Adapting your business to AI by creating AI assets is all the rage right now. At this point, you’ve heard it said a million times on social media, in keynotes at conferences, in strategic roadmaps: “Adopt now or be left behind.” Everybody seems to feel the gravity of AI and that it’s primed to upend business. But frequently, the understanding of specifically how it will do so, as well as which data or AI assets have lasting value, doesn’t go much further than that.
Now that we have AI that speaks English fluently, can it do any task we currently do in English? It’s easier than ever to imagine your daily tasks being done automatically across the various apps open on your computer. As most people interested in AI have come to realize, though, we’re not quite at that point yet, despite the seemingly infinite potential that we all felt exposed to the first time we used ChatGPT.
We’ve broken through to an incredible new plane where computers can operate and understand numbers, text, images, and audio. But AI doesn’t yet know how to do the things humans do. There are many high-minded and fundamentally human things that we rightly believe computers cannot and should not do. However, routine and repeated day-to-day tasks have suddenly fallen into the domain of machine-learning-based AI systems.
In this article, we’ll be looking into 5 ways that businesses can create AI assets to stay ahead in the AI revolution.
Use the links below to jump to each section:
As we move into this new technical paradigm, AI systems and the ways that their processes are architected will continue to change at a constantly accelerating speed. An ever-growing collection of researchers and engineers will seek new model architectures, training methods, and data modes to create more capable models.
All but the largest technology-focused companies will adapt the state-of-the-art AI models trained by advanced research teams to their use cases rather than building their own from scratch. However, the playing field will constantly shift as the capabilities of these systems grow in different and hard-to-predict directions. So, the question becomes: How should businesses play on it in the meantime?
To answer this question, a better set of questions would be: What assets do companies possess now that drive their success? What assets and strategies can they create moving forward that will retain their value even as AI disrupts knowledge work?
Many of the best minds in the world are working on the above question now within their respective areas of expertise. Stepping back and taking a macro view of the components of today’s AI systems, it’s possible to extrapolate a set of general language-based assets that researchers and engineers use as their starting points for the systems they create.
In essence, these are text-based data assets, and just as humans have much to learn from reading high-quality sources, the best AI will be the AI that has access to the best information, even as the ways it learns from this information continue to advance.
[quote_widget]
As there are countless niches that businesses successfully operate in, there are also countless forms of valuable AI assets that can be created.
[/quote_widget]
However, rather than continuing to talk in abstract terms, we’ll zoom in on a non-exhaustive list of five specific forms data assets might take. Here, we’ll focus on use cases centered around language models, as these are likely the most familiar and immediately applicable AI implementations for most readers today. However, for each of them, it only takes a little imagination to picture how these assets could be created using other data modes like tables.
In each of the five, we’ll provide business and technical descriptions of these to bridge the gap between the technology and what it’s useful for.
[first_letter]
Firstly, there’s the most well-understood use case of AI today, thanks to the meteoric rise of ChatGPT. This is AI’s ability to operationalize or extract information from text data you already own.
[/first_letter]
Obviously, many different types of value can exist in text. It can take many forms; maybe the text needs to be transformed into a different format or translated for a diverse user base. Users might need it shortened via summarization or extended to include explanations or definitions. Maybe sentiment or details about a particular object can be extracted. If your company has any text data that you interact with regularly, it’s worth asking if there’s a systematic use case to operationalize this text.
In technical terms this is called prompting a large language model (LLM) – think GPT-4, the model that answers questions when you use ChatGPT. Prompting is a method by which you explain the task that you want an LLM to do instead of providing it with explicit examples of the task being done.
You can prototype these methods by giving ChatGPT prompts like, “Translate the following technical document into business language and provide examples of how it might apply to XYZ,” followed by the text you have.
This class of AI asset requires the least work or attention. If the LLM can perform the task being asked of it directly, the only work left is to build a software pipeline that does this repeatedly.
Knowing what LLMs can do out-of-the-box with the text you have will continue to be a separating factor for businesses embracing AI. But keep in mind that if a model can do the task you ask of it with no modification besides a well-written set of instructions, then it can likely do the same for your competitors.
With this in mind, it’s best to see the transformed output text and a company’s knowledge of how to extract it from its existing text data as the AI asset. Moving forward to the following types of AI assets, this won’t be the case. Instead, there are ways for companies to create concrete datasets and implementations that enable them to scale their AI capabilities beyond what publicly accessible models are already capable of.
[first_letter]
While there are a select few industries that are panicking today, for the most part there is an understanding that today’s AI by itself cannot replace the capabilities of most businesses. That’s at least partly because these businesses possess specialized domain-specific know-how and have created practices, policies, and procedures around these.
[/first_letter]
Just as importantly, businesses have access to markets, skills, and information that aren’t publicly available for AI models to be trained on. This data is internal to a company and is rightfully held close to the chest in most cases. These are already critical intellectual assets for the company, but they can also become AI assets that their owners can scale significantly with the right approach.
The more of this know-how that can be carefully centralized into knowledge bases and fed to secure and proprietary models, the more this information can be quickly accessible to team members and, eventually, to AI that can act in line with it.
The most relevant technical implementation of knowledge bases today is Retrieval Augmented Generation (RAG) with LLMs. RAG is a method in which LLMs generate text by:
In this way, large amounts of information or instructions can be given to a model, exponentially increasing its usefulness within a given business domain.
