Enterprise AI—Principles and Best Practices

Enterprise AI refers to the application of artificial intelligence to enhance business operations within large organizations. Most organizations start with data science teams experimenting with AI for various use cases. The rules are loose, and the stakes are low at this stage. However, transitioning from a proof of concept to a production environment requires deeper thought and design considerations. Enterprise-level AI projects must meet three critical requirements:

1. Maintain consistent performance at scale.
2. Meet all legal and ethical requirements for the AI use case.
3. Transparent in their decision-making to build user trust. 

AI in production involves implementing bias mitigation measures, ensuring the explainability of AI decisions, and adhering to ethical standards. The following sections explore frameworks and concepts to ensure that enterprise-level AI projects are robust, scalable, and trustworthy.

Summary of key enterprise AI concepts 

Enterprise AI strategy

Many successful enterprise AI strategies follow a phased approach that treats AI operationalization as another change management process. It allows for greater user adoption and adherence, which is crucial for success. For example, Dana Farber Cancer Institute recently deployed GPT-4 to over 12,000 employees over six months. They started with specific use cases for a small group of advanced users, gradually expanding based on lessons learned from this cohort.

Your enterprise AI strategy should pinpoint key areas where AI enhances efficiency, reduces costs, and generates new revenue streams. Consider the following key factors.

Powering data engineering automation for AI and ML applications

Learn how Nexla helps enhance LLM models

• 
  
  
  Enhance LLM models like GPT and LaMDA with your own data
• 
  
  
  Connect to any vector database like Pinecone
• 
  
  
  Build retrieval-augmented generation (RAG) with no code

Use case

What specific business problem is the AI intended to solve? AI can broadly automate or augment tasks, targeting higher efficiency or cost reduction. Aligning AI initiatives with business objectives ensures that AI investments yield tangible business value.  Any project must create long-term business impact and revenue at the enterprise level.

Data

What data is needed, and how will it be sourced and accessed? Consider whether the data is proprietary or public. If public, is it copyrighted, or does it fall under fair use provisions? If proprietary, does the AI application restrict access only to authorized users?

Hypothesis

What assumptions are being made, and how will AI address these assumptions? These will vary depending on the business context.

AI model

What type of AI model is suitable—predictive, generative, or descriptive? This decision is tied to the business problem and hypotheses. The choice of model, ranging from text-based to multimodal LLMs, influences further decisions like building versus buying the model and self-hosting versus using APIs.

Actions

What actions will the AI drive, and how will these actions impact business processes? Remember to factor in access controls and privileges for the AI model.

Outcome metrics

What key performance indicators measure success? Metrics should evaluate both the project and the model’s performance. For the project, include KPIs for roadmap adherence, milestones, user adoption, and revenue targets. For the model, focus on accuracy, security, privacy, and other customer success criteria.

Enterprise AI governance

As AI is a socio-technology, it carries ethical risks alongside traditional ones. AI systems raise concerns regarding bias, data ownership, privacy, accuracy, cybersecurity, and other ethical risks. These issues can lead to discrimination, loss of consumer trust, and even material and physical harm. Governance mechanisms aim to mitigate these risks and realize business value with AI.

Ethical concerns in AI

Robust governance frameworks are critical to successful enterprise AI adoption. They ensure organizations develop specific policies on responsible use, development, and deployment. They also consider practical ethical, moral, and regulatory compliance aspects. AI governance is not a one-time effort but an initiative that evolves with an organization’s AI maturity.

Regardless of maturity, a comprehensive AI governance program specific to your context is vital. It involves a cross-functional team developing and implementing the following key components.

Principles, policies, and guidelines

Establish clear and comprehensive principles, policies, and guidelines for the entire AI lifecycle, from development to procurement, deployment, and use. Many governments and international organizations have laid out AI principles and guidelines. Here are some examples:

OECD’s AI Principles

The Organisation for Economic Co-operation and Development (OECD) developed the first intergovernmental standard on AI in a participatory manner, getting input from experts from diverse domains. OECD’s AI Principles aim to guide AI actors in developing trustworthy AI by covering themes such as fairness and privacy, explainability, accountability, security, etc.

IEEE’s ethically aligned design 

Created by The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, IEEE’s Ethically Aligned Design provides recommendations for corporates to support ethical systems and practices around AI funding, development, or use. 

While these frameworks provide high-level guidance, enterprises must develop their own AI principles based on their industry, domain, and the types of AI systems being deployed or procured. Your principles should be linked to your organization’s missions and values and guide the types of AI systems and use cases it should pursue.

Risk assessment and management

Develop a comprehensive risk assessment and management framework to identify, assess, and mitigate potential risks associated with AI deployments, including:

• Technical risks—identify and mitigate risks related to technical failures, security breaches, and system vulnerabilities.
• Ethical risks—assess and address ethical risks such as unintended bias, privacy violations, and misuse of AI systems.
• Operational risks—evaluate and manage risks related to the deployment and ongoing operation of AI systems, including data quality, model drift, and human-AI interaction.

Managing these AI risks is challenging. While many frameworks are available, The Artificial Intelligence Risk Management Framework (AI RMF v1.0) developed by the National Institute of Standards and Technology(NIST) is considered the standard for AI risk assessment and management frameworks.

Core functions according to NIST AI RMF v1.0(Source)

The framework offers insights into mapping, measuring, and managing risks within the governance framework. Perform inventory due diligence and risk assessments, including third-party products/services. 

