Glasswing Ventures’ Enterprise AI Value Creation Framework

The high-growth AI market is expected to reach $2.7 trillion by 2032. This means that over the next eight years, one can anticipate a 20% annual increase in AI use cases, AI-native solutions, and AI services. With this rapid expansion comes the responsibility of founders, enterprise business leaders, and investors to recognize which AI use cases will bring the highest return on investment and the necessary criteria AI-native solutions must meet to support these use cases.

As an AI and frontier tech-originalist investor, Glasswing Ventures has rigorously and diligently evaluated countless AI technologies for over a decade. We draw on our technical and operational backgrounds in frontier tech to analyze how AI models are structured, the foundations and quality of the data on which they are trained, and the resulting potential for meaningful business impact.

Today, we are proud to unveil and open source Glasswing’s proprietary Enterprise AI Value Creation Framework. This framework represents the culmination of years of domain experience backed by technical and business expertise. It is designed for AI pioneers and enterprise adopters — builders and buyers alike — to utilize in their evaluation, selection, creation, deployment, and performance tracking of the impact of AI-native products on enterprise use cases. With this framework as a resource, founders, business leaders, and investors should feel empowered to discern between quick-win, lower-impact AI use cases and long-term AI initiatives with higher impact potential while also being cognizant of the necessary criteria to build AI-native solutions that effectively support their use cases.

The Three Criteria Necessary to Create Value with AI

Building AI products is not a monolithic process. Artificial Intelligence represents the outcome of a broad set of architectures, techniques, data sets, and training algorithms working together to produce the desired output for a particular use case. Glasswing’s Enterprise AI Adoption Framework deconstructs the three key criteria that determine the potential value of an AI-native platform to an enterprise business. These criteria are data, architecture, and impact.

How to Read the Enterprise AI Value Creation Framework

Depicted as a chart with an x- and y-axis, the Enterprise AI Value Creation Framework represents a comparative analytical model. The criteria that impact value creation are indicated on the horizontal axis (x-axis), while the potential resulting value of deployed AI tools and platforms is positioned on the vertical axis (y-axis).

We have designed the framework to represent a mathematical function encompassing aspects of utility theory, where the expected utility (positive business impact) is maximized under given constraints and inputs. This is illustrated as the function f(x), which is summed across all variables — Data Quality to Propensity to Adopt — to predict the value of AI solutions on identified use cases.

To provide examples of how to read the Framework, we have included two use cases: AI in compound manufacturing processes (represented by light blue) and AI in content marketing (represented by dark blue). The area under the use case’s line demonstrates the impact of AI’s access to various business areas on value creation and the associated risks.

Note: These examples are purely illustrative.

1: Data

While training data is the backbone of any successful AI model, simply “having data” will not guarantee an impactful outcome. Rather, the data must be clean, maintained, and relevant to the use case. Glasswing’s Framework outlines the elements one should look for in the data on which an AI model is trained to ensure the model leverages the right data in the right way for the appropriate use case.

Data Quality:

At the most basic level, the higher the quality of data an AI model is trained on, the higher the quality of output the model will produce. The data must also be robust enough to encapsulate all aspects of the business outcome that the AI model is being designed to deliver.

While every AI model benefits from high-quality data, some use cases require higher-quality data than others. For example, AI models used in compound manufacturing must be trained on exceptionally high-quality, thorough data due to the severe consequences that result from an error. AI models used in content marketing, on the other hand, can be trained on lower-quality data because the repercussions of an error are less impactful to the overall business.

In use cases involving lower-quality data, there must always be an option for a human to remain in the loop to review any output before it is published. This practice helps ensure accuracy and reliability, as humans can catch errors, biases, and inconsistencies that automated systems might overlook or misinterpret due to poor data quality or a lack of input data.

Data Availability:

When evaluating AI-native solutions, one should not only look at the quality of data on which the AI model is trained but also the quantity of data available for training. Data availability refers to the data to which an AI model has access. The availability of the data directly impacts the specificity, security, and accuracy of the model.

Private data refers to data that is unique to a business and difficult for the public to access, such as patient information. AI models used for compound manufacturing should be trained on vertically specific, private data to protect intellectual property, ensure model accuracy, and comply with the industry’s stringent regulatory standards.

