The Messy Truth About AI Data Management (And What to Do About It)

Tue, 12 May 2026 16:23:42 Z

Data will always be unclean. It's just a matter of degree.

I internalized that on day one of my master's program in data science, when a professor warned us that roughly 80% of our time would go to preprocessing and cleaning, not building models.

Years later, as Principal Product Manager for AI, ML and Analytics at Ivanti, I've found the guidance holds up remarkably well in practice.

As my team and I work to bring AI out of the lab and into production for IT and security teams, AI data management matters more than ever. Ivanti’s 2025 Technology at Work Report found that 42% of office workers use generative AI tools at work, up 16 points in a single year. Among IT professionals, adoption reached 74%.

The appetite is there. So is the hesitation. Many IT leaders know their data isn’t clean, their systems are fragmented, and their governance hasn’t caught up. The good news: you don't need perfect data to adopt AI.

You need a clear data management for AI strategy built around what you already have.

Why IT data is never perfect

In enterprise IT, data quality issues aren't anomalies. They're the baseline reality of AI and data management. Tickets get categorized inconsistently. Asset inventories are incomplete. Critical information lives in silos across systems. And unstructured text in support tickets and survey responses defies neat categorization.

Ivanti's research confirms how deep this goes. Our 2026 Autonomous Endpoint Management Advantage Report found that 89% of IT professionals say siloed data negatively impacts operations, with 39% saying silos cause inefficient resource use.

Our Tech at Work Report tells a similar story:

38% of IT professionals cite tech complexity as a significant barrier to effective operations, up four points year over year.
Nearly half (46%) say new software deployments actually drive-up ticket volume rather than cut through the noise.

Add that 48% of organizations still run end-of-life software, and the picture becomes clear: this is a data environment that's messy by design.

As David Pickering, Ivanti's Product Marketing Director, told me: when data is formatted differently across systems, entered inconsistently, siloed by department, shaped by years of acquisitions, you’ll find agentic AI workflows that span those systems quickly run into trouble. You can't tell an AI which data to trust if you don't know yourself. And without that foundation, even well-designed automations will fall apart at the seams.

In other words: "Garbage in, garbage out" still applies. But pristine data isn't coming anytime soon. Any serious approach to master data management and machine learning must account for the mess, not wait for it to resolve itself.

The decision framework — choosing your data management strategy

There are two primary paths for data management for AI in IT. Both are valid, both have trade-offs, and many organizations will use both for different use cases.

Path 1: Manual/programmatic cleaning

When my team introduced ticket classification for Ivanti’s ITSM system, we were training a model to categorize service requests. That demanded clean, well-labeled training data. So, we built a step into the workflow that gave administrators the opportunity to review and clean data before it fed the model. That human review made a measurable difference in accuracy.

This path works best when you're training or fine-tuning a custom model, ingesting data into a knowledge base or working with structured datasets where quality standards can be defined. The trade-off is time and resources. The outcome is high accuracy and full control.

It also works best when baseline data hygiene is already in place. Many organizations aren't there yet: just 35% track device age or location, and only 37% track patch status.

Path 2: Generative AI processing

Sometimes manual cleaning isn't feasible. I learned this working on Ivanti's survey analytics. Survey responses are some of the messiest data any IT team encounters: freeform text, inconsistent formatting, wildly varying detail. Cleaning that manually at scale isn't realistic.

Instead, we used large language models to identify themes, patterns and sentiment across incomplete and unstructured inputs. We could summarize entire surveys, flag satisfaction drivers, and surface actionable insights fast.

This path is ideal for high-volume unstructured data, situations where manual cleaning simply isn't possible, or any scenario where the cost of cleaning exceeds the value of the output. It does require access to capable large language models and validation that the use case is a fit.

Choosing between the two strategies

The decision comes down to data volume and variety, time constraints, accuracy requirements and how much control you need over where your data goes and how it's processed.

Fine-tuning a model where precision is critical? Invest in cleaning. Working with large volumes of unstructured input where speed matters? Lean into generative AI. The goal is deliberate choice, not inaction because the data isn't perfect.

Building AI-ready infrastructure for data management

Cloud services are essential here, and I don't say that lightly. When my team built a digital experience score to measure, quantify and improve digital employee experience, cloud was the critical enabler. It served as our integration hub, bringing together service tickets, device telemetry, application performance, and security signals.

That level of multi-source integration isn't feasible at scale without cloud infrastructure. Cloud also enabled us to run a hybrid AI model that processes both text and numeric telemetry simultaneously. Supporting thousands of devices and users at that complexity level isn't feasible on-premises.

Beyond compute, AI-ready infrastructure means tackling master data management for machine learning. Organizations need a single source of truth across systems. Data formats need to be standardized, particularly when growth through acquisition introduces legacy platforms with different conventions.

Data governance complicates the picture further. Regulations like GDPR and CCPA impose strict requirements on how personal data is processed and where it can be transmitted. For global organizations, that means AI pipelines need to account for regional jurisdictional differences, particularly when evaluating whether to use external AI services or keep processing in-house.

