Featured
Accelerating AI/generative AI with assetization
Workflow integration with Brix by QuantumBlack, AI by McKinsey
McKinsey research indicates that, in today’s fast-paced digital world, the most successful organizations are those leveraging artificial intelligence and machine learning (AI/ML) to restructure their operating models, focus on high-impact areas, and stay competitive.
It is known that projects using AI/ML are a key driver of business value. However, the complexity involved in building, deploying, and managing it at scale should not be underestimated.
Our recent book, Rewired: The McKinsey Guide to Outcompeting in the Age of Digital and AI, (Wiley, June 2023) explains the benefits of architecting reusable solutions as a set of modules, or assets. Keeping digital assets in a central repository makes them accessible across the organization. This also increases efficiency through accelerated development, and thus innovation. Conversely, when digital assets are spread across multiple platforms, discoverability, reuse, and governance are negatively affected.
Scaling AI effectively relies on assetizing code, insights, and resources to streamline operations and enhance productivity. For AI transformation projects, the benefits of assetization include faster time to market, lower costs, reduced risk, greater consistency, and easier adaptation.
Reusing assets is crucial for accelerating gen AI projects. The core types of sharable digital assets are code, data, and models:
Front-end code
- User interface components: Reusable UI elements that leverage AI capabilities, such as chatbot interfaces, or process mapping for generative AI (gen AI) agents.
- Visualization libraries: Tools and scripts for creating AI-driven data visualizations like interactive charts to highlight insights from machine learning models, or dashboards that display real-time predictions.
Back-end code
- Use case components: Modular, independently deployable services that handle specific AI functionalities, such as model inference, data preprocessing, and feature extraction.
- Full use cases: Components that can be used together to solve a complete business use case, such as customer life-time value prediction.
Data
- Curated datasets: Cleaned and structured data collections that are ready for analysis and model training. Examples include demographic data, transaction records, and sensor data.
Models
- Pre-trained models: AI models that have already been trained on large datasets and can be fine-tuned for specific applications. Examples include image recognition models, natural language processing models, and recommendation systems.
- Model templates: Standardized configuration template for developing new models, including baseline architectures and hyperparameters.
In the past, organizations have focused on sharing technical details about training models, pre-trained classical ML models, and reusable functional code. With gen AI, that focus is shifting to sharing prompts, pre-configured agents, and reusable code, with an increasing emphasis on front-end code such as chatbot UIs. For gen AI projects, instead of sharing how to train models, teams are now either sharing how to use them or sharing prompts that produce better and more creative results.
The challenges of assetization and reuse
Teams can often get stuck in a cycle of creating everything from scratch, instead of ‘standing on the shoulders of giants’ and reusing code from elsewhere in the business. Imagine if, instead of using the TensorFlow library, each team rewrote it from scratch. The pace of development would be astronomically slow; the quality may be questionable; and the level of risk for the newly developed library would be high.
Engineers use third-party libraries like TensorFlow or scikit-learn when developing new use cases or training models. However, they often don’t adopt this approach for their internally developed code and datasets, to the detriment of their development velocity. Instead of publishing and reusing best-in-class internal assets, such as data extraction pipelines, teams often fall into the trap of developing them from scratch. There are a several reasons why teams “reinvent” rather than reuse:
- Lack of asset discoverability due to fragmented sources: Digital assets are often stored across multiple platforms, such as code on GitHub, JFrog, or other artifact repositories, and data in systems like Snowflake, Databricks, or other application-specific databases. Models are stored in even more diverse sources, ranging from MLflow and cloud catalogs to local folders on individuals’ machines. This fragmentation reduces the ability for teams to find assets to reuse.
- Difficulty accessing assets: Permissions and security protocols can impede teams from reusing assets because of access restrictions across different platforms. Even within a single platform, there can be barriers to sharing between teams, which can become a logistical challenge.
- Challenges in using and integrating assets: Digital assets developed internally often aren’t developed with the intention of being reused. Consequently, they may have inconsistent documentation, varying standards, and a lack of user-friendly interfaces. Even if it’s possible for teams to find and access these assets, they can be difficult to integrate and reuse effectively.
The challenges above serve to reduce the time to value. The opportunity costs can set back organizations in their gen AI ambitions since expertise and time is diverted unnecessarily.
Brix: A centralized platform for sharing assets
In a recent blog post, we introduced Brix, which solves the challenges described above. Brix, from QuantumBlack Labs, is designed to centralize assets for discoverability and accessibility. In the rest of this article, we’ll illustrate how Brix integrates with existing tools, including code repositories (GitHub), data warehouses (Databricks, Snowflake), model catalogs (MLflow, Unity Catalog) and beyond.
Brix can transform time to value for AI projects, and reduce operating costs and inefficiencies of AI teams via:
- Centralized gen AI/AI assets: Brix provides a unified platform where all digital assets can be stored, discovered, and reused.
- Quality and reusability assurance: Brix has features to assess code quality, and to document and refactor code upon upload, ensuring quality and consistency across assets. The snowball effect means that, the more reuse an asset receives, the better the code and documentation becomes. An asset is refined over time through increasing numbers of reviews and improvement suggestions.
