An Intelligent Strategy for AI Architecture

Artificial intelligence (AI) has a rich history that began sometime in the 1950’s with the theoretical concepts of machine learning. The idea was that computers could do more than just what they were statically instructed to do, that they could actually be programmed to learn from the data fed into them. From this conceptual foundation the focus shifted in the 1980’s to expert systems which did not yet have the capacity to learn but could use rules applied to data in order to make decisions. The field progressed from expert systems to neural networks which attempted to mimic how interconnected neurons functioned in the human brain. In the 1990’s more advanced learning algorithms like support vector machines (SVM’s) began to take off and then in the 2000’s this led to deep learning which leveraged multi-layered neural networks. These advances enabled break-throughs in natural language processing (NLP), speech recognition and image recognition.? In the 2010’s we saw the advent of reinforcement learning where AI models could improve in real-time as they operated via trial and error. The application of these technologies culminated in the development of large language models (LLM’s) such as OpenAI and GPT-3 which could generate human-like text by predicting the likelihood of the next word or words given the previous words used. This may not sound like much but it is enormously powerful. The evolving capabilities of GenAI have enabled a whole new generation of intelligent applications that can answer complex questions, analyze data, read documents, transcribe meeting notes, write programming code and generate brand new content from existing text, images, audio and video. The question for us as technologists is how do we use these new capabilities safely and effectively? While architecture doesn’t have all the answers it can help provide a framework along with some of the guardrails for organizations to successfully apply GenAI to create business value.

In architecture I always like to start with grounding principles. As it turns out GenAI can actually help us with this so I asked ChatGPT what it thought. My prompt (prompt engineering is a critical new skill to learn) was “As an enterprise architect what are some of foundational principles you should apply to the usage of GenAI?”.? The model came back with a lengthy response that was actually pretty good. I integrated some of its “thinking” into my own which is grounded in the three pillars of my architecture practice: “Outcome Driven”, “Capability First” and “Platform Centric”. Here is what we came up with:

1. Business Alignment
2. Data Governance
3. Scalability and flexibility
4. System Interoperability
5. Risk Management
6. User Experience
7. Ethical usage
8. Continuous Improvement

That is good list so let’s take these one at a time and think about each of them.

Business Alignment

Like anything else in technology our attention should first be on the business value of what we do (i.e. Outcome driven). There is so much going on in the GenAI space that it is very easy to get distracted and fragmented to the point that no real progress is made on anything (anti pattern of chasing shiny objects). In order to avoid this it is important to start by focusing on a few key business use cases with well defined problems to solve. In this way organizations can realize value from GenAI while they learn and mature the technology. A light weight review process setup with key organizational stakeholders can be used to vet subsequent use cases as they arise.

Data Governance

Garbage in, garbage out as they say. Good results from GenAI models are dependent on their algorithms, their training and the corpus of data that they ingest. Poor data will generate inaccurate, incomplete or wholly incorrect responses (often called hallucinations). This makes data quality engineering a critical capability (capability first) to have in place within an overall data ownership and data governance framework for the enterprise. Additionally, data security, privacy and compliance concerns must also be addressed. This is not just with private models within an organization but with 3rd party models embedded in external platforms (e.g. SaaS). The classification (e.g. public, confidential, highly confidential etc.) of data ingested by models must be understood as well as how that data is used, if it will be retained (i.e. not transient to a request) and how it will be protected. These criteria should be evaluated as part of a GenAI governance process along with the proposed use cases.

Scalability and Flexibility

Both scalability and flexibility are solid architectural principles that should be applied to any system architecture and design. In the context of GenAI these can be best realized with a private platform approach (i.e. platform centric). An AI platform would include components such as a UI Layer, an API layer, a model integration layer and a model customization (or contextualization) layer. This architecture enables the usage of multiple models (public, private, vendor etc.), the ability to secure the software environment, control over the rollout of new GenAI capabilities, observability of models and the means to protect enterprise data. A private platform also allows for managed inner source development at scale to drive innovation.?

