(eamesBot/Shutterstock)
Now that AI is a board-level topic, organizations are rushing to achieve successful outcomes, but enabling that success requires planning. According to Gartner, more than 60% of AI projects fail to deliver on business SLAs and are often abandoned because of poor data quality, weak governance, or lack of contextual relevance. While AI/ML models receive much of the attention, the truth is that they are only as good as the data that feeds them. If organizations can’t trust their data, they can’t trust their AI.
This is where data observability comes in. Moving beyond simple monitoring or data quality checks, data observability continuously assesses the health, trustworthiness, and representation of data throughout its lifecycle. It ensures that data pipelines produce outputs aligned with business expectations and are suitable for training and operating AI/ML models.
Yet, data observability has also been caught up in the hype. Gartner’s Hype Cycle for Data Management 2025 notes that while observability rose quickly, it’s now in the “Trough of Disillusionment” as organizations struggle to make it practical and valuable. The lesson: observability isn’t just a tool you buy; it’s a discipline and culture that must be embedded into data practices to go along with the tool.
If organizations want to get data observability right and position themselves for AI success, they need to apply the following five steps:
1: Treat Observability as Core to AI Readiness In the traditional sense, high-quality data means that anomalies are scrubbed away, which isn’t enough for today’s AI/ML models. For example, in analytics we might cleanse outliers to create neat reports for human consumption. But for training an AI/ML model, those anomalies, errors, and unexpected events are vital. They help algorithms recognize the full range of real-world patterns.
(innni/Shutterstock)
Data observability ensures data pipelines capture representative data, both the expected and the messy. By continuously measuring drift, outliers, and unexpected changes, observability creates the feedback loop that allows AI/ML models to learn responsibly. In short, observability is not an add-on; it is a foundational practice for AI-ready data.
2: Embed Observability into DataOps Practices – Data observability is most effective when paired with DataOps. Just as DevOps brought continuous testing and monitoring into software delivery, DataOps embeds testing, validation, and governance into the data pipeline itself.
Rather than relying on manual checks after the fact, observability should be continuous and automated. This turns observability from a reactive safety net into a proactive accelerator for trusted data delivery.
As a result, every new dataset or transformation can generate metadata about quality, lineage, and performance, while pipelines can include regression tests and alerting as standard practice. It also ensures that failures or anomalies can be detected and flagged before they reach business users or AI/ML models.
3: Automate Governance Enforcement – Often blamed for slowing things down when it comes to AI, governance is always a non-negotiable. Regulations, risk controls, and business SLAs all demand that data feeding AI/ML models be governed in context.
(Aree_S/Shutterstock
The key is automation. Rather than policies that sit in binders, observability enables policies as code. In this way, data contracts and schema checks that are embedded in pipelines can validate that inputs remain fit for purpose. Drift detection routines, too, can automatically flag when training data diverges from operational realities while governance rules, from PII handling to lineage, are continuously enforced, not applied retroactively.
Automated governance is critical, as it creates trust that data flowing into AI/ML models complies with the right standards without slowing innovation.
4: Enable Cross-Functional Teams – Observability isn’t just a technical concern for data engineers. Its true value comes when business, governance, and AI teams share the same view of data health. Organizations should adopt multidisciplinary groups that combine business domain experts with technical staff.
What Gartner refers to as Fusion, these teams ensure observability solutions don’t just report row counts or freshness, but connect to business value. It checks for things such as are customer records are complete. Are operational KPIs trustworthy? Are AI/ML models being trained on representative datasets?
Embedding observability across roles creates shared accountability and accelerates feedback loops. Everyone sees the same picture, and everyone contributes to trusted outcomes.
5: Measure Business Impact, Not Just Technical Metrics – It’s tempting to measure observability in purely technical terms such as the number of alerts generated, data quality scores, or percentage of tables monitored. But the real measure of success is its business impact. Rather than numbers, organizations should ask if it resulted in fewer failed AI deployments. Created a faster time to insights and decisions? Reduce regulatory or reputational risk? Establish higher trust in AI/ML model outputs by executives and end users?
By framing observability metrics in terms of outcomes, data leaders move the conversation from “IT hygiene” to a strategic enabler of AI success.
Why the era of “good enough” data is over
(kkssr/Shutterstock)
As AI becomes embedded in every business process, data must always be trustworthy, representative, and continuously monitored. The days when data was considered good enough are over because AI demands more. Data observability provides the discipline to achieve this, not as a point solution, but as an embedded capability across DataOps, governance, and business teams.
Organizations that follow these five steps will find that observability accelerates AI adoption, safeguards trust, and unlocks faster value. Those that don’t risk joining the majority of companies facing AI projects that stall before delivering meaningful results.
About the Author: Keith Belanger is Field CTO at DataOps.live with nearly 30 years in data. He has led multiple Snowflake cloud modernization initiatives at Fortune 100 companies and across diverse industries, specializing in Kimball, Data Vault 2.0, and both centralized and decentralized data strategies. With deep expertise in data architecture, data strategy, and data product evangelism, Keith has spent his career bridging the gap between business goals, technology execution, and community influence. He blends foundational principles with modern innovation to help organizations transform messy data into scalable, governed, and AI-ready solutions. Recognized as a Snowflake Data Superhero, Keith contributes actively to the data community through conference talks, blogs, webinars, and user groups.