- Curated access: Brix enables easily controllable organization-wide access to a curated set of assets. The right people have access to the right assets at the right time.
The impact of Brix at McKinsey
At QuantumBlack, AI by McKinsey, we faced the challenge of data science component reuse a few years ago, as we grew in numbers. There was significant time pressure on our data practitioners, who found they were tackling problems solved elsewhere in the company. QuantumBlack Labs set out to build Brix, and since launching it, we’ve gathered some data on usage:
- Assets shared on Brix: ~700 unique assets; 3,000+ versions overall.
- Total files shared on Brix: >15,000
- Total lines of code shared on Brix: >10 Million.
A concept case study example
QuantumBlack Labs helped a leading global transportation business codevelop a customized version of Brix. The goal was to reduce the time to value of new gen AI/AI use cases from three months to 1 month.
The ‘digital marketplace’ was designed to centralize best-in-class code, curated data sets, and models, that could easily be taken ‘off the shelf’ by any colleague. New use cases were composed of 40–80 percent reused assets, rather than new code. This reduced time spent querying and cleaning data, writing code, and fine-tuning models.
Before the project, the business housed its data and models across local machines and on two separate platforms: Snowflake and Databricks. Each platform had distinct user groups, multiple layers of access, and different internal processes for access granting and management. Code was spread across numerous repositories, each with their own bespoke access restrictions. Not only were data and code siloed across platforms, but even within platforms and teams. This meant that data scientists, analysts, and engineers couldn’t find existing data and code developed by others in the organization. They were constantly reinventing their own solutions within their workspace and platform.
The new digital marketplace, based on Brix, was designed to be accessible to over 1000+ colleagues. It featured integration with the company’s existing information sources: Snowflake, Databricks (including Unity Catalog), MLflow, and GitHub. It formed the basis for the development of designs, requirements, and architectures to be executed by the product development team and achieve the organization’s goals of reducing the time to value by 60 percent or more, for new AI use cases.
What are the signs that your organization would benefit from Brix?
The following are clear signals:
- A critical mass of engineers that are building AI/ML projects, usually 50 or more.
- Data, code, and models spread across multiple systems
- Teams that are constantly reinventing solutions to problems others have solved such that there are low levels of data, code, or model reuse.
- New gen AI/AI use cases are taking three or more months to develop.
How to foster a culture of reuse
Reusing assets can be a powerful way to accelerate time-to-value, enhance solution quality, and mitigate risks in new GenAI projects. But how can organizations foster reuse at scale?
Incentives can be compelling. QuantumBlack’s experience shows that positive incentives (carrots) can be more effective than punitive measures (sticks). However, overcoming the ‘not invented here’ syndrome, where teams are reluctant to use externally developed assets, is a significant challenge.
When developing a culture that encourages reuse, it is important to recognize that it needs to cover both contribution and consumption of assets. Therefore, a dual-lens approach is required. These methodologies can be effective:
Align interests: Ensure that employee incentives are structured to promote asset contribution and asset reuse, and disincentives are eliminated. For example, incorporate requirements for reuse into employee evaluation frameworks, goal-setting processes, and product development criteria. A practical approach could be to mandate that new use cases consist of at least 20 percent reused inner-source components.
Celebrate and radiate: Transparently highlight individual contributions and the level of asset consumption (comments, shares, and downloads of individual assets). Greater recognition through gamified leadership boards, automated personal reports, and town hall announcements can also be used to celebrate contributors and consumers. This not only acknowledges individual efforts, but also encourages a broader culture of reuse.
Sustainable reuse at scale often cannot be ‘imposed’, but a ‘choose to use’ philosophy can make a systemic difference. This means not only providing the tools to reuse, but also making changes for people and processes to encourage reuse within the organization. Practical steps can include reducing barriers between teams by organizing regular inter-team meetings and creating shared communication channels. Other options include implementing monorepos to centralize codebases, and establishing cross-team working groups to focus on common goals.
In short, achieving sustainable reuse at scale requires more than just technical solutions. There needs to be a holistic approach involving people-related changes and process improvements. By aligning interests, celebrating contributions, and fostering a culture of reuse, organizations can effectively bring about a culture of reuse over time.
Summary
In an era where digital assets are pivotal to AI development and organizational success, Brix changes how teams operate. By integrating existing storage into a centralized, platform-agnostic solution, Brix helps organizations overcome the complexities of managing multiple sources of digital assets. The goal of Brix is to enable teams to focus on reuse rather than reinvention. With Brix, a business can streamline its AI workflows, boost efficiency, and ultimately achieve better business outcomes and build a competitive edge.
Discover how Brix can transform your organization’s digital asset management and reduce time to value for new AI use cases. Contact Marc_Solomon@mckinsey.com to learn more and schedule a demo.
Thanks to all who contributed to this article: Marc Solomon, Jo Stichbury, Yetunde Dada, Rory Walsh & Joanna Sych.
