Uncovering the Hidden Risks of Poor IT Data in AI Deployments

I’ve worked in IT for over 20 years, and I can tell you one thing with certainty: if your data is a mess, your AI deployment is going to be a mess too. No matter how powerful your tools are, bad data will tank your results faster than you think.

In fact, Gartner predicts that by 2028, 99% of Infrastructure and Operations-led investment in agentic AI without system data and operating model improvements will fail to achieve sustainable return on investment. That doesn’t surprise me.

You might be wondering how that could be possible, how nearly every single AI deployment can fail just because of bad data–but think of it this way. The reason is simple: when data multiplies, mistakes do too.

Before AI, you entered one input, and you got one output. Even if you entered five inputs, you typically still got one result. That's all multiplied with artificial intelligence. With AI, you can enter one input but get millions of outputs. Whatever mistakes bad data caused before, they are exacerbated with AI.

If they're going to fuel successful AI initiatives, IT leaders need to not only access clean data but figure out what they have in the first place and how they want to specifically use that data.

Bad Data Kills Operational Impact

You've heard the phrase “garbage in, garbage out,” and never has it applied more strongly than we talk about bad data and AI. Unfortunately, the problem doesn't just affect a small number of IT teams. I've worked with countless IT teams, and I don't think I've ever met a CIO, ITVP, or IT Director who thinks they have great data.

Poor IT data isn't just an inconvenience. It undermines IT service and asset management, leading to “inaccurate impact analysis, delayed issue resolution, and inefficient change management in Dynamic environments,” per Gartner.

There are three areas where I see this go wrong most often.

Automation Failures

You'll never get good suggestions or results without good data. Remember that AI learns the longer it runs. As you try to implement automations using inaccurate or incomplete information, the model can learn to do things that don't actually work for the business.

Without accurate, real-time data to inform algorithms, you're stuck with:

• Automation Inefficiencies: If the AI is using incorrect data, your processes won't work as you intended. Missing pieces of data can inform better outputs, but you may not be able to manually figure out what those data points are.
• False Outputs: The AI can only work with what it has access to, and if the data is working against itself, you'll never see successful outputs.
• Incomplete Processes: In some cases, the AI won't have enough data to be able to complete automation tasks in the first place.

In the end, you'll have spent all this time trying to create these automations to improve your operations, but you'll end up spending more time fixing them when poor data doesn't yield the results you expected. Without trusted data, automation becomes guesswork.

Security Vulnerabilities

Agentic AI is designed to run in the background. The whole point of using this form of AI is to need little to no human interference. However, that means you need to know that the data you're feeding it isn't opening you up to any moral, legal, or technical security risks.

To that end, you have to ask the unexpected questions: what happens when the AI decides that it's more cost-effective to drop certain security measures on certain networks because it will allow more throughput or reduce the number of calls? That’s why data governance, permissioning, and audit visibility are non-negotiable.

There's also the risk of hallucinations or bias. Since agentic AI connects with multiple systems, as it takes in more data, you open yourself up to those risks if the data being used is incorrect or insufficient.

To avoid this, and ensure security never dips, IT teams need to not only improve data governance but also set permission controls and access restrictions to limit the exposure of sensitive data that can be used to improperly feed AI algorithms.

Budget Overruns

I’ve worked on projects where we had to justify every dollar spent on IT improvements. AI raises the stakes. According to recent research, businesses using bad data in AI models saw global revenue impacts of over $400 million on average. How are you supposed to justify that?

When bad data produces bad models, you have to spend time fixing those. You waste time and money on these deployments that don't come to fruition. In some cases, the AI can autonomously decide to work in a certain way, scaling processes or making purchases way outside your budget. But you still have to go back to your executive board with nothing to show for your efforts and explain those overruns.

That's seriously concerning when so many IT operations struggle to find the “sweet spot” of demonstrating short-term and long-term value. I’ve seen the tension firsthand. There's just this dichotomy between “‘I need some results” versus “this is a great idea.” Even if you know developing an AI model will improve things long-term you're not really in the business of AI; you're in the business of delivering IT.

If you do get approval for an investment in AI deployments, but you go over that budget without good results, good luck getting approval for your next big initiative.

You may be thinking these are abstract problems that may not necessarily affect your business, but poor data has caused severe consequences for even some of the biggest tech giants.

Lessons Learned the Hard Way

AI is based on trust–trust in the data and outputs. Without that trust, you not only face failed deployments, but you hinder future adoption as people build skepticism in AI’s ability to do what they want.

For example, consider enterprises who heavily rely on CMDBs to manage their IT asset data. The techniques used to manage these databases are static and require heavy manual input that many IT teams don't have the resources to keep up with.

Essentially, almost immediately after data is added to the system, you end up with gaps in accuracy that, according to Gartner, “may lead to obsolete data, misaligned resource tracking and inadequate service visibility, affecting decision making, troubleshooting and the agility of IT operations.”

Those bad results mean 63% of businesses don’t trust their CMDB to provide accurate, up-to-date information.

We’ve already seen bad AI deployments play out in real-life happenings.

