Experimentation is not innovation
According to most industry experts, generative AI’s potential is near limitless, and the simple implementation of LLMs like OpenAI’s into existing workflows will be able to radically transform major enterprises.
When looking from this perspective, executives are presented with a paradox. There is an incredible first mover advantage for those who quickly grasp AI’s potential to create business-wide efficiency gains. However, the scope of this potential transformation is so great that executives need to fully consider how AI can best improve the millions of tasks and processes within their business.
As a result of this scale, enterprises get stuck in cycles of brainstorming, proof-of-concept creation, and pilot stage trials. Rarely, if ever, are new systems properly deployed into real production environments where actual users, systems, and APIs can interact with the AI features and functionality that have been developed by foundational companies to the tune of tens of billions of dollars. Instead, this endless experimentation fuels the pursuit of achieving an idealized best-possible outcome. Perfection is the enemy of measurable progress.
Experimentation without implementation is not innovation.
Experimentation without clear paths of deployment into production systems is crippling AI’s adoption amongst enterprises. Without real-world feedback, it’s impossible for enterprises to generate the data needed to evaluate outcomes, iterate effectively, and understand ROI.
So, if endless experimentation isn’t correct, how can enterprises actually begin to implement AI tools? Before we can understand how the implementation of AI solutions into existing processes can provide practical, measurable benefits to enterprises, we need to have a strong idea of the two fundamental superpowers that genAI systems possess on an enterprise level.
GenAI’s two superpowers
Firstly, AI systems can intelligently process data at a speed and scale no human-led process ever could. Enterprises gather millions of data points every day that can be fed into AI models to generate hyper-specific results, like knowing on which day a specific type of shirt sells better in a selected city. This data analysis is impressive and no doubt provides strong value, but it’s still within the bounds of current practice.
AI’s true power comes from its ability to integrate billions of disparate datapoints (captured in both internal and external datasets) into one model. Sticking with the shirt analogy, enterprises could leverage AI’s data processing capabilities to conduct a range of experiments and determine variables as specific as where in the store a specific shirt sells best, or if it sells better on rainy or dry weekends.
Better data means better outcomes for both businesses and consumers.
AI agents fed with hyper-personalized data can make suggestions – and even execute purchases! – to users that match their exact needs, consider regional nuances, and fit their pricing preferences.
While many have long expressed fears about AI’s ability to replace human labor, AI is directionless without a human at the helm. I firmly believe – no, I have experienced – that AI’s second superpower is its ability to accelerate and improve human capability, thinking, and creativity, not to replace it.
Most enterprises are fundamentally driven by a complex web of structured processes that ensure the consistency, repeatability, and reliability of delivering products and services to others. These processes are vital to ensuring a business’s operations are streamlined and that any new variables can quickly be accounted for.
Within these processes, there are routine tasks that are ripe for AI augmentation. For example, a worker building internal training schemes could use an AI tool to brainstorm potential answer sets for multiple-choice questions that look correct but are actually incorrect. Here, the AI’s creative capabilities based on its training data make the overall process more efficient, while also ensuring that the human actor continues to serve as the subject matter expert responsible for establishing the correct answers.
While not all tasks within a process will benefit from integrating AI (based on experience, about 60 to 75% of them probably won’t), its ability to accelerate workflows and provide better outcomes resembles magic to the untrained eye. To stay competitive, enterprises must develop a reliable process for moving ideas from experimentation to implementation.
How enterprises can quickly ideate & implement AI solutions
Executive sponsorship is essential to any project’s long-term success. With a senior leader personally invested in overseeing the implementation of AI systems, any potential external or internal hurdles can quickly be overcome. Senior oversight also provides the best vantage point from which to conduct the first step: a business impact ranking that identifies the highest-impact processes in an enterprise’s workflow.
If they are to have the most impact, AI initiatives must target core systems and value drivers, not peripheral functions. If the initiative targets non-core areas, it risks becoming a low-visibility hobby project that makes little impact and may reaffirm the mindsets of those skeptical about the potential impact of AI. The key is to start from the organizational core and consider how AI could optimize tasks within those workflows. For a payments-focused fintech, this would be their Payment Management System – for a hospital, it would be their internal Electronic Medical Records Management System.
With this foundation, AI teams can then assess the technical complexity of those processes and ascertain how realistically AI can be integrated. While simpler processes are easier to augment, they may not be the most impactful. To have their best shot at success, executives need to aim for the sweet spot between impact and ease.
This whole process should take no more than four weeks – any longer and an enterprise risks paralysis by analysis. Setting a clear timeline from ideation to deployment can help businesses avoid falling into the trap of indefinite brainstorming or design perfection. Enterprises must remember that experimentation without implementation is not innovation.
Multiple proof-of-concepts should be launched and run in parallel. Those that fail will provide invaluable learnings for future iterations. Once a viable concept has been refined, it should move quickly through the engineering – Q&A – user acceptance – deployment pipeline. With this framework, useful AI tools can be ideated, refined, and implemented within 3 months.
Building an AI-ready workplace
This three-month timeline works for every enterprise, regardless of scale, or if they’re a legacy provider with established systems or a newly emergent fintech. As with any process, however, there are ways in which it can be enhanced.
For executives, the focus of AI augmentation should be on routine tasks within the complex web of structured processes. The more easily a specific task can be targeted, the more easily AI tools can be introduced – so the extent of an enterprise’s success with AI is in part dependent on its internal architecture.
More than just software and systems, executives should strive to make every aspect of their business (and the resulting workflows) as modular as possible. The increased flexibility of modular systems enables enterprises to optimize their processes much more efficiently. Instead of needing to untangle a complex web of processes, these well-segmented modular systems enable enterprises to easily target individual tasks within workflows for improvement. This agile structure also enables enterprises to more easily manage workloads as needed to account for the efficiencies unlocked by implementing AI tools. This flexibility also means enterprises can easily use multiple tools in parallel and mix and match models, making comparative testing easier, and leaving space for new products to be integrated. Combined, this adaptability provides enterprises with a significant edge in finding the solutions that are best suited to their business.
Three factors to consider
In probabilistic AI systems, some degree of imperfection is inevitable. Executives are aware of this (the many hallucinations of popular consumer LLMs are well-storied), and yet launches are often endlessly delayed in the pursuit of an impossible perfection. Enterprises should set a fixed release date, launch boldly, and iterate based on live feedback.
For a nascent market to grow, a wide variety of tools need to launch, compete, and die, leaving only a few well-refined products remaining. This survival of the fittest can bring uncertainty to businesses that don’t want to build a system with a specific tool, only to have it acquired six months later and shut down. To avoid this, executives should favor tools with clear long-term viability and strong third-party backing.
Conversely, executives should avoid being locked into one vendor’s closed ecosystem. This would be an unnecessary hurdle. Many AI tools and models provide few functional differences. Any serious enterprise should always seek to avoid dependencies on any single external factor – be that an AI model, cloud computing partner, or data provider.
Experimentation and implementation
The implementation of AI tools does not need to require years of research, or cost millions. A streamlined process focused on identifying the smallest possible AI feature that could make a notable business impact – and the quickest possible path from ideation to deployment – can efficiently and affordably unlock real value from AI tools. Enterprises must remember – experimentation without implementation is not innovation.
