Enterprise AI: Its Linkage to Enterprise Architecture
Author: Deepa Naik?
Even though Artificial Intelligence(AI) has been much talked about in recent years - backed with increasing research and a plethora of companies building products in this space; adoption of AI especially in the enterprise context is lacking and has not been meeting the same pace (source: OReilly AI Adoption in the Enterprise 2021).
The technology leadership / CTO / CIO in most organizations are struggling to bring AI based enterprise applications out of the labs to productionized environments and building strong business case around AI investments.
In this article we try to understand some of the bottlenecks for AI adoption in the enterprise and explore how the practice of enterprise architecture can guide a way forward for the leadership teams.
Challenges for Enterprise AI Adoption
One of the primary reasons for lack of AI adoption in the enterprise context, is the nature in which AI works. AI implementation is highly dependent on the type of data that your organization possesses, what is your core competence as far as data is concerned, a variety of technological skills, the technology capability that the organization holds and the kind of business problem that you are trying to solve.
Moreover, if you decide to adopt the buy as against the build option, or decide to just consume AI instead of developing it inhouse, then it boils to the challenges that your IT and operations team faces to integrate these systems.
In some cases, it might also need a cultural shift for your organization as a whole. To top it all, deriving the ROI or the value proposition of introducing AI application(s) or solutions(s) into your enterprise has its own roadblocks.
Figure 1: Bottlenecks to AI Adoption ( source: OReilly AI adoption Report 2021 )
Enterprise Data Context and Strategy – a bird eye view
Most enterprises build and scale up as group of systems (enterprise applications) over time; built with point-solutions to solve specific problems. It becomes difficult to look at data in totality as an enterprise data context or the data strategy. Not only does AI deal with a vast amount of data, but the variety it deals with – structured as well as unstructured; text, images, audio, video etc. Moreover, with an increasing number of applications operating as distributed systems (integrated with external systems and comprised of multiple processes that must communicate to perform work), event based or streaming/real time data, services with data on the inside vs. data on the outside, the data challenge is multi-fold.
Having an enterprise-level bird eye view of your "data treasure” is a key ingredient absent with most organizations
Heterogenous Technical Skills and Capability
Another important aspect is the versatile skill-set requirements for the development of AI based software.
The skills necessary for a successful AI project, are tied closely to the type of data sets available and accessible, the machine learning models chosen and the algorithms that will be needed.
For example, if you have unstructured text data or plan to use chatbots – it deals with NLP (natural language processing); with recommender systems it deals with graph databases, predictive algorithms, labeling of data and supervised learning; with IDP( intelligent document processing or image processing it gets into pattern matching and so on and so forth.
So, depending on the use case and the type of problem at hand, you are dealing with a distinct set of technological skills and resource pool.
Infrastructure Needs and Productionizing AI Systems
Whether on premises or cloud or hybrid, one cannot ignore the complexity involved in productionzing - deploying and monitoring AI system. Infrastructure sizing becomes critical not just where the data is hosted, but the processing and computing for the high volume in question becomes a key, if its real time or data at rest. ?Some amount of thought needs to be invested in to the ETL (extraction, transformation, loading) processes, data quality procedures to separate out the transactional data vs. the model data. AI systems need to be aligned with existing CD / CI or similar IT operations processes followed by production teams.
ROI and Business Value Proposition
Identifying the appropriate AI use cases in order to derive a business value proposition (Business Use Case) is one of the most challenging piece as it requires deep expertise both on the domain / business as well as the technology / data side of things.
Getting together a good solutions team to understand stakeholder requirements, work on existing pain areas and exploring opportune for new AI paradigms becomes a key aspect of the value stream generation.
The ROI for AI projects is generally more complex than other technology projects – one issue is that there is not too much historic metrics to estimate in terms of productivity numbers as the AI technical advances are fairly recent. Also, the very aspect of data including softer aspects of data quality, reliability and so on put in its own overheads during the actual implementation. The development vs. operations aspect needs to be considered and that puts in more challenges for estimating one time vs. on-going costs.
Integrating AI Applications in the Enterprise
Moving from developing AI products built as stand-alone or in silos to having them integrate as a part of the enterprise application(s) portfolio is a crucial aspect.
领英推荐
How does the AI application or solution fit in into the enterprise context?
One needs to consider more aspects to glue-in the AI solution and weave it into your business process or business service to provide value to your customer and stakeholders.
Way Forward – the EA and AI handshake
Having looked at the top challenges for enterprises to integrate AI systems into their enterprise application(s) portfolio, let us now explore how focusing on the enterprise architecture practice will help resolve some of these.
To start off, let’s get some definitions to be on the same page for our discussion
Enterprise - An enterprise is any organization that uses software systems
Enterprise Architecture - is a practice, a collection of skills, that aligns technology strategy with business strategy.
Architecture - in this context, means the complex way in which computer and computer systems are organized and integrated
Anchoring AI Use Case in the Enterprise Architecture
As per OpenGroup’s TOGAF technical architecture framework standards the basics of enterprise architecture can be constituted as a series of layers or levels. Some of these key layers include
Figure 2: Accelerating Enterprise AI Adoption - The AI - EA Handsake
In most practical scenarios these can be further sub divided into
Business Architecture Layer
Data/ Information Architecture Layer
Application / Systems Architecture Layer
Technology / Infrastructure Architecture Layer
The AI Use Case is intertwined across all of these layers
Most of the AI product companies focus on the data and application layer. However, in the Enterprise AI context, business managers need to consider the AI Use Case covering all the four tiers
The hand-shake between the the realms of the technology teams across these tiers becomes important for successful Enterprise AI adoption.
With evolution of architecture patterns like headless architecture, microservices, hexagonal patterns; popularity of of APIs, messaging and event based platforms and like Kafka and distributed data management concepts like data mesh; low-code no-code paradigms and AI products, platforms and solutions becoming ever more mature; it is time for the enterprise apex leaders / technology leadership to look at AI technology not in isolation, but rather how it can be anchored and cemented within the context of enterprise architecture.
Artificial Intelligence and Analytics (AIA) - Cognizant
2 年Excellent article, the bottlenecks are so relevant. Another important aspect I am seeing is the lack of skilled UI developers who can integrate with a AI based services architecture working in real-time, which gives the Business Users the flexibility to change the inputs on the screen and see how the prediction varies, as well as visualize the data with interactive plots and charts, rather than having some static probability figures or prediction figures shown on the screen (Take it or not, its your problem, kind of attitude!). That key aspect to involve Business Users all through the process, of data exploration, model building is somewhat missing. Python based UI like streamlit might be more relevant here, as traditional UI will fail to show some of these interactive plots and charts, and this is where there is a gap. Continuous model monitoring and retraining is also something Business is keen about, and it seems current technology is not well equipped for the needs of the Business here. Process compliance and legal aspects are also taking a heavy toll on the timeline of AI implementation in production.
SDE III at Walmart Global Tech | Innovating for Global Impact | MS @ UConn | M.Tech @ MIT Manipal | Ex-JPMorgan Chase & GE Healthcare
2 年Excellent
Sahastraar - Groundbreaking Startup for Training, Applied Research & Consulting in Project Management & Industry 4.0
2 年A good article. Thanks for sharing.