3 steps to get your data AI ready

While many organizations have already run a small number of successful proofs of concept to demonstrate the value of gen AI, scaling up those PoCs and applying the new technology to other parts of the business will never work until producing AI-ready data becomes standard practice. Gartner’s recent prediction that 60% of AI projects that run without AI-ready data will be abandoned by next year sheds light on the time bomb IT leaders need to immediately diffuse.

For many CIOs, preparing their data for even one AI project is a tall order. “As they embark on their AI journey, many people have discovered their data is garbage,” says Eric Helmer, chief technology officer for software support company Rimini Street. “They aren’t sure where it is among hundreds of different systems in some cases. And when they find it, they often don’t know if it’s in a state that can be used by AI. This tends to put the brakes on their AI aspirations.”

To prepare for the oncoming tsunami of requests to support the use of AI in projects across the business, CIOs should take these three steps to ensure making data AI-ready becomes standard practice.

Give up on using traditional IT for AI

“The ultimate goal is to have AI-ready data, which means quality and consistent data with the right structures optimized to be effectively used in AI models and to produce the desired outcomes for a given application,” says Beatriz Sanz Sáiz, global AI sector leader at EY. AI-ready data is not something CIOs need to produce for just one application — they’ll need it for all applications that require enterprise-specific intelligence.

Unfortunately, many IT leaders are discovering that this goal can’t be reached using standard data practices, and traditional IT hardware and software. “It’s nearly impossible to clean up data across a sprawling estate of disconnected systems and make it useful for AI,” says Helmer. “If you go into an HR system and delete duplicate records or clean up the data in any other way, the changes may not be propagated to all relative data stores, creating data inconsistencies.”

To regularly train models needed for use cases specific to their business, CIOs need to establish pipelines of AI-ready data, incorporating new methods for collecting, cleansing, and cataloguing enterprise information. Further Gartner research conducted recently of data management leaders suggests that most organizations aren’t there yet. Two thirds of the organizations included in the study of over 1,200 either don’t have the right data management practices for AI or are unsure if they do. So IT leaders who plan to increase AI adoption clearly need to rethink how they manage data.

So far, most organizations have been relying on traditional systems that were already struggling to support the production workload, according Jason Hardy, CTO for AI at Hitachi Vantara. Now with AI workloads added on, the result is a lot of downstream problems that affect normal day-to-day operations. CIOs need to revamp their infrastructure not only to render a tremendous amount of data through a new set of interfaces, but also to handle all the new data produced by gen AI in patterns never seen before. “The AI revolution is forcing a modernization of the data center across all industries,” says Hardy.

Modernization had already begun at scale by around 2018, according to Sáiz. New technology became available that allowed organizations to start changing their data infrastructures and practices to accommodate growing needs for large structured and unstructured data sets to power analytics and machine learning. They started using data virtualization, which reduced the need for large data warehouses by decoupling data consumption from origination. Now with agentic AI, the need for quality data is growing faster than ever, giving more urgency to the existing trend.

Use AI to improve data, and knowledge to improve AI

The good news is AI is part of the solution, adds Sáiz. For example, gen AI can be used to generate synthetic data, and other forms of AI can be used to help analyze and improve data quality. Some organizations use AI to analyze data distribution by identifying values that aren’t within a reasonable range, and then filling in missing values. AI can also help engineers localize problematic data sets, applying different techniques to determine the probability that a given value is realistic. “We’re seeing ‘AI for data’ as one of the largest applications of AI in the enterprise at the moment,” says Sáiz. “The fact that the revolution of data and the revolution of AI are happening in parallel produces a win-win situation.”

AI can also be used to enable a much more decentralized data infrastructure by having a centralized intelligence that employs agentic AI to manage the decentralized infrastructure. Hundreds of thousands of agents can enforce standards and ensure data consistency, which, according to Sáiz, is one of the biggest challenges companies face in regard to data infrastructure.

For example, AI can help ensure the systems of records of a particular client are consistent in all systems including CRM, contact center software, and financial applications. “To maintain consistency, whenever there’s a customer interaction with a contact center or with the web, all systems get the change in near real time,” says Sáiz. “Where you used to have more latency and lots of manual checks before, now it’s all driven by AI, which constantly checks on the state and the master data set to determine, based on intelligence, whether a record needs to be updated in the whole system.”

According to Sáiz, knowledge is becoming more important than data because it helps interpret the data. A knowledge layer can be built on top of the data infrastructure to provide context and minimize hallucinations. “If somebody in telco runs a forecasting model, the variables, inputs, and results will be different than running the same model for financial forecasting,” she says. “The more you focus on knowledge, the more accurate your AI.”

Apply an iterative approach to transformation

Some IT leaders feel overwhelmed by the challenge before them, thinking they need to get all their data in a perfect state before starting their AI journey. But Hardy says a better approach is for them to change their data management practices and infrastructure in an iterative fashion. “Once you put the foundational principles and practices in place, you can make the transformation one project at a time,” he says.

One of the foundational principles is cybersecurity, which is the primary concern for CIOs, according to Hardy. IT leaders need to make sure not only that the data used to train models doesn’t violate any rules around data privacy, but that models generate responses consistent with the access rights of the user. “AI systems need to know who’s asking the question so the right level of information is brought back and no additional information is exposed,” says Hardy.

The risk of exposing intellectual property also has to be mitigated, especially when AI is offered as a cloud-based service. “Based on how you interface with the service — and the types of data, sovereignty requirements, sensitivity requirements, and regulations — you’ll probably decide that some of the data should never be in the cloud,” adds Hardy. “Putting in place guidelines will help you decide on a case-by-case basis what stays on prem and what goes to the cloud.”

According to Helmer, a governing body should be established to help ensure best practices are being followed. Anybody who develops or deploys an AI application has to adhere to a set of rules that are consistent not only with data quality but retention policies, data dependency policies, and all appropriate regulation.

“As you go through the journey, decide the outcome you’re going for with each project,” says Hardy. “Then figure out what data you need, and the systems you need to interface with, to get to that outcome. Instead of trying to boil the ocean before you see any return, focus on your data transformation one outcome at a time.”