With the current unmistakable surge in interest in modern Artificial Intelligence (AI) tools and applications, many organizations around the world are considering how to best leverage the rapidly multiplying AI technologies and tools to bolster their processes and enhance their market
success.

In the current environment, the first thing that comes to mind when AI is mentioned, is the AI genre known as Generative AI, which is largely based on Large Language Models (LLMs), and exemplified by well- known applications such as ChatGPT, Claude and Gemini, to name but a few. We can say that those models are statistical generation or prediction models, enabled through training on massive amounts of data available in the public domain, especially on the Internet. The applications based on LLMs effectively chart superhighways into publicly available human knowledge and accumulated, documented, expertise.

Beyond Generative AI: The Rise of Reasoning Models
Nowadays, we have an even newer genre: Large Reasoning Models (LRMs) that are supposed to ‘reason’ to find solutions to problems that were not included in their original training sets. But those systems are still in their infancy.

Yet, AI is far broader than the generative and reasoning models currently dominating headlines.

A Brief Historical Perspective: From General Intelligence to Narrow Expertise
The early days of applied artificial intelligence attempted explicit representations of general human knowledge and reasoning. However, this proved to be unrealistic as the software and hardware technologies at the time could not cope with the sheer magnitude of the task of engineering the constituents of general intelligence into machines. The quest for integrating human intelligence was then directed to narrower domain of expertise, such as disease diagnosis, game playing and financial fraud detection, etc.

This required significant efforts to “engineer” and represent expert knowledge, followed by the requisite effort to refine, validate and maintain such knowledge; again, the promise was far greater than what was actually delivered. Some applications that carry strong AI and machine learning flavors such as those achieving accurate image recognition, machine translation and speech recognition, were largely bottom-layer algorithmic, meaning their cores relied on algorithms to obtain results by processing mathematical structures that mimic biological neural networks, achieved great success.

However, this still did not lead to automated systems that achieved high intelligence-requiring task performance, like perhaps writing or improving a poem. The current gen AI still can’t be considered to be the harbinger of the goal of artificial general intelligence (AGI). Hallucinations and the built-in biases of these systems are well-known. Attempts to improve and configure more intelligent systems continue.

It is my prediction that we shall return to combine programmatic artificial intelligence, based on explicit encoding of knowledge and some kind of ‘inference engine’, as in the earlier knowledge-based systems, with systems that have generative, or reasoning AI components based on large language (or foundation) models.

Matching the Right AI to the Right Organizational Context

This myriad of AI technologies demand that their successful exploitation needs carefully matching the capabilities and limitations of each type of AI to fit the right part of the organization’s operating model, followed by deliberate, and careful assembly of capabilities to support the chosen types, as well as a process model for how the initiatives are to be sequenced in time, making allowances for experimentation and re-work.

This requires organizational modelling from knowledge and reasoning perspectives, followed by configuring operating models with AI-oriented transformations, and an experimentally led process along the lines of agile software development or design thinking. It would probably be best to follow and incremental/ experimental approach to deploying various types of AI to allow for local expertise and focused teams to emerge. The evaluation of experiments requires agreeing explicit evaluation frameworks so that the results of the experiments are appropriately judged and the results fed into subsequent stages.

Architectural Thinking: A Strategic Imperative
The deployment of modern AI should be managed through architectural thinking. By architectural thinking I mean a style of thinking that pays attention to the surrounding organizational and business landscapes, the nature of the users and their tasks, as well as the facilitation of future evolution and change. The AI architect must consider the deployment of an appropriate, supportive, eco-system. A favorable organizational and eco-system (human, technical and procedural) and capacities are needed. People’s skills and attitudes need to be oriented towards capturing and curating organizational knowledge and learning.

Technologies that support the curation, sharing and expansion of knowledge need to be determined and implemented. Here are some questions to answer when embarking on AI deployment:
1. Does the human and organizational climate encourage the sharing and preservation of knowledge and experiences?
2. Do the appropriate technologies, such as specialized databases/ data warehouses and sources, and their supporting elements exist?
3. Do we have the right leaders and leadership styles?
4. Do people have the mindsets to seek opportunities for deploying AI and knowledge systems in useful ways?
5. Do we have supportive norms, policies and strategies in place?

Architecting for Sustainable Impact
Based on authentic answers to the above question, leaders can decide on what needs to be done, and AI architects can set about “architecting” an operating model that heeds the surrounding “landscape” and deploys the right types of AI in the right places in the organization to reap the best benefits. This should be done while also designing a robust process for experimentation, monitoring and refining the AI-focused interventions. Successful deployment of AI needs careful “architecting”.

This article is by Galal H. Galal-Edeen, professor of information systems architecting, Heikal Department of Management, Onsi Sawiris School of Business, The American University in Cairo.