Generative AI has gained wide interest recently, with the explosion in the public domain of ChatGPT in November 2022. This ascending technology promises to solve many of our most intractable problems, and we’ve all been amazed at what large language models (LLMs) can do. The promise of generative AI, with models trained on more data than a human being could read in a lifetime, seems almost magical.
It’s no wonder many companies are joining the race to adopt AI and find creative ways to harness the power of the new technology. Unfortunately, not all companies will realize success. A recent study by Gartner predicts that a third of generative AI projects will be abandoned by the end of 2025. While AI adoption is accelerating, organizations are still figuring out how to manage it, and few companies have established the necessary infrastructure to deliver consistent ROI and long-term value and use.
Some challenges are common across different fields, but others are unique and must be addressed by leaders in their respective industries. Organizations urgently need to establish clear frameworks for effectively managing AI in their fields to ensure its responsible use, and organizations that do this first will have a clear advantage. An earlier example is the AI maturity roadmap established by companies like McKinsey and GMSA for the telecoms industry.
Other industries, such as finance, healthcare and legal, where client confidentiality and data integrity are of the utmost importance, have an even greater burden for responsible AI development. The legal industry, in particular, has an additional need and responsibility to demonstrate AI defensibility in courts of law.
There are three key areas to focus on when building out a company’s AI infrastructure to ensure long-term success: Effective data management, customer-centric AI development and cross-functional collaboration with AI governance.
Data Management
Having high-quality data is essential for leveraging AI effectively. Most models rely on recognizing features in the data for their tasks, but they will only be reliable if the datasets used to train them match the data for which they will be used. For predictions from human-labeled data, models will only be as trustworthy as the labels. Even the most sophisticated and advanced models are still subject to the rule of garbage in, garbage out. Couple this with the need to have large quantities of data for some models for training and evaluation, and the problem can be daunting.
The impact of poorly trained AI can be consequential. In the healthcare industry, flawed AI can result in misdiagnosis. In the finance industry, incorrect risk assessments may result. For a specific example, consider the complex litigation challenges many of the largest Fortune 500 companies and their legal departments face. Suppose these companies use AI for contract review that was trained on inconsistent or poorly managed data. In that case, they will likely make poor risk assessments, miss key clauses, and fail to identify substandard legal reviews and assessments.
Gartner estimates the average financial impact of poor data quality on organizations is $9.7 million annually. Training AI models using high-quality, relevant data enables more accurate insights and supports better decision-making, thus improving the effectiveness of AI deployment.
To address this key need for high-quality data for development and testing, companies should build scalable, adaptable data systems that will support AI as their business grows. It can be helpful to adopt advanced data management tools to automate the process of cleaning, processing, and organizing data. Finally, creating strong governance frameworks that define how data is collected, stored, and maintained across the organization will ensure AI models will be able to scale safely and reliably with the business.
Addressing Customer Pain Points
AI development requires large investments of a company’s time and resources to deliver new features to customers, so it’s critical for companies developing AI to be sure they’re aligned with customer needs. It may be tempting to roll out flashy AI features that, on the surface, seem magical, but sometimes, features that seem magical to the AI developers are not useful for the end user. Perhaps the features are useful, but it might not be obvious to users why or how they can benefit from them. In the legal industry, for example, AI can be used to facilitate document review or summarize depositions. If the AI features are not implemented in a way that addresses the specific needs of the legal teams, they simply won’t adopt the new technology. By focusing on solving real-world pain points, organizations can be sure to deliver immediate value to users in a way that users will adopt the new technology.
To address this need, companies can do user research to identify very specific challenges and use cases to guide AI development. Investing in continuous feedback loops will help to improve AI models and adjust features based on real-world usage and experiences. Lastly, focusing on user-friendly interfaces will allow the AI to enhance rather than complicate the user experience.
Collaboration and AI Governance
Regulations concerning the use of AI are very much still in the early stages, but understanding how regulations will evolve and being able to respond to future requirements will certainly present new challenges. Beyond federal or international rules, industries are developing standards or expectations for the responsible use of AI. Companies are also sensitive to the concerns or explicit requirements of their customers concerning AI. The use of AI raises new security concerns, particularly regarding how to ensure the safety of customer data when using generative AI models hosted by external vendors.
To stay abreast of regulation and other requirements, companies should create cross-functional teams within their organizations that bring together legal, data engineering, and AI expertise to inform AI strategy and development. Having diverse viewpoints to consider the space from all angles will help to prepare against likely future developments. In addition, implementing governance frameworks to provide oversight on model development will ensure fairness, accountability, and transparency throughout the process.
To conclude, successful AI deployment requires good data management, customer-centered design and implementation, and robust governance to ensure responsible and effective use of AI. Companies that do this will be better positioned to fully realize the potential of AI.