannot provide the rapid feedback loop it was designed to deliver. Timeliness requirements vary significantly by use case and must be defined explicitly before data readiness can be assessed.

Accuracy. How closely do recorded values correspond to actual physical conditions? This dimension requires calibration verification, comparison against reference standards, and domain expert review. Accuracy assessment is particularly important for sensor data that has been in continuous service — calibration drift may have introduced systematic biases that are invisible to automated validation but consequential for model performance.

Representativeness. Does the available data cover the full range of operating conditions, product variants, seasonal patterns, and anomalous events that the AI model will encounter in production? A dataset spanning only normal operating conditions will produce a model that performs well under normal conditions but fails precisely when the organization needs it most — during process deviations, equipment degradation, or unusual production requirements.

Traceability. Can each data point be traced back to its source, with a clear audit trail documenting any transformations, aggregations, or corrections applied? In regulated industries — food processing, pharmaceuticals, automotive — traceability is not merely a data quality consideration but a compliance requirement. HACCP and ISO standards mandate documented data provenance, and AI systems built on untraceable data create regulatory liability regardless of their predictive accuracy.

Governance. Are there established procedures for data ownership, access control, quality monitoring, and lifecycle management? Gartner's research confirms that 63% of organizations either lack or are unsure whether they have appropriate data management practices for AI [5]. Without governance structures, data quality improvements achieved during an AI project will degrade over time as operational practices revert to pre-project norms.

Figure 2. Typical data readiness assessment results across manufacturing organizations. Governance (37%) and representativeness (35%) are the weakest dimensions, consistent with Gartner's finding that 63% of organizations lack appropriate data management practices for AI [5]. Completeness scores highest but still leaves 40% of organizations below the adequate threshold. Data based on author's observations across industrial AI implementations.

We recommend that this assessment be completed before any AI vendor selection or model development begins. In our experience, approximately 40% of organizations that approach us for AI implementation discover through this assessment that their immediate priority should be data infrastructure improvement rather than AI model development — a finding that ultimately saves months of effort and significant resources that would otherwise be spent building models on inadequate data.

Author's Assessment and Conclusions

The relationship between data quality and AI quality is not merely correlational — it is causal, direct, and non-negotiable. No algorithmic breakthrough and no increase in computational resources can compensate for fundamentally flawed data. This remains the most important and most underappreciated truth in applied artificial intelligence.

Several conclusions emerge from our experience implementing AI solutions in manufacturing environments.

First, data quality investment must precede AI investment. Organizations that attempt to address data quality issues concurrently with model development invariably discover that the two activities compete for the same resources and attention, with data quality consistently losing the competition. The most successful AI implementations we have observed are those where data quality was treated as a prerequisite phase with its own budget, timeline, and success criteria — not as a subordinate activity within an AI project plan.

Second, the 80% data preparation figure is not a problem to be solved — it is a reality to be managed. Efforts to reduce data preparation time through automation and tooling are valuable, but they will never eliminate the fundamental need for careful, domain-informed data curation. Organizations that budget for 80% data preparation and 20% model development will consistently outperform those that allocate resources in the inverse ratio, regardless of the sophistication of their modeling approach.

Third, data readiness assessment should become a standard practice — as routine as financial due diligence before an acquisition or structural engineering assessment before construction. The seven-dimension framework presented in this article provides a starting point, but each organization should adapt and extend it based on its specific industry, regulatory context, and operational characteristics.

Fourth, the data-centric AI paradigm articulated by Andrew Ng and validated by extensive research represents a genuine shift in how AI projects should be conceived and managed [4]. Rather than treating data as a fixed input and the model as the variable to be optimized, successful industrial AI projects treat the model as relatively fixed (proven architectures for the use case) and invest the majority of effort in systematically improving data quality, completeness, and relevance.

Finally, we consider it important to note that data quality is not a one-time achievement but a continuous discipline. Sensors drift, systems are upgraded, processes change, personnel rotate. An AI model that performs well on high-quality historical data will degrade if the ongoing data pipeline does not maintain equivalent quality standards. Data quality monitoring must be built into operational procedures with the same rigor as equipment maintenance — because in an AI-enabled manufacturing environment, it is equipment maintenance.

The organizations that will extract the most value from artificial intelligence in manufacturing are not those with the most sophisticated algorithms or the largest compute budgets. They are the organizations that recognize data quality as a strategic asset and invest in it accordingly — systematically, continuously, and with the same executive attention devoted to any other critical business function.

References

  1. Sambasivan, N., Kapania, S., Highfill, H., Akrong, D., Paritosh, P., Aroyo, L. "Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI." CHI 2021. https://research.google/pubs/everyone-wants-to-do-the-model-work-not-the-data-work-data-cascades-in-high-stakes-ai/

  2. Anaconda. "2020 State of Data Science." 2020. https://www.anaconda.com/resources/whitepapers/state-of-data-science-2020

  3. IBM. "The True Cost of Poor Data Quality." 2025. https://www.ibm.com/think/insights/cost-of-poor-data-quality

  4. Ng, A. "Why it's time for data-centric artificial intelligence." MIT Sloan Management Review, 2022. https://mitsloan.mit.edu/ideas-made-to-matter/why-its-time-data-centric-artificial-intelligence

  5. Gartner. "Lack of AI-Ready Data Puts AI Projects at Risk." 2025. https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk

  6. McKinsey & Company. "Clearing data-quality roadblocks: Unlocking AI in manufacturing." 2024. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/clearing-data-quality-roadblocks-unlocking-ai-in-manufacturing

  7. MIT CSAIL. "Introduction to Data-Centric AI." https://dcai.csail.mit.edu/

data qualityai data preparationmanufacturing datadata-centric aidata readinessindustrial aidata cleaningsensor datadata governance

AQTeck solutions on this topic

We design and implement these systems at industrial enterprises in Ukraine and the EU.

Related Articles

Need a consultation?

Contact us for a free consultation.