Conversely, public data refers to data that is more widely available (horizontally available) and not specific to an individual business. AI-powered content marketing tools function well on public data due to the variety of answers one may look for when leveraging the product, as seen with ChatGPT.

Threshold of Adequate Performance:

The threshold of adequate performance for an AI solution refers to the benchmark an AI model must hit to be considered sufficiently operational. There are several questions that must be answered to determine such a threshold. For example, what are the consequences of errors made by the model? What does “state-of-the-art” look like for AI in this use case, and how easily can it be achieved? What is the balance between error rate and inclusion of humans in the loop?

Think of it this way: When one leverages an AI product to draft emails, they review (or should review) the email copy prior to hitting send, making sure the email is tonally, grammatically, and factually correct. If they do not review the output, they risk sending an email with hallucinations, which will hurt the reputation of their personal and professional brand. On the other hand, an AI malfunction within the workflow of pharmaceutical drug production will cost millions of dollars and threaten human lives. For this reason, the threshold for deeming an AI model effective and valuable to compound manufacturing must be much higher than the threshold for content marketing.

When exploring AI products for such important and sensitive use cases as compound manufacturing, one should look for AI models that incorporate features like feedback loops, duplicate subsystems or components of systems, and Machine Learning for predictive maintenance.

2: Architecture

Where AI is the “what,” its architecture is the “how.” The architecture of an AI application refers to the unique combination of models, guardrails, data, and systems that make it function. Understanding certain nuances of an AI model’s system, regulatory sensitivity, and model performance will help one understand the potential impact of an AI solution on its use case.

System:

AI systems can be broadly categorized as open-ended or discrete, though in practice, many AI solutions incorporate both types of code.

Open-ended AI systems continuously analyze and process data, incorporating new data as it is introduced. Open-ended systems are responsible for the continuous learning functionality of AI models, where the models consistently improve with the introduction of more data. Content marketing tools operate sufficiently on open-ended systems because of the way in which their models can adapt and adjust — such as AI writing tools adapting their outputs’ tone of voice based on inputted writing samples. Because open-ended systems are highly adaptable and do not rely on specific datasets, they are relatively simple to remove from the broader AI model.

Discrete AI systems, on the other hand, operate on distinct data sets. They follow a strict formula for computing inputted data, allowing for outputs such as pattern recognition. Discrete AI systems are essential to compound manufacturing, which relies on distinct and countable variables to produce the correct product. Because of their uniqueness, discrete AI systems are essential to the proper functioning of their AI models and are difficult to remove.

Regulatory Sensitivity:

As the phrase suggests, regulatory sensitivity refers to the extent to which an AI model complies with the existing regulations of its use case. While different regulations may apply to different use cases within one industry, such as those governing AI used on the manufacturing floor versus the AI used by the marketing team of a manufacturing company, all AI systems must operate within the legal frameworks of their industry.

For example, AI used to optimize production on a manufacturing floor must adhere to rigorous safety and operational standards, as any malfunction could directly impact product quality, worker safety, and compliance with industry certifications. Conversely, the AI used by the marketing team of that same manufacturing company operates under fewer constraints, as its outputs — such as campaign targeting or customer engagement — pose less risk to human safety or product integrity. At the same time, the more the content marketing solution is tailored to the unique aspects and regulations of the manufacturing industry, the more value it will create for the business.

Sensitivity to regulatory frameworks is not just a compliance requirement; it’s a competitive advantage. Companies that design AI systems with regulatory foresight not only mitigate risks but also build trust and credibility in markets that are increasingly attuned to ethical and lawful AI deployment.

Model Performance:

The performance of an AI model should be measured in relation to the business metrics of its use case. This could be a measure of accuracy  — how close the model got to predicting the correct value, or it could be a measure of precision and recall, i.e., how often the model correctly detected an interesting event versus how often it provided a “false positive.” Quantitative metrics are a cornerstone of evaluating AI performance.

In high-stakes environments like compound manufacturing, achieving high performance is a mandate. AI systems predicting chemical formulations, monitoring quality, or ensuring consistency in production must adhere to extremely high standards. A single false positive — such as approving a flawed product — can have cascading consequences, from operational inefficiencies to compromised safety. Here, performance is synonymous with trust, making near-perfect accuracy non-negotiable.