System Interoperability

The GenAI ecosystem is broad and rapidly growing with new models, features and vendor extensions being released at a very rapid pace. Every SaaS system we rely on will soon have GenAI capabilities embedded within them as part of their product offerings and it will be up to technology teams to determine how those capabilities are properly used. The platform approach to GenAI development leveraging API’s as a foundational building block will allow enterprises to integrate with structured and unstructured data source connectors, tap into 3rd party systems, optimize prompts, add or remove models, apply retrieval-augmented generation (RAG) to make responses more accurate, enhance existing systems and create whole new GenAI enabled applications. ?

Risk Management

It is always a critical concern for any enterprise to manage risks. This includes security risks, data risks, business risks, technology risks and of course architectural risks. As with concerns for system interoperability, scalability, flexibility and data protection a platform approach is the best way to consistently manage these risks along with a GenAI governance framework to decide how the platform will be used. A crucial part of the GenAI governance framework is the partnership of architecture with other teams such as compliance, legal, security, technology and privacy to determine the appropriate guardrails and controls that must used to evaluate each new GenAI capability, model, vendor integration and use case. If managed properly the benefits of GenAI will far outweigh the risks.

User Experience

Achieving a positive experience with GenAI depends on multiple factors including clear communication, intuitive UI design, good data quality, effective model training, retrieval-augmented generation (RAG) for better contextualized responses and a system architecture that meets non-functional requirements (e.g. performance, scalability, resiliency, security). Another important factor is user training. Users need to understand how to interact with AI, provide efficient prompts (i.e. prompt engineering) and leverage options for customizations that cater to their preferences. Both the AI system and its consumers must learn together to create an optimized user experience.?

Ethical Usage

GenAI is a very powerful technology and as Uncle Ben told Spiderman in the movies “With great power comes great responsibility”. Core to this responsibility is the ethical usage of GenAI capabilities. This means transparency into how we are using LLM as well as its limitations. As part of this transparency we need to demonstrate how models are trained, develop algorithms that are fair and unbiased (i.e. avoid reinforcement of existing bias in the training data), respect user privacy (data protection and consent for data usage), have accountability for detecting and correcting errors, ensure human oversight (always having a human in the loop), provide accessibility for all users (i.e. accommodating users with disabilities) and consider the social impacts such as potential job displacement and new skills training.

Continuous Improvement

Success with any technology is a journey and it is no different with GenAI. Depending on the industry involved (especially if it is a highly regulated one like Financial services) it may be beneficial to start as a fast follower rather than on the leading edge. This combined with an initial focus on lower-risk use cases to improve internal productivity (e.g. co-pilot efforts) means that an incremental approach to establish GenAI, enhance capabilities over time and respond to evolving regulatory and technology landscapes would be needed. The incremental approach would involve continuous monitoring of GenAI systems (platform implementation makes this easier) to optimize algorithms, validate controls, ensure data quality and resolve errors (hallucinations). It would also mean the gathering of continuous user feedback with a priority on improving user experience, data privacy and security to build user trust.

That is quite a lot to consider, but it is not really surprising given the impact, complexity and risks associated with GenAI technology. As impressive as what the current technology can do may be, it is really only the beginning. While we are focusing now on how to use large language models (LLM) the next frontier is spatial computing. In spatial computing we will give AI more capabilities with new sensory dimensions such as sight with large vision models and sound with large audio models allowing us to interact with digital information in the real world. This will enable improvements in augmented reality (AR), mixed reality (MR) and enhanced virtual reality (VR). The emerging spatial computing devices like AR glasses and VR headsets will bridge the gap between the physical and the virtual worlds by creating immersive virtual environments or integrating information into our real-world experiences.

With all of the advancements in GenAI I guess I better watch out or I could be replaced by a robot enterprise architect. Actually I am not really worried about that at all. While GenAI is very impressive, it is tool for us to use and not a threat. It has the capacity to improve our productivity, raise the quality of our work and to help us to make informed decisions faster. It is an assistant or a co-pilot to make us better at what we do but not a replacement for human judgment, creativity or skill. It is an invaluable aid that will be required in order to be competitive. As professor Richard Baldwin from the Geneva Graduate Institute said “AI won’t take your job, somebody using AI will take your job”. There are truly great opportunities ahead with GenAI and if we employ an intelligent strategy for AI architecture we will be well positioned to take advantage of them.