[first_letter]
The next two classes of AI assets are more involved and likely require targeted development efforts to create. The payoff is potentially higher, though, as these allow active involvement in the model creation process.
[/first_letter]
Data that shows what a useful AI looks like by providing specific examples is incredibly powerful in the hands of AI engineers. With enough such data, they can train general-purpose models to pick up and mimic the behaviors that are shown in the data.
For an LLM, if you have thorough examples of a task being done well, you can teach an AI with these. In practice, a wide range of texts have the potential to be such an asset: from messaging conversations, to code, to desired document formats, to large collections of documents about a niche domain of expertise. If an associate-level employee could look at these previous texts and create a new one with minimal supervision, an LLM probably can too.
Using such a dataset to teach a model is called fine-tuning. To do it, engineers take a general-purpose language model like Meta’s Llama 3.1, which has been trained on vast amounts of data and which therefore already possesses strong capabilities in language understanding, and teach it how to approach a specific domain or task.
Just like a good teacher in a classroom lets students try the problems themselves and then corrects their answers, so too does the LLM training process ask models to give their best attempt at a problem, then corrects them so that the next time they respond closer to the “correct” answer you specified in the fine-tuning dataset.
The amount of data needed to get good results scales in proportion to how different the task is from what the model was originally capable of doing. The output of this process is a unique and proprietary model that can take inputs and create outputs in the specific way that you’ve taught it to, and the ability to create more of these every time a new and smarter open-source model is released.
[first_letter]
An essential step in creating and utilizing AI for high-value applications is to generate insights into their performance which can be constantly monitored. This makes it much easier for business users to keep an eye on models and ensure they’re aligned. It’s also essential for the AI to be able to evaluate itself.
[/first_letter]
Therefore, the real trick of this is to identify, as neatly and holistically as possible, the underlying requirements for the AI system being deployed. This might be user engagement, reported satisfaction, safety, compliance with policy, functionality with downstream systems, or countless other metrics.
In technical terms, these metric-gathering processes are referred to as model evaluations (or evals for short). In contrast to other data assets that we’ve defined so far, evals are not text-based assets but rather technical processes by which business value is defined quantitatively.
As such, they require businesses to get involved in the analytical process of translating qualitative information into quantitative. For some, this might be a familiar process of defining KPIs for internal use; such processes can be adapted from report-building to LLM evaluation. For others, it might take a process of reflecting on what measurable subgoals are involved in their formula of success and then translating those into measurement systems. Once these have been defined, the upward potential is immense!
Having good metrics provides the freedom to confidently swap the newest models in and out of your AI systems, or to quickly discern whether the other data assets mentioned above are working effectively to achieve desired outcomes and avoid potential pitfalls.
[first_letter]
The final data asset we’ll look at is more analogous to previous technical shifts like the introduction of app stores for mobile devices. In this framing, LLMs can be understood as a new type of platform in which users take action in digital space by interacting with AI assistants, which then carry out their directions.
[/first_letter]
The fast-growing capabilities of today’s language models have essentially guaranteed that the next generation of ChatGPT, Siri, Google Assistant, Alexa, and many hopeful new entrants will be far more capable of taking on complex responsibilities for their users.
This presents a new opportunity for businesses to step into what will be a growing ecosystem of services on these platforms by adapting their offerings to give these assistants access to them. B2C companies like travel agencies, tutors, and coders have already built such implementations to integrate their services with AI assistants on OpenAI’s GPT Store, the first such marketplace.
The technical terms for such integrations today are tools or plugins. OpenAI has also introduced a similar implementation called Actions.
While the names and format by which these integrations are built are liable to change over the coming months and years, what’s likely to stay constant is that third-party businesses can build these by providing their products to the AI assistants via APIs, which take a set of inputs that the model can deliver to them and return a set of outputs in response to that request. Alongside this, the assistants require a set of written instructions on how to use the given API. If done right, any model can understand the process and translate between the user’s needs and the company’s offerings.
As users shift their digital behavior to utilize these assistants more and more, companies that build offerings around this new paradigm will be poised to bring their services and expertise to this growing demographic.
Language-capable AI is heralding a new era of technical disruption, and the business environment that will develop in the new playing field has yet to be seen. However, what we do know today is that by using a set of inputs and frameworks, first movers can control and align AI with their individual goals.
[quote_widget]
Business leaders who understand the value of these AI assets, and find ways to map their domain expertise to them, are poised to become key players in this new field.
[/quote_widget]
Not only will such people create business value for their own companies, but they’ll also contribute to the way progress is shaped in this nascent field, which is always looking for new examples of how to successfully turn the vast possibilities of AI into functional AI systems.
In this new field, the AI/ML team at RapidScale is a group of AI-focused engineers, scientists, and solution architects who partner with businesses to turn their data into AI assets, and their AI assets into fully-captured business value.
We engage with customers on LLM-based and general machine learning solutions at all steps of the journey: from early-stage consultation and strategy development, to validating and building solutions, to optimizing AI/ML operations to ensure efficiency, safety, accessibility, and performance at scale.
[first_letter]
We’re constantly looking for win-win engagements where our technical expertise can pair with successful businesses to create net positive outcomes for all. To set up an introduction and discuss your business use case for creating AI assets, send a message to our team now.
[/first_letter]