Compliance mechanisms

Establish compliance mechanisms to ensure adherence to relevant legal and regulatory standards, including:

• Personal health information is subject to data protection and privacy laws, such as the General Data Protection Regulation (GDPR) or the Health Insurance Portability and Accountability Act (HIPAA). 
• Intellectual property and copyright laws 

Several high-profile legal disputes have occurred between AI companies and copyright holders over the use of copyrighted material to train AI models, like The New York Times vs. OpenAI, which is still ongoing at the time of writing this article. To avoid such legal risks, develop processes for regular audits, documentation, and reporting that demonstrate compliance.

Governance processes and structures

Typically, organizations establish a cross-functional AI governance board with a clearly defined operating model, roles, and responsibilities. The board is responsible for oversight, decision-making, and accountability within the AI lifecycle. For example, the board can

• Incorporate ethical considerations and risk assessments at every stage of AI development.
• Implement mechanisms for continuous monitoring and evaluation of AI systems.
• Develop incident response and remediation processes to address AI-related issues or failures.

Governance structures thus actively minimize harm and ensure accountability.

Engineering enterprise AI—from principles to practice

In enterprise AI, the focus often falls heavily on developing AI models. However, neglecting data engineering can result in sub-par outcomes despite having advanced AI models.

Data readiness for enterprise AI

Poor data quality or unoptimized data pipelines can lead to inaccurate model predictions and affect the overall performance of AI solutions. A good data engineering strategy involves a comprehensive approach to managing the entire data lifecycle, including data collection, transformation, and pipeline management. The concept of a data flywheel can be applied to achieve and maintain high-quality AI outputs. 

A data flywheel ensures continuous improvement through a feedback loop: data flows into the system, is transformed into structured formats, and is fed into AI models. The outputs of these models are then monitored for performance metrics, and any degradation in performance is flagged. Insights gained from monitoring are used to refine data ingestion and transformation processes, thus enhancing data quality. Improved data quality leads to better model performance, which, in turn, generates more reliable insights and decisions.

Figure: The data flywheel cycle

The benefits of the data flywheel are continuous improvement, scalability, and resilience. The automated nature of the process allows the system to handle increasing data volumes without much manual intervention.

For instance, the monitoring and feedback stages of the data flywheel help detect changes in data patterns, known as data drift. In Retrieval-Augmented Generation (RAG) applications, the system can quickly adapt to new trends or shifts in the underlying data, maintaining the relevance and accuracy of the generated content. We have covered RAG in production in detail here.

Integrating ethical AI in engineering

Integrating ethical AI practices into the daily workflows of engineering and design teams is challenging. It involves:

• Ongoing team training on responsible AI practices
• Implementing tools and processes to embed ethical principles into system design and auditing.
• Fostering a safety-first work culture that combines diverse expertise across engineering, design, legal, ethics, and domain teams throughout the AI lifecycle.

Moving from enterprise AI strategy to practical implementation

Recommendations for enterprise AI development

We recommend the following for successful enterprise AI development.

Establish strong data pipelines

Data quality is a key factor for enterprise AI success — incomplete or missing data, inaccuracies and inconsistencies, etc. affect the outcomes of AI-assisted systems. If your data is siloed, not transparent, not adequately governed and documented, it isn’t ready for AI. 

Don’t let data become a bottleneck in enterprise AI projects. Nexla’s Data as a Product makes moving your data from anywhere to your vector databases easy with speed and reliability. Nexla also automates time-consuming data engineering tasks such as managing data credentials, integrating data, transforming, and delivering ready-to-use data.

Build robust AI infrastructure

The importance of AI infrastructure lies in its role as the foundational backbone to develop, deploy, and scale artificial intelligence solutions effectively. A robust AI infrastructure is crucial for several reasons. It ensures efficient ingestion and handling of large data volumes for training and maintaining AI models. It provides the computational power necessary to process complex algorithms and run large-scale machine learning models. AI infrastructure supports integrating and deploying AI solutions across business functions, be it through APIs or software products. 

You can read our chapter on AI infrastructure, which covers the subject in detail.

Implement security throughout your LLM lifecycle

Algorithmic filters and controls should be employed during model training, input prompts, and output generation. For example, you can implement:

• Input filters to prevent risks like prompt poisoning
• Output quality checks to avoid hallucinations or toxic content
• Privacy controls to restrict data access.
• Explainability tools for transparency
• Observability monitors to detect deviations from expected patterns. 

Please refer to the article on LLM security for an in-depth discussion. 

Incorporate human involvement and feedback

Automated safeguards alone are insufficient. Human oversight and interventions are equally crucial, especially for high-risk AI applications. Prior to launch, rigorous stress testing and “red teaming” by diverse experts in security, adversarial machine learning, and responsible AI should be conducted to uncover vulnerabilities and edge cases. Dedicated toolkits like Fiddler Auditor and PyRIT can assist but cannot replace manual testing.

Similarly, after deployment, ongoing human monitoring is vital. Abnormalities detected through algorithmic monitoring can inform and focus these human review processes. You can consider including live support to intervene when models misbehave. Periodic auditing of model responses and controls to roll back or take down problematic models is a must.

Conclusion

Taking AI systems from proof-of-concept to production deployment is a complex endeavor that requires robust safeguards and processes. Comprehensive human oversight through governance structures, monitoring, and auditing is essential, especially for high-risk AI applications. Data engineering must be prioritized to succeed with AI.