Despite all the hype, most companies are not, OpenAI, Amazon.com, Tesla or Netflix if it comes to Artificial Intelligence (AI). They don’t have massive infrastructures. They don’t have several Data Science and Data Engineering teams. The aren’t fully data driven, what ever this means.
But even before ChatGPT saw the light of day, AI was a topic and still is, besides Generative AI. Current IT systems already offer a lot, which is already used in different ways. Today we can identify three worlds of AI (at least) as I show here:
Fig. 1: The three worlds of AI
Every „World“ delivers AI today. This comes with some implications:
Different operational models are necessary or in place
AI & Data Governance is complex through distributed usage of AI
A unified perspective is often missing to develop d Data and AI Strategy
Technical AI relevant capabilities develop at a rapid pace
There are specific capabilities in each of these worlds that support AI in their respective context.
Fig. 2: AI Capabilities are different in each world
When implementing use cases, decisions are often made on a situational basis in favor of the appropriate platform, without considering the overall picture.
Let’s discuss what that could mean based on customer examples I have seen in practice. All the examples are on their way to see how GenAI can be used. But we also see that GenAI can not replace most of the established cases or the maturity is not there.
Example - Operational AI in a Retail World
In the first example, we have a retail company for which logistics plays a very important role and is considered as a core process. At the same time, the various areas have an extremely high degree of autonomy.
Fig. 3: Example 1 - Operational AI in retail
This means that the AI capabilities of the respective operational system are primarily used. The customer also has a central team that creates forecasts, for example, on the central platform. However, the hurdle for this is significantly higher. The AI strategy tends to follow the respective operational system.
This results in decentralized responsibility for AI with the following observations:
In-house developments with a tendency towards standard software
Strong operational focus for AI
High departmental orientation
Logistics as a core process
Innovation along the provider roadmap
After years of being focused operational/decentral this results in the following challenges:
Low maturity of the data platform due to strong opertional and decentral focus
Decentral BI fontend strategy, means many local tools
Partly massive MS Excel ecosystem for data exchange
Overload of operational systems with external data
The overall data architecture is a Big Ball of Mud due to local initiatives and many point to point connections
Overall we observed a decentral AI responsibility not coordinated leading to all the effects described.
Example - Multi-Platform Approach for Data & AI
The second example is also strongly business-driven in the financial service subsidiary of a larger enterprise. In order to achieve greater flexibility (with reduced IT governance), a separate environment was set up for R and Python, which is technically provided by central IT, but the administration and governance and responsibility lies with the business departments.
Fig. 4: Example 2 - Business-oriented data platform in financial services
There is also a central DWH for standard reporting and planning with a stronger governance. We made the following observations within the organisation, why this was needed:
Department-oriented AI & reporting platforms due to very individual needs and high domain knowledge
Central platform support for standard reporting and planning processes
Local governance and resource steering give high flexibility
On this business-driven approach we observed the following challenges:
Low maturity of the overall Data approach
Low IT governance and data governance maturity for Data & AI
Multiple versions of truth
Overall we observed a missing overall AI strategy which results in rising efforts, complexity and technical debt.
Example - Making use of modern BI tools
My last example is from a sales planning project of an automotive supplier. Coming from traditional reporting and OLAP tools, the new project brought new tools which leaverage AI. Augmented analytics, which uses natural language processing (even before ChatGPT) and the automation of reports and functions, has been developing for years now.
Fig. 5: Example 3 - Augment analytics with AI in a modern BI environment
AI can thus be used supporting Business Intelligence, as well as for data analysis and machine learning tasks itself.
Often, this still fails due to trivial results, a lack of data quality or simply a lack of knowledge (literacy) in dealing with the new possibilities.
We observed the following situation during the project:
IT-centered data platform for BI is delivering a single source of truth
Extension of the classic reporting use case through AI
AI can support in ad hoc situations or in a dynamic environment
But the approach also shows some trade-offs:
Very limited and specific set of methods within the BI solution
High data and AI literacy was necessary to really make use of the AI functionalities
Having a central data warehouse as source is often showing missing flexibility
Overall we experienced a missing central AI strategy supporting users for realizing the full potiential of available AI capabilities.
Conclusion
Several years ago the AI world was easier, as AI was not applied to all the different systems. Today we have to handle multiple worlds, offering multiple solutions or tools, each delivering specific AI capabilities.
We can not handle the side effects described by only looking on the technology level. A holistic approach should cover the following aspects:
Clear vison and goals for the company
An operating modell, supporting central and decentral demands with the necessary flexibility
A supporting Data & AI Governance approach giving guardrails where necessary
A holistic data architecture considering the above mentioned aspects
While the right data architecture is always individuell, modern data architecture patterns like Data Mesh and Data Fabric consider already many of these aspects and give orientation on best practices.
AI is a rapidly changing field. What are your experiences with AI in an enterprise context?