Our Autonomous Endpoint Management research found that just 32% of IT professionals use a unified endpoint management system. Without consolidated visibility, AI and automation can’t reach their potential. Effective AI data management starts with visibility: you can't automate what you can't see.

Best practices for IT teams implementing AI

When it comes to data management for AI, adopting tools without developing the processes to support them is one of the most common mistakes I see.

Establishing Knowledge Management Practices

Ivanti’s ITSM platform uses AI to generate knowledge articles from past tickets and incident resolutions. The productivity gain is real. But it doesn't eliminate the need for management discipline.

Articles still require review and approval cadences, version control and clear ownership.

Despite 86% of IT professionals agreeing that AI is important to efficient operations, fewer than half use it for high-value scenarios like predictive maintenance or automated incident response. The gap in AI and data management isn't technology. It's process maturity.

Validation and governance

Validation is just as important on the output side as data quality is on the input side. AI-generated results need to be checked, especially as organizations move toward agentic AI, where autonomous systems act on decisions in real time. The question isn't just whether the data coming back looks right. It's whether the system is taking the right actions.

Measuring AI performance matters too: how often it's being used, how accurate it is and where it's failing. Ivanti's 2026 State of Cybersecurity Report found that 92% of security professionals say automation effectively reduces mean time to respond. That effectiveness, though, depends on continuous monitoring and tuning.

Using AI as a catalyst for better data practices

AI doesn’t just consume good data practices. It drives them. By lowering barriers to content creation and analysis, AI frees teams to build the governance frameworks they’ve deferred. When generating a knowledge article takes minutes instead of hours, the team can invest that time in approval workflows and quality assurance.

This is especially valuable when junior technicians get real-time AI guidance, enabling them to contribute at a higher level while senior staff focus on strategy.

Our Autonomous Endpoint Management Advantage Report found that 62% of IT professionals feel overwhelmed by day-to-day operations, and one in four say a colleague has resigned due to burnout. AI that augments human expertise helps teams scale without that cost.

The path isn’t always clear, but the strategy can be

Perfect data is a myth. That shouldn’t stop you.

Manual cleaning for structured, high-precision use cases. Generative AI for unstructured, high-volume scenarios. Both require intentional investment in cloud infrastructure, governance and process development.

As AI models continue evolving, incorporating not just statistical pattern recognition but explicit rules and structured reasoning, the barrier to AI-ready data management will keep dropping. The organizations that move now, clear-eyed about their data’s imperfections and equipped with a strategy to manage them, will capture the most value.

Measuring Digital Experience with Artificial Intelligence

Thu, 21 Apr 2022 15:25:10 Z

With the expansion of remote and hybrid workforce and an ever more complex technology set, it is more important than ever to manage the digital employee experience. But how do we measure digital employee experience? One of the most common ways is to use surveys – for example post-incident surveys or annual customer satisfaction surveys. But these surveys, on their own, can be reactive, time-delayed and limited in scope. Is there a way to consistently, automatically measure digital experience that covers the broader IT functions?

The answer is yes! In the Ivanti Neurons Workspace 2022.2 release, we are introducing the Digital Experience Score (DEX Score in short). We start with a curated set of indicators that affect digital experiences, pump them into our proprietary scoring engine and generate a DEX Score between 0-100.

Let's look at the input to our scoring engine – indicators – first. Indicators can be, for example, metrics like hard disk storage, Boolean flags like antivirus protection and events like when a scan happened. We want a scoring system that is representative of the rich digital experience across key aspects of IT functions, so we selected indicators from service management, application, device and security.

Obviously, when it comes to our scoring engine, we can't reveal too much of our secret sauce, but what I can say is we use a hybrid model – meaning we use more than one statistical and machine learning technique to calculate the score.

Our scoring engine does not require system admins to manually adjust the threshold of what makes an indicator good or bad. For some of the qualitative device indicators such as hard disk storage, we analyze past readings and infer what the normal range of behavior is using statistical models. Once we know what the normal range is, we can calculate how different or similar a new data point is from normal behavior. This idea of quantifying normalcy versus irregularity is at the heart of our scoring engine.

But what about textual indicators like incident subject and description? How do we identify the latent emotion and convert them into a qualitative measurement? We decided to apply sentiment analysis to open incidents that are linked to a device. Our sentiment analysis model is a deep neural network to classify incident texts into positive, negative, and neutral sentiments. The model is pre-trained, so no additional training is required from customers.

Lastly, we do not just analyze the indicators individually; instead, we use a data mining algorithm that elegantly takes into consideration the interactions between those indictors. That means even though individual indicators can fluctuate and spike, or not tell you much on their own, together they tell us how much a device’s behavior deviates from normal. Another advantage of this data mining algorithm is it picks out which indicators are contributing to the abnormal behavior, so we surface them as potential issues.

If we do not use any of the fancy data science terms, what we are doing is comparing and correlating indicators across time and with each other, then condensing it into a single, holistic score. With no manual machine learning training and no threshold setting, are you ready for unprecedented insights into the digital experience across IT functions?

Ivanti Blog: Posts by