![Search icon](https://hackernoon.imgix.net/search-new.png?w=19&h=19)
The hype for Gen AI has settled as organizations have begun to lay the groundwork for seamless adoption. But there are decisions that must be made, dilemmas to be solved, and questions to be answered. We list the top 10 Enterprise Gen AI dilemmas facing the C-suite leaders.
The world came to know about Gen AI only through licensed models such as Open AI’s GPT or Google’s Bard. These systems shared very little details about the underlying code or how their systems were trained. But the Geminis and ChatGPTs indeed helped end-users understand Gen AI’s immense possibilities for the larger public.
But soon came the emergence of open-source models such as Meta’s Llama. These models are built on a shared architecture, reveal the underlying codebases, and are pretty much transparent about the data being trained. So which way Enterprises would sway when it comes to embracing these foundation models? Open-sourced or licensed - which would really benefit enterprises in the long run. Various aspects related to performance, accuracy, ethics, explainability, and IP protection would enable stakeholders to decide whether to go with closed or open-sourced models.
Large language models (LLMs) are known for their high performance, comprehensive understanding, and versatility. However, they are lavish energy spenders, which means only the largest companies have the funds to train and maintain energy-hungry models with billions of parameters. Could it be that these enterprises are better off adopting smaller language models that are comparatively energy-efficient, cost-effective, and even privacy-friendly?
Among these smaller models, domain-specific LM is fast getting traction because of the obvious advantages - you can run your small model on modest hardware yet build powerful customized models trained on proprietary data. Industries such as healthcare, legal, and finance could immensely benefit from these domain-specific smaller language models.
Have you ever asked ChatGPT as to how it arrived at a decision? While the outputs of ChatGPT can be evaluated for their relevance and accuracy, the inner workings of how specific responses were generated were never transparent. This could be a challenge for enterprises when they work with LLMs as these models given their vast architecture with several layers of neurons, make it hard to trace the decision path. For any input, the model may not be able to tell how it arrived at a particular output. This is what we call a Black Box AI.
Again on this front, smaller language models (SLMs) are better poised to allow human users to comprehend and trust the results created by their algorithms. This explain ability provided by smaller language models can help developers ensure that the system is working as expected. SLMs may go even further to help enterprises meet regulatory standards.
The enterprise layer is crucial to building successful Gen AI applications. Remember one can’t take the-box foundation models such as Open AI’s GPT or Anthopic’s Claude, and instantly create Gen AI applications for their enterprise. They have to manually build an enterprise orchestration layer that enables the infusion of Gen AI into their normal applications. Would organizations invest considerable resources into building this layer? This is where Enterprises must work with effective partners to create a customized business layer that truly connects enterprise applications with Gen AI models.
Enterprise leaders emphasize that Gen AI offerings must demonstrate a clear business value proposition. But for a technology that is in its nascent stages, yet with tremendous transformative potential - how can enterprises measure value? Should organizations overlook ROI in the initial stages and instead focus on how Gen AI infuses innovation and helps them in strategic positioning and differentiation? Or should enterprises be shrewd enough to prioritize investing in areas that can bring easy wins, and gradually get into complex projects?
The adage cannot be truer in the age of generative AI where the reliability of Gen AI output majorly depends on the quality of data being sent as input. How can organizations ensure the highest data quality standards when training Gen AI models? How seamless/difficult would it be for organizations to ensure that their custom AI abides and operates based on their internal data?
This situation is further compounded when we realize that on average, organizations only capture 56% of the data they create. Even more alluding is that about 77% of the data collected is either obsolete or trivial. So, Organizations, if at all, if they wish for seamless large-scale Gen AI deployments, must thoroughly capture, classify, and clean data before feeding into their AI systems.
On one side, enterprises are concerned about the accuracy of Gen AI outputs. On the other side, employees have an inherent fear of technology potentially replacing them or making their jobs obsolete. Together, this cumulative mistrust could pose a barrier to Gen AI enterprise adoption. But fortunately, these concerns have not stopped organizations from embracing Gen AI in their Pilots or small-scale experiments. But if these pilots are to be scaled, then organizations have to build trust in numerous dimensions across worker empathy, transparency, input/output quality, etc.
Can Chatbots go all the way to completing the task apart from making sensible conversations? The ChatGPTs and Geminis have shown us how their question-and-answer systems can give rich contextualized answers. But can the chatbot help us complete a task altogether? For example, can it make reservations for us? Can it plan a trip or even book a ticket? The dawn of sophisticated virtual agents is not far and that would transform customer experience in ways never seen before.
The ambiguity in the regulatory environment continues to prevail. However, the EU has taken a more conservative approach to AI-generated content. In Dec 2023, a provisional agreement on the Artificial Intelligence Act was reached, prohibiting indiscriminate scraping of images to create facial recognition databases, biometric categorization systems (with potential for discriminatory bias), "social scoring" systems, and the use of AI for social or economic manipulation.
Because of the impending election, there is still time for America to decide on AI governance and security, but that could only do little to stop the stellar developments continuing to happen in the country. Yet organizations cannot afford to be complacent when it comes to implementing guardrails for trust and security.
Shadow AI typically arises when a seemingly enthusiastic employee, in the desire to learn or complete an action through an AI application, feeds sensitive information to a public-facing model and thus exposes business secrets. In another case, the employee could train the model with copyrighted material, potentially exposing the company to legal consequences.
In conclusion, the potential benefits of Gen AI are immense, but realizing them requires a delicate balance of technological innovation, ethical considerations, and strategic foresight. As the technology matures, organizations must navigate these challenges proactively to harness Gen AI's power while mitigating risks. The road ahead is undoubtedly complex, but with careful planning and execution, enterprises can position themselves at the forefront of this transformative era.