Enterprises use several different technologies, including AI, to automate their processes and create value. They are currently experimenting with generative AI to determine whether it can provide significant and enduring value. Is the enterprise’s excitement about generative AI justified? Is the size of the investments being made and planned warranted? What is their strategy missing and should be in their action plan?
Over the years, enterprises have used AI to automate various processes. The goal of each such automation was to improve productivity, reduce costs, increase efficiency, enhance the customer experience, identify new revenue streams, and grow existing ones. For example, Robotic Process Automation has been used to automate several back-office processes to increase productivity while decreasing labor costs. Machine learning is being used extensively in online advertising to increase productivity, advertisement effectiveness, and decrease costs. A survey of business executives conducted by IDC during October 2023 across North America, Europe, and Asia revealed that increasing operational efficiency through the application of AI is the top goal. But does the application of generative AI have the potential to create additional value for the enterprise beyond what discriminative AI is already providing? The answer is yes but not without requiring further analysis. For example, it has already been reported that generative AI applied to software engineering leads to improvements in many areas including developer productivity. But it has also been reported that in some cases it is more cost-effective not to automate a task because implementing, operating, and maintaining an AI application, including one that utilizes generative AI, is more expensive (today) than the existing human-operated process.
Selecting the task to automate and derive the desired form of value should be the first priority. During this year when many enterprises are experimenting with generative AI the task selection should also be driven by the scope of the experimentation. For example, if the corporation applies generative AI to software development, a recommended early experiment is to use it to generate comments to legacy routines. Such an application improves the productivity of IT staff and reduces the cost of legacy system maintenance. We recommend that early process automation efforts focus on processes whose product has a language-like structure and where existing foundation models and LLMs can be successfully applied with minimal or no refinement, e.g., computer programming, or translating business manuals,
A successful generative AI application needs four ingredients to provide value: people, models, data, and infrastructure. The people involved must bring three capabilities. First, they must be able to ideate on how to reimagine and redesign the process or task to be automated. In this way, they will be able to innovate to achieve significant value, rather than merely achieving smaller efficiencies. It is unlikely that straight (step-by-step) automation of a process that was designed to be performed by humans will result in major value creation. Instead, the process will need to be redesigned with a view of taking advantage of AI’s strengths. Second, they must bring AI technology know-how. The field of generative AI is evolving rapidly. While having people with strong AI foundations is a good start, corporations must ensure that their AI experts are keeping up with the developments in generative AI. They will use their expertise to determine:
- whether to utilize an existing model or build a new one; if they decide to use existing models, they must select the right models (already many LLMs are available in addition to the foundation models),
- the appropriate approach to refine the selected LLM(s) to successfully address the automation’s requirements given the corporate and other available data, and
- how to implement and scale the application that automates the process.
Third, they must understand how to address the risks associated with generative AI applications. These risks range from incorrect output generated by the models used due to “hallucinations,” copyright infringement and other legal risks based on the provenance of the data used to develop or refine the model(s) incorporated in the application, privacy violations and/or biases due to the data used to train the model(s), the introduction of viruses that negatively impact the application’s security and the cybersecurity of the users’ systems, to properly scaling a pilot application to a broadly deployed system.
When ChatGPT first burst into the scene, enterprises felt that to succeed in generative AI they only needed to work with a single foundation model and its associated chatbot. Fast-forward to today enterprises must understand the strengths and weaknesses of several foundation models, some of which are multimodal, and a larger number of proprietary and open-source LLMs that is constantly growing to determine how to best take advantage of generative AI. A clear mapping between task and model doesn’t yet exist and is needed.
Diverse data types are important for refining/fine-tuning existing models, tailoring them to the enterprise’s needs, and incorporating information about products, services, clients, partners, etc., but also for building completely proprietary language models that address specific enterprises that general-purpose LLMs cannot address even after refinement. Cataloging the relevant data inside the corporation, determining what additional data to license, integrating, cleaning, and curating the data are important tasks that must be performed for value to be achieved.
To realize the various forms of value that are possible from the application of generative AI during each stage of the journey, we recommend that an enterprise takes the following actions before starting any project:
- Select a process with an impact on the enterprise, whose automation will be highly valued. Determine the form of value the project will generate at each stage and the use cases that will make it most impactful.
- Reimagine and redesign the selected process in a way that generative AI’s current capabilities can be used effectively to achieve the expected value. Align this strategy with the corporation’s overall digital strategy.
- Allocate the necessary budget to address the needs of each stage. Just because there are available free generative AI resources, it doesn’t mean that a corporation’s experimentation with the technology can be done for free.
- Assign an owner for the effort, making sure that this person will be available for the duration. In this way, the enterprise will benefit from the knowledge and experience gained throughout the effort.
- Develop a measurement framework for assessing the value achieved at every stage. Most organizations start a project without such a framework and then try to back-fit the obtained results, which often proves impossible.
- Keep up with and understand the generative AI technologies as they rapidly evolve and adapt the corporate efforts accordingly.
