5 Ways to Prepare Your Data Infrastructure for AI

“Although organizations are keen on AI and its benefits, they don’t have the process and data governance required to deal with data challenges it brings along,” says K.J. Burke, Field CTO, Hybrid Infrastructure at CDW Canada.

In an enterprise setup, AI systems highly depend on the integrity and availability of organizational data to drive business outcomes. If this data or the systems supporting it are unfit for AI processing, it may result in unintended, potentially harmful results.

Picture a salesperson using AI to generate a new product brochure based on unreleased sales data. Without transparency and guardrails in place, they may end up divulging confidential information in the public domain, which can cause business risks.

Therefore, for enterprises planning to adopt AI, the first step is to gauge the readiness of their data infrastructure, which includes the physical and software components for consumption, storage and sharing of data.

Organizations noted that data governance, traceability, agility and scalability need to improve to enable an AI-friendly architecture in their IT environments. However, the lack of data infrastructure readiness is perceived as a barrier to fully realizing the value of AI investments.

“Increasing cloud adoption has created new data siloes that can make it harder to ensure security across the distributed data estate,” Burke remarks.

Hybrid IT architectures with low interoperability can give rise to data siloes where data stored on one system (for instance, public cloud) may not be readily accessible by the other systems (on-premises servers).

This data sprawl creates security challenges that cause bottlenecks for AI initiatives as it’s harder to protect the data across environments, each of which follow separate security measures.

To overcome this, organizations must create a holistic data management and governance strategy and work on consolidating security controls for various IT components. The key objectives of this strategy are described below.

Encrypt your source data

AI systems interact with large amounts of data, often transferred between systems or stored in the cloud. Encryption protects the data during these transfers and in storage, ensuring that the data remains unreadable and secure even if a breach occurs.

Encryption also helps protect against insider threats. While employees or administrators may have access to systems, encrypting the data ensures that even if someone with access attempts to misuse the data, they cannot easily decrypt it.

Implement robust RBAC and identity management

There is a special need to build overarching security policies comprising role-based access control (RBAC) and identity management to prevent sensitive data from falling into the wrong hands.

Organizations must configure data storage systems with fine-grained access controls, both for AI and human agents. Whether it’s training data in the cloud or local files on an employee’s PC, each data request should be validated before granting access.

Using a centralized access control solution can go a long way in enforcing security policies across hybrid architectures. It can help implement security policies across IT components with a unified control plane that’s easier to manage and secure.

Secure your AI models

Many organizations fine-tune open-source foundational models to build their own generative AI applications, which can be riddled with vulnerabilities. The same issues can also be found in custom-trained models.

Equipping models with adversarial defence mechanisms is critical to prevent attack scenarios like model inversion. Such attacks can deceive the model into revealing sensitive business information.

At the same time, it’s essential to secure access to environments where AI models are stored and hosted to prevent infiltration attacks.

Ensure third-party vendors pass security checks

Smaller organizations that cannot train or fine-tune their own models usually rely on third-party vendors such as OpenAI to obtain AI features. These vendors may be able to access your organizational data to produce AI outcomes.

If you’re working with third-party vendors, ensure that your data and AI pipeline adhere to strict security practices, including encryption, access control and regular security reviews.

“Data curation is critical as organizations look to drive value both for analysis and AI asset creation. Also, as organizations better curate and refine their data, it becomes even more valuable. So, it is more critical that data resiliency is improved.” Burke says.

As per the Canadian Hybrid Cloud Report, 35 percent of respondents reported they need functionalities such as data masking, redaction, etc., that could prevent sensitive data from becoming a part of AI training data.

Whether an organization trains its own AI models or leverages pre-trained models in conjunction with retrieval augmented generation (RAG), sensitive data should be kept away from the process.

Personally identifiable information (PII) can leak into the datasets used for AI training, which can put customer trust at risk. If customer-facing AI systems are trained on or have unchecked access to such data, the chances of mishandling data become more significant.

That’s why organizations must incorporate data curation techniques to reduce the risks of leaking sensitive data into an AI model. Here are some techniques that could help:

Data masking and redaction

Data masking techniques replace sensitive elements like names, social security numbers or credit card numbers with pseudonyms or random characters before entering the data into an AI model.

Redaction, on the other hand, completely removes or blacks out sensitive portions of the data. These techniques ensure that the data used for training is representative but anonymized or de-identified.

Data anonymization

Data anonymization techniques transform datasets so that individual or sensitive details cannot be linked back to the original data sources.

Anonymization removes or obfuscates direct identifiers (such as names, addresses and phone numbers) and indirect identifiers (such as gender or zip codes) to prevent re-identification of individuals.

Advanced techniques like differential privacy ensure that aggregate data is anonymized while still providing accurate insights.

By using these techniques, organizations can effectively prevent sensitive data from being used for AI training while still extracting valuable insights from their datasets.

When we talk about good quality data for AI training, we want data that’s accurate, complete and consistent. As Reginald Hernandez, Field Solutions Architect at CDW Canada remarks, “Quality of data and analytics are the core foundations of effective AI decision-making.”

Poor-quality data can lead to biased or incorrect results, affecting decision-making, resulting in lost opportunities, inefficient processes and even compliance issues. 

According to the Canadian Hybrid Cloud Report, more than one-third of organizations (36 percent) said they will prioritize data quality for decision-making in the next 12 months.