On the other hand, accuracy benchmarks in lower-stakes use cases such as content marketing are a bit more flexible. AI models generating ad copy, personalized recommendations, or campaign strategies don’t face life-or-death stakes. Rather, these models focus on creativity, engagement, and relevance. A slightly off-target suggestion or a less-than-perfect phrase might not be the ideal output, but it’s unlikely to cause significant harm. Precision in marketing supports optimization rather than absolutes.

3: Impact

While understanding the data and architecture of AI models will provide one with the technological know-how to build or buy an effective AI solution, recognizing which solutions are easy to integrate, are likely to be adopted, and have the potential for revenue impact is essential to ensuring one allocates resources to the AI initiative that brings the most value to their business.

Workflow Integration:

Adopting any AI technology disrupts the established workflow of a business function. Some AI solutions replace existing tools in one’s tech stack, in which case careful planning is required to not only decommission the old system but also to ensure that the new AI solution seamlessly communicates with the technologies already in place. Such planning may involve significant backend integration efforts as well as user training to close any knowledge or functionality gaps.

Alternatively, the AI solution may supplant manual tasks previously handled by employees, in which case a strategic realignment of roles is required. Decision-makers must redefine employees’ responsibilities to leverage their human skills in areas where AI cannot efficiently perform.

The ease of AI integration varies significantly by use case. AI solutions in content marketing have gained rapid traction due to their relatively straightforward integration into existing workflows. These systems are often user-friendly, require minimal adjustments to current processes, and can dramatically enhance efficiency and creativity in content generation. AI implementations in industries like compound manufacturing, on the other hand, are more burdensome, as they demand extensive planning, a deep understanding of existing processes, and significant resources to be integrated effectively.

Revenue Impact:

The impact an AI platform has on a business’s revenue can be either direct or indirect. When an AI model has direct revenue impact, its impact is concrete and measurable, often reflected in metrics such as increased customer acquisition, heightened product sales, reduced customer churn, or significantly reduced cost. AI models providing peripheral revenue impact will not have immediately measurable results in financial statements but are crucial for long-term growth.

In specialized fields like compound manufacturing, AI-native solutions often directly impact revenue by optimizing production processes and reducing the frequency of costly errors. Glasswing’s portfolio company, Basetwo, for example, has built a platform that helps its customers see a 20% reduction in process deviations, thus increasing the amount of viable product produced from each pharmaceutical batch and measurably creating a more cost-efficient manufacturing process.

Conversely, AI applications in content marketing typically deliver a peripheral impact on revenue. These tools refine content strategies and enhance user engagement, gradually helping companies build a stronger brand presence. Although these benefits might not immediately and directly impact revenue figures, they lay the groundwork for sustained growth by improving productivity, fostering a loyal customer base, and strengthening one’s market position.

Propensity to Adopt:

The likelihood of enterprise teams to adopt an AI platform directly ties to the two previously described impact criteria. Generally speaking, if employees can clearly understand how an AI solution will improve their workflows and the overall business, they will be more willing to accept change. For example, despite the significant effort required to integrate AI systems into compound manufacturing processes, their potential to deliver transformative business impact is high enough for employees to engage with the technology and justify any initial challenges.

While AI solutions for content marketing are simpler to deploy and weave into existing workflows, the weaker ties to revenue lead to lower motivation among the broader enterprise to fully and completely adopt these products. With such a use case, it is evident that the ease of integration alone does not guarantee widespread acceptance of AI technology unless the impact on key performance indicators is evident and aligned with strategic business goals.

The Glasswing Perspective

The current AI landscape is vast, and it will only continue to expand in the coming years. Founders, business leaders, and investors who take the time now to understand the importance of data quality and accessibility, threshold of adequate performance, AI systems, sensitivities and performance, workflow integration, propensity to adopt, and the potential revenue impact of AI solutions on varying use cases will be the ones who cut through the noise and lead this next generation of technology adoption.

By following Glasswing’s Enterprise AI Adoption Framework and understanding the data, architecture, and impact of AI use cases and solutions, AI pioneers should feel empowered to make AI investment decisions that lead to positive business outcomes. If you have any questions or comments about the Framework, please do not hesitate to reach out to info@glasswing.vc.