Custom Generative AI Models Are the Optimal Path for Enterprises
-
Custom enterprise generative AI promises security and performance advantages, but successfully developing models requires overcoming challenges in data, infrastructure, and skills.
Consumer-facing pre-built generative AI models like ChatGPT have attracted widespread attention, but customized models may ultimately prove more valuable for organizations in practice.
Most companies' generative AI initiatives are in early stages. According to the latest research from TechTarget's Enterprise Strategy Group (ESG), only 4% of organizations currently have generative AI in mature enterprise-wide production. The majority (54%) remain in early deployment, pilot, or planning phases.
As enterprises increasingly explore generative AI, many recognize the value of aligning models with their specific data and use cases. The same ESG survey reveals a preference for customization, with 56% of respondents planning to train their own custom generative AI models rather than relying solely on one-size-fits-all tools like ChatGPT.
However, successfully developing custom enterprise generative AI faces significant challenges in areas such as data management, security, and system integration. To address generative AI's risks and limitations while leveraging the advantages of custom models, enterprises need a targeted approach to deploying this emerging technology.
There are several compelling reasons to develop custom enterprise models:
Privacy and security risks of generative AI are a primary concern for enterprises. One IT/cloud manager in the finance, banking, and insurance industries described existing generative AI tools as "too much of a security risk" in response to ESG's survey.
"Many available AI technologies are free or enabled by existing vendors—organizations have no opportunity to proactively review the technology's risks based on data privacy, security, compliance, confidentiality, and intellectual property considerations," said another respondent, a business VP in the telecommunications industry.
When it comes to internal data used to train models, fine-tuned proprietary models can provide better oversight for security-conscious organizations. With internal models, organizations can maintain control over sensitive data rather than sharing access with third parties.
Models tailored to a company's specific tasks and data may also produce more relevant outputs with fewer hallucinations. This can alleviate some organizations' concerns about obtaining accurate, fair, and representative outputs from third-party models.
"There are concerns about the accuracy and completeness of AI reports," said a consultant in the healthcare and health services industry in response to ESG's survey. "How do we confirm and validate data sources? Additionally, there's the issue of algorithmic bias."
The process of training models on target datasets—here, information about the organization and its industry—is called fine-tuning and can yield more accurate results for relevant tasks. AI tools customized to address specific business problems and workflows can improve efficiency and reduce integration issues. In short, this means custom models may require less extensive oversight while producing outputs that better align with business needs.
However, adopting generative AI is not without challenges. ESG survey respondents identified widespread obstacles to generative AI implementation (custom or otherwise), primarily lack of employee expertise and skills (39%), ethical and legal considerations (e.g., bias and fairness) (32%), and concerns about data quality (31%).
Creating effective custom generative AI is an exceptionally complex task that requires overcoming three main obstacles: insufficient high-quality data, difficulties in integrating models into legacy systems, and the shortage of artificial intelligence and machine learning (ML) talent.
Without adequate high-quality, extensive, and well-integrated data, training accurate proprietary models can be difficult or impossible. Transforming messy enterprise data into usable training corpora is a labor-intensive process that involves building pipelines to ingest and prepare proprietary data for labeling and input into models.
This process of cleaning and labeling enterprise data is inherently resource-intensive. Additionally, constructing and maintaining model pipelines that can adapt over time requires substantial ML expertise, which remains scarce among current IT staff. A quarter of respondents cited technical complexity as a barrier to generative AI implementation in their organizations, and the limited supply of qualified machine learning and data science professionals exacerbates these technical challenges.
Any organization pursuing proprietary generative AI needs in-house ML experts to refine data management practices and build training pipelines for custom models. Post-deployment, ML operations (MLOps) skills are also required to monitor model performance, address data flaws and errors, and handle integration issues. However, due to intense corporate competition for the relatively small pool of ML talent, hiring these team members can itself pose a significant obstacle.
Enterprise generative AI is still in its early stages. To avoid costly mistakes, organizations need to cultivate patience, realism, and a long-term vision.
Building custom generative AI models is a complex and costly proposition that may not be the right choice for every business. When evaluating whether internal model development initiatives are worthwhile, companies should weigh the potential benefits of generative AI against the required resources.
"We're still at a fairly early stage in AI adoption," said a business manager in the transportation industry in response to an ESG survey. "While many people are talking about AI and recognize its potential, they... are reluctant to adopt [or] invest effort in training AI."
Respondents expect their organizations will need to invest broadly to support generative AI initiatives, including in training and employee skills (47%); information management (44%); and data privacy, compliance, and risk (37%).
However, for businesses that need and can invest in ML infrastructure and talent, building custom generative AI can provide a competitive advantage. Respondents anticipate wide-ranging benefits from generative AI in their organizations, including improved or automated processes and workflows (53%), enhanced data analysis and business intelligence (52%), and increased employee productivity (51%).
Comprehensive planning is crucial for the success of generative AI initiatives. As a first step, take time to carefully identify isolated areas where custom generative AI tools can deliver particularly high returns. Next, build prototypes and conduct extensive pilots before scaling to broad deployment.
In ESG's survey, top enterprise generative AI use cases included data insights, chatbots, employee productivity and tasks, and content creation. For example, an executive in the computer services industry reported their organization uses generative AI to create content ranging from social media marketing to technical ebooks to presentation slides.
Overall, the key is to start small with focused, goal-oriented models that can gradually expand in scope after proving their value. Don't expect to build a massive internal ChatGPT; fine-tuning models for specific tasks using internal datasets is faster, requires fewer resources, and is more likely to demonstrate short-term returns.