To assess a technology trend, follow the money. According to the Artificial Intelligence Index Report 2024 from the Institute for Human-Centered AI (HAI), Stanford University, investments in Generative AI have ‘skyrocketed’ from $2.82 billion in 2022 to $25.2 billion in 2023. The annual report has been among the most reliable sources of AI-related information. Our key takeaway from the report: Generative AI has gone from being cool to super hot in 12 months; enterprises of all shapes and sizes should get their act together around the technology if they have not done so already.
IDC has predicted that over 40 percent of core IT spending will be allocated to AI-related initiatives by 2025. Enterprise spending on generative AI services, software, and infrastructure is forecasted to grow from $16 billion in 2023 to $143 billion in 2027, with a CAGR of 73.3%. Gartner says that by 2027, more than 50 percent of the GenAI models that enterprises use will be specific to either an industry or business function—up from approximately 1 percent in 2023.
Where are we today? A study by O’Reilly Media, Generative AI in the Enterprise, released in late 2023, found that only 33 percent of the enterprises polled were not using Generative AI. Two-thirds of the respondents (67 percent) used it, of which 41 percent had used it for over a year. Delaying the adoption of Generative AI into the enterprise will not help.
The challenge is to figure out what makes a successful Generative AI implementation. Even enterprises that have begun their Generative AI journey may want to peruse the answer.
5 Criteria for Successful GenAI Deployment
Based on our market experience and lessons learned, enterprises building a generative AI solution from the ground up should focus on five critical dimensions central to a successful AI deployment:
- Quality of Training Data:
A well designed generative AI strategy will prioritize the quality, accuracy, variety, and size of the datasets used to train the selected foundation LLMs. The data must reflect diverse business-related scenarios that let the model perform precisely and reliably.
The data must be sourced, curated, labelled, and annotated carefully. All four tasks must ideally be performed by domain specialists familiar with the business who can ensure the accuracy, coverage, and completeness of the dataset to meet the use cases identified by the enterprise. In addition, they should take steps to address gaps in the training data, class imbalances, and other biases before the data is applied to the model’s neural network.
The source of the training data is critical. Data acquired from unverified public sources such as social media, blogs, and websites presents the risk of being flawed, false, and incomplete, resulting in factually wrong or biased output. Any model that produces an output with restricted truths or reinforces skewed truths will be unusable. The data could also include copyrighted material and PII, which can lead to legal consequences and raise serious ethical questions. - Selecting a Foundation Model:
Generative AI is moving fast. Every day, new applications and models appear, providing more choices and GTM options for the enterprise. How does one evaluate the pros and cons and select the right model for the business?
This is where an AI solution architect working with industry experts should be leveraged to identify and evaluate the right foundation model for the use case/s. The AI expert creates a baseline, grounds the model in the training data, and checks the output against the benchmark. Improvements in the dataset can help enhance the output. By comparing the output of various models, it is possible to arrive at a short list to pick from.
Model scoring for accuracy, relevance, validity, fluency, risk, safety, bilingual evaluation understudy (BLEU), Turing-style tests, etc., are then recommended to arrive at a final selection of the foundation model. Your generative AI strategy must also consider critical aspects such as scalability, performance, and commercials. - Customer Experience Design:
Any new technology introduces a learning curve for the user. An intuitive CX that works well and handles issues gracefully is essential. The answers to the following questions are crucial for shaping a generative AI strategy that simplifies user adoption and accelerates the business impact of the technology.
How can the customer experience be humanized? Do the responses from the model consider the user’s desire for simplicity, complexity, and flexibility? Is the model context-aware? Does it provide personalized outputs? Does it consider previous interactions with the user? Is it well integrated with the larger business or the industry environment to provide a deeper and more meaningful interaction to the user? Is the output unusable, and in these instances, can the user provide feedback that realigns the model?
- Continuous Model Training and Enhancement:
Continuous improvements are a vital aspect of any generative AI strategy. Systems and processes are required to ensure the model is continuously trained on the latest data and the application is refined for new or changing use cases.
The ideal approach is to design a CX process that tracks user behavior and captures user opinion and active feedback. Real-time data analysis allows a business to evaluate the value delivered to the user. The analysis is used to train and improve the model. MLOps becomes central to this process with automated data collection, processing, and mapping that result in new training datasets.
In parallel, CI/CD and DevOps allow for the continuous delivery of new features and products that is key to Gen AI App Management. This strategy, for continuous model training and App enhancement, should focus on reducing the cost of resources (human and computational), minimizing disruption, and reducing the need for coordination between the AI team and the domain experts. - Scaling and Performance:
Cost is a significant factor in ensuring a Generative AI model can scale and its performance remains reliable. Recurring costs include license fees for the underlying pre-trained LLM, Cloud Infrastructure charges, and data processing costs. While GPU and Cloud storage for data are well understood, network/bandwidth costs will add up and be a significant factor as the GenAI models are continuously refined. Collation and continuous processing of large volumes of data need to be accounted for in your generative AI strategy as key cost components.
GenAI Application development and integration with business platforms are one-time costs. The paucity of skilled resources contributes to higher initial costs of solution development.
Achieving scale and reducing costs requires a multi-pronged approach: Should the workloads be in a public cloud, on-premise, or a combination? When should quantization be used to minimize memory and computational requirements? These (and many similar) decisions around processes and tools can result in faster inference, better scalability, and lower cost.
Extracting value from every dollar spent on Generative AI is complex but can be achieved with the right generative AI strategy and technological knowledge.
Also read: Supercharge Insurance Underwriting with AI and GenAI
Embrace the Fail-Fast Approach
There is an urgency to get Generative AI into the enterprise workflow. Business efficiency and productivity can be elevated by several factors at the hands of AI. Making early mistakes is better than procrastinating over Generative AI adoption. As Nassim Nicholas Taleb, the author of The Black Swan and Antifragile, once noted, “Early failure is tuition. Late failure is punishment”. However, as always, creating a roadmap in collaboration with a competent technology partner can save money and person-hours and help avoid tuition and punishment. A well-thought-out generative AI strategy can turn early missteps into valuable learning experiences and expedite business outcomes.
Talk to us to create a roadmap for Gen AI adoption
References
https://aiindex.stanford.edu/wp-content/uploads/2024/04/HAI_AI-Index-Report-2024.pdf
https://www.gartner.com/en/articles/3-bold-and-actionable-predictions-for-the-future-of-genai