Here, data quality has two different meanings in an AI system:

• Quality of the data used to train an AI model: Selective datasets that are specifically curated to train a model on specific capabilities such as data analysis, speech recognition, etc.
• Quality of the data an AI model interacts with: General organizational data such as spreadsheets, documents, etc., that an AI model can interact with to accomplish given tasks. This data may or may not be a part of the AI training data.

If you’re training your own AI model, it’s critical that you use clean data so that it doesn’t contain impurities in its learning. On the other hand, if you use a pre-trained model like ChatGPT, you must not subject it to poor-quality data as it may spew out inaccurate responses.

Improving data quality involves a systematic approach that ensures the data fed into AI models, and the data they interact with, is fit for consumption. This process typically includes profiling, cleansing, standardizing and monitoring data.

Data profiling

Data profiling is essential to assess the current state of the data and understand its structure, quality and statistical characteristics. Data engineers use techniques such as statistical summaries to reveal the spread of data across an organization, which helps them build a transformation roadmap.

Data cleansing

Data cleansing involves fixing the inconsistencies in data to make it ready for AI use. The process is highly subjective and depends on what’s wrong in a given data set. Missing values are added, duplicates are removed and false datapoints are corrected to bring the data to a certain quality benchmark.

Data standardizing

Standardization ensures that data conforms to consistent formats and units across datasets, which is critical for uniform AI model training. The focus is on making data values across various formats such as currency types, dates, addresses, consistent throughout the data set.

Data monitoring

Data monitoring ensures that data quality is maintained over time. Organizations must set up data quality metrics to make monitoring viable alongside automated quality checks to catch any decline in overall data quality.

“Before even introducing AI into their IT environment, organizations need to start building a data management plan that can align people, process and technology for curbing AI risks. Establishing an AI centre of excellence (CoE) that allows for a shared approach among the IT and different business teams can help organizations in identifying synergies, addressing interdependencies and working toward common business objectives,” Burke says.

A data management plan can help organizations not only ensure that their data is of high quality, trustworthy and secure, but also adheres to legal and ethical standards.

Two critical components of this plan are data management policies for creating a single source of truth, as well as robust compliance frameworks that can help control data privacy, usage and protection.

Master data management (MDM)

Master data management helps centralize and standardize critical data across the organization, creating a single source of truth. By consolidating master data (such as customer profiles, products, etc.) into a unified system, MDM ensures consistency, accuracy and reliability across all AI-driven applications.

It also helps improve data readiness with the following benefits.

• Improved data consistency: Ensures that all systems in an organization refer to the same consistent set of data points by providing a holistic data view, often referred to as a 360-degree data application.
• Fewer data anomalies: Helps in removing duplicate records and maintaining data integrity, reducing the risk of skewed AI model training caused by redundant or conflicting data.
• Better data integration: Creates a framework where data from different sources (ERP systems, CRMs, databases, etc.) can be integrated seamlessly.

Data compliance policies and frameworks

Building a strong compliance framework is vital for ensuring that AI models adhere to provincial or national regulations, avoid penalties and maintain trust with users.

The three steps organizations can take to build data compliance are as follows.

• Regular data audits: Conduct regular audits to ensure that AI models and the data they rely on are compliant with all relevant regulations. This involves checking for the use of personal data, adherence to retention policies and monitoring data processing activities.
• Data provenance and lineage: Track the flow of data throughout its lifecycle, from collection to processing and use in AI models. Data lineage helps organizations ensure transparency, allowing them to trace how data was sourced and used in AI systems.
• Ethical AI: Establish ethical guidelines for AI model development and use, ensuring that AI systems do not reinforce biases, discriminate or invade privacy.

Even after investing in the underlying infrastructure, AI is still unchartered territory for many Canadian organizations. They need to run several test pilots and PoCs before they can be confident about an organization-wide AI implementation.

This is where the expertise of AI solution providers, such as CDW, comes into play. They bring the necessary experience, skills and knowledge to meet the unique data challenges of an organization.

AI solution experts can help organizations navigate several key aspects of the data readiness journey.

• Data governance framework: Establish policies and procedures for data management, ensuring compliance with privacy and regulatory requirements.
• Identify generative AI use cases: Work out the most fitting generative AI use cases that can make the best use of organizational data.
• Change management and skill development: Adapt to new data-driven workflows and prepare teams to effectively manage data for AI projects.
• Control the costs associated with AI projects: Guide IT teams on how to optimize costs for AI infrastructure, talent and licensing for sustainable AI development.

“We often see scenarios like cloud sprawl, where organizations massively expand their cloud resources, which can inflate cost burdens quickly. This unchecked growth can take place due to lack of expertise or understanding of complex cloud projects.

“CDW can help organizations take a strategic approach to deploying an AI-ready infrastructure that’s cost-effective as well as helps them achieve their intended AI outcomes fitted with compliance and governance,” says Hernandez.

AI solution experts can help organizations navigate change management, a key aspect of integrating AI into workflows without disrupting existing operations. They facilitate the adoption process, ensuring employees are equipped with the skills and understanding needed to work alongside AI systems.

Our long-established expertise in the hybrid infrastructure space and key technology partnerships with leading vendors can help you kickstart your AI projects confidently. Whether you want to train an AI model from scratch, implement risk-free AI workflows or source data centre technology, CDW can help you meet your unique needs.

You can also explore the key trends and strategies, and receive key takeaways to help you succeed, in our 2024 Canadian Hybrid Cloud Report.