Raffaele De Piano
Principal Data Architect at Zühlke
AI proofs of concept often look convincing until they meet live data. Then come the inconsistencies, permission gaps, silent errors and loss of confidence that stall deployment.
If your AI initiatives are slowing down as they move towards production, the first explanation is often the wrong one. People tend to blame the model, the vendor, or the lack of internal adoption; but in reality, one of the most common causes is that the data foundation is not strong enough for real-world use.
That is why data quality has become essential to prevent AI failure and scale safely.
When data quality is weak, the effects show up quickly: incorrect outputs, biased decisions, security exposure, compliance surprises, and a loss of trust that can stall an otherwise promising AI initiative.
This article focuses on that question.
This article is part of our 'Value from AI now' series on the three key challenges organisations face when scaling AI initiatives. Explore the full framework here.
How data quality impacts your AI initiatives
Data quality refers to the accuracy, completeness, consistency, reliability, and relevance of your data and it has a direct bearing on AI decision-making and outputs.
Low-quality data compromises decision making, which can impact everything from strategic decision-making to customer interactions. It’s also one of the main causes of GenAI project failure.
The classic adage ‘garbage in, garbage out’ rings very true when it comes to building AI systems. Even when businesses can deliver early wins from the likes of GenAI RAG applications, they soon realise that data access and quality are critical for turning functional prototypes into scalable, product-grade solutions.
When data doesn’t have high quality, it’s easy for an organisation to start seeing some familiar risks:
- confidently wrong outputs based on incomplete or outdated information
- biased outcomes caused by non-representative data
- operational disruption when automated decisions rely on inconsistent sources
- regulatory or compliance exposure when permitted use, provenance, or accountability are unclear
What 'good' AI data quality actually means
If you train a model on poor or irrelevant data, the results will likely be disappointing. But labelling data as either 'good' or 'bad' in this way might be too simplistic. To truly understand data quality, more nuance is required.
It comes down to the specific problem you’re trying to solve. A data source might be an excellent fit for one AI model and worthless for another. That’s why choosing the best data for each task is so important.
To mitigate risks and truly become an AI-empowered business, companies must focus on collecting data that’s not only accurate and timely but also directly relevant to the challenge being addressed – whether you’re an insurance firm looking to streamline underwriting and claims processing, or a retailer looking to enhance real-time strategic decision making.
In a nutshell then, ‘good’ AI data quality is data that’s:
- relevant to the use case
- accurate and reliable
- complete and consistent enough for the level of risk involved
- timely enough to support real operations
- secure and accessible under the right controls
- compliant with the organisation’s obligations
- understandable across teams
AI tools can learn to identify data anomalies and predict potential future quality issues, minimising errors before they escalate.
A framework to ensure data quality for AI initiatives
We’re witnessing an increased focus on data quality across our clients and the many industries we work with. More organisations are recognising that data quality is critical for the success of AI initiatives.
To make data truly AI-ready, organisations should approach it methodically, focusing on alignment, validation, and governance.
1. Align: make sure the data fits the use case
Before a system goes anywhere near production, leaders should be clear about what data it depends on and why.
That means asking:
- Which sources is the system actually using?
- Are we allowed to use this data for this purpose and in this jurisdiction?
- Are the definitions and semantics consistent across teams?
- Is the data representative of what the system will face in real operation?
- Does the data support the actual business objective, or just the demo?
This is where many projects already carry hidden risk. A model may appear effective in a narrow environment, but once the real data landscape shows up — fragmented systems, conflicting definitions, unclear permissions — the reliability of the whole initiative starts to weaken.
Alignment is about stopping that earlier.
2. Validate: test reliability before the user does
That means more than a one-off quality check before launch. It means understanding what “good enough” looks like for the workflow in question, then checking consistently for missing fields, outdated values, duplication, anomalies and drift.
Validation should answer questions such as:
- Is the data consistent across systems?
- Are key fields complete enough for the level of risk involved?
- Are there anomalies that suggest quality is degrading?
- Do we have clear thresholds for when the data is not good enough to rely on?
- Will somebody act when those thresholds are breached?
Automation can help a great deal here. It can surface anomalies earlier, detect patterns humans might miss, and make monitoring more efficient. But automation does not remove the need for judgement. Validation only works when the organisation has agreed what it is validating against, and who is accountable for the response.
3. Govern: keep the data trustworthy over time
Good data does not stay good by accident.
Sources change. Definitions shift. Access rights evolve. New use cases introduce new pressures. Teams come and go. Models are updated. Without governance, the data layer quietly degrades until performance drops, confidence disappears, or a visible incident forces the issue.
That is why governance matters.
In practice, it should cover:
- clear ownership and stewardship
- traceability of where data came from and how it changed
- version control over important data assets
- compliance with internal policy and external regulation
- review processes when usage, structure or business meaning changes
This is also where the link to AI trust becomes obvious. No governance framework can compensate for data nobody can explain. If leaders want AI that is safe, auditable and ready to scale, the data underneath it must be fit, permitted and traceable.
The questions boards should ask before approving scale
Before approving or expanding an AI initiative, senior leaders should be able to get clear answers to five questions.
- Do we know exactly which data sources this system depends on?
- Are we allowed to use this data for this purpose, in this market, and under these conditions?
- How representative is the data of the people, scenarios and edge cases the system will face in production?
- Who owns ongoing data quality, monitoring and change?
- How quickly can we trace a problem back to a source, transformation, permission rule or change event?
These questions help ensure that AI initiatives begin with a defensible and sustainable data foundation.
Lay the data foundations for reliable, production-ready AI
High-quality data is the foundation of successful AI initiatives foundation of successful AI initiatives, ensuring that prototypes develop evolve into scalable, reliable, and impactful solutions.
However, many businesses rush this step, either by underestimating its importance or only recognising its value when problems arise.
To truly unlock the potential of AI:
- Focus on quality from the start. Accurate, complete, and consistent data sets are essential.
- Invest in governance and culture. A robust framework ensures that quality becomes second nature across your organisation.
- Leverage the right tools and approaches. Automating processes with AI tools can enhance efficiency and reliability, but it needs to be paired with accountability.
If you want to understand how robust your organisation’s data foundation for AI really is, Zühlke can help assess your current readiness and identify where improvements in AI data governance, validation, and risk management will have the greatest impact.