Air Canada’s Chatbot Lied
A passenger sued Air Canada after they were given false information by their chatbot and forced to spend hundreds of unexpected dollars on a flight. While Air Canada argued they couldn't be held liable for the information provided by a chatbot, that was rejected in court, and they were forced to pay costs and damages to the passenger.

ChatGPT Made Up Fake Cases
A lawyer attempted to use ChatGPT to find cases to support his argument, but the platform gave him cases that didn't exist, with false names and docket numbers. When that was discovered in court, the attorney was fined and the case was later dismissed, losing the plaintiff any chance at damages.

IBM Watson Couldn’t Comprehend
IBM Watson was intended to provide AI-driven cancer recommendations to doctors. Unfortunately, the AI was only trained on hypothetical cases, not real ones, so it couldn't understand complex medical situations and provided inaccurate, potentially dangerous recommendations instead.

You might be thinking, “These are very specific cases that don't apply to me or my business,” and maybe you're right. But let me throw out a scenario I can imagine happening to any IT team.

Let's say you plug in some code to allow your AI to make autonomous decisions on how you scale your cloud infrastructure. It's pulling data, albeit in complete data, to build its “mindset” and eventually goes very “iRobot” on you, taking things you never thought of–because you're human–into consideration.

Now it determines certain dimensions should make certain decisions, and all of a sudden you have a bill for $13 million when you thought you were going to pay $5 million. It technically did what you asked: used the data at its disposal to scale your infrastructure. It just lacked the financial data to do it in budget.

So how are your IT teams meant to access stronger, more reliable data, mitigate data risk, and have successful AI deployments?

Strategies to Avoid the Chaos of Bad Data

Establishing a solid foundation means re-examining the current systems and processes you use to manage IT data and adopting a new mindset that supports continuous long-term improvement.

In simpler terms? You need to know not only what data you have, but where it came from, who touched it, and how current it is.

Here’s what I recommend I&O leaders do to build a sustainable data foundation that supports agentic AI.

Work Out Your Value Areas

You may be tempted to wait until you have nearly perfect data to determine where you use it, but that wastes time. Perfect data doesn’t exist. In reality, there are going to be areas where you'll never give AI full control, but there are others that you can.

As you're working to clean up your data and improve visibility, determine areas where you'll actually get value and materially benefit from implementing autonomous AI. This can guide your processes and even influence where you make changes within your IT operations.

Modernize Legacy Systems

Gartner research shows “an average of 41% of systems are beyond end of life or support and an average of 38% are incompatible with target architectures and new projects.”

IT leaders need to switch from legacy systems, like static CMDBs, spreadsheets, and patchwork integrations that lack the capacity to store, connect, and manage strong data. If you’re building on static systems, you’re building on sand.

Instead, you need to opt for systems that allow for:

• Federation: Instead of forcing data into one system, simply pull it from where it already lives.
• Consistent Enrichment: Make data more useful for automation by continuously adding new contexts and attributes.
• Real-time Synchronization: Allow data to flow across all systems, not only a single dashboard.
• Normalization: Resolve conflicts and discrepancies at the source and push them across all systems.

Strengthen IT Asset Management

Knowing where your IT assets are, their usage details, their maintenance needs, and their cost throughout the entire life cycle is vital for spend management, risk reduction, regulatory compliance, and the user experience. But, you need a way to manage and update that data across all your systems to reap those benefits.

Strong IT asset management enhances visibility throughout your entire technology ecosystem and improves data cleanliness and inventory control.

By integrating advanced ITAM solutions with your current IT, security, and business systems, you create a unified asset inventory. This eliminates data fragmentation, inaccuracies, and discrepancies caused by isolated management tools, allowing you to run your IT operations with a reliable system of record.

Improving Observability Practices

Many enterprise IT teams still lack the maturity needed to know what they have and where it is– but that's one of the core functions of their role. Even once they figure that out, they're typically not tracking the metadata associated with quantifying the results of their operations.

So many teams I talk to can tell me their processes, but not how well those processes work. Can you answer basic questions like, if you change a workflow, do you know if it improved anything? Are you tracking equipment return rates, time-to-value, or user satisfaction?

For example, consider the constant onboarding and off-boarding of people within your business.

When someone leaves, you want their equipment back. You can automate the notification that they need to return their laptop, but maybe you hypothesize that if you automatically send them a FedEx label, you'll receive the equipment back sooner. What if you automatically send a box as well? You need to track those results– and have the data to do it.

Then using that data, you can further improve AI algorithms and drive faster operations by identifying bottlenecks, moving on from a ticket-based system via automations, and spend less time debugging models to work as you want.

Deploy AI Built on Strong Data

The last thing you want is to be another statistic of a failed data deployment. You don’t have to be.

To minimize the risks and consequences that come from using incomplete, inaccurate, and outdated IT data, invest in solutions that connect your asset data and improve your visibility into their entire landscape.

You have the opportunity to avoid the chaos that comes from poorly-fueled AI initiatives.

Download Gartner’s Innovation Insight: Agent-Native I&O report to get expert recommendations on how to assess your readiness for agentic AI and stay ahead of challenges to empower successful deployments.