At this point many companies are trying to move from the initial Task Assistant style of AI usage to the Knowledge Agent and Business Problem Solver style (see the report Nordic AI archetypes, which distinguishes four archetypes of AI usage in Nordic organisations). Whereas, the Task assistant is simple to roll out since you just give access to the employees and train them in using an LLM, the other usage archetypes require more. The knowledge agent for example requires you to curate knowledge, often from many different sources, and make it available to the agent. The Business Problem Solver handles entire segments of the value chain and the processes and interactions with humans needs to be carefully orchestrated. Unfortunately, most companies fail to move beyond the initial step of making AI accessible. It is a major challenge to move to more complex and more valuable implementations from pilot to production. It turns out the Enterprise Architecture (EA) leaders have a head start.
To have a half decent chance of getting any agentic system to do something helpful you need to be able to understand how you work and document the workflow. It is needed to determine where humans are needed to make a decision, for example, or to figure out when to invoke the various agents. If your organisation was so lucky as to have already documented its processes as part of the enterprise architecture, maybe even in BPMN, you are in luck and poised to leverage this good work for quick agentic AI results. As an example one company in the pharmaceutical industry had updated their ways of working and documented all their processes in detail in one central repository. Based on these BPMN diagrams, they were able to generate hundreds of SOPs with the help of an LLM in days, which, needless to say, saved hundreds of hours.
If your organisation also had the foresight to have developed a logical data model that documents what people call the different information structures that are used in business problem solving and maybe even described it in a data catalogue, there is a similar boost to your AI adoption, because it canbe used to explain the agentic AI system how information and data are related. That way you can control its urge to hallucinate and make up relations between things that are not helpful. You can also use it if it has problems finding exactly where information is in the underlying data sources since you can map the logical data model to the physical data model. With this in hand leaders have begun to build interfaces to data bases that use information from the data catalogue to let end users retrieve data from operational databases without any SQL knowledge.
Knowing what systems an organisation has and what data they contain will help you quickly map out where to get what information and what systems to connect to do something. When it is time to scale the solution this is crucial. To connect the brain of the agent with systems that do something, like sending emails, scheduling appointments or paying invoices, integrations are needed. Model Context Protocol defines these interactions and MCP servers can be set up with the right systems and the right data access the first time. This is the backbone of any autonomous or semi autonomous system, if we don’t know what systems and APIs to connect to and how, the agentic system is never going to be more than just a fascinating reviewer or conversation partner.
EA is thus the operating system for successful agentic AI: business processes tells it How to work, Data models are its vocabulary, and system architecture defines its hands and feet that allows it to actually do something.
Business process modelling, data modelling and system architecture modelling are core Enterprise Architecture skills that will allow you to leverage your existing solutions more easily, accelerate development and scale deployment beyond the pilot phase. These disciplines will help your organisation reach the next steps of your AI journey and move to the Knowledge Agent and Business Problem Solver level, and absolutely critical to ever achieve any kind of autonomy.