Reconstructing the Enterprise: Data, State, Policy, and AI
The interesting thing about GenAI is not that it can answer questions or write code. It’s that it can reverse-engineer 'meaning', a topic I’ve posted about before.
If you drop a strong model into a real enterprise environment (with access to schemas, APIs, logs, documents, and event streams) it can infer, with surprisingly high fidelity:
➡️ what the actual system is
➡️ what the entities are
➡️ how they change state
➡️ what the invariants are
➡️ what “must not happen”
➡️ and what approvals or conditions gate progress.
Most enterprise systems are not complicated because the business is complicated, they are complicated because the plumbing is.
In my experience, over time, message buses, ETL pipelines, batch jobs, REST APIs, and workflow engines become the real system, while the data model and rules they were meant to serve get buried and become more difficult to maintain over time.
However, if you strip the plumbing away, what’s left is almost always something much simpler:
➡️ a set of entities
➡️ a set of states
➡️ a set of transitions
➡️ and a set of rules about who can trigger them and when
i.e. the semantic layer of the system - the intent and meaning of why the software was built in the first place, not the existing implementation.
This is exactly what an AI can reconstruct, and with the models getting better, and smarter, its starting to become viable to rebuild a competent 'clean' functional twin of a real enterprise application.
In exiting Enterprise legacy architecture agency lives in orchestration i.e. BPM engines, message brokers, routing rules, cron jobs etc.
An AI-native architecture moves agency to intent:
➡️ The model decides what should happen.
➡️ The policy layer decides whether it may.
➡️ The event system decides what actually did.
That is a clean, auditable, legally defensible chain of causality.
This is why identity, authentication, and policy are not optional. They are the moral and legal boundaries of autonomy. You can let a model reason all day, but you only let it act through a cryptographic identity, an authorization check, and a policy decision that becomes part of the permanent record.
That is how you get autonomy without chaos.
Once you see enterprise systems as:
➡️ world model
➡️ rules
➡️ state machine
➡️ policy
instead of:
➡️ integrations
➡️ APIs
➡️ workflows
➡️ microservices
the rebuild problem collapses.
You don’t have to migrate the plumbing. You have to reconstruct the semantics, and that is exactly what large models are starting to be really good at.
That’s the real changed in the enterprise that we’re living through. Not better chatbots, but the first tools that can read a tangled, decades-old enterprise and infer what it means well enough to recreate it cleanly.