An AI agent reads a report and suggests a useful next step. Before that suggestion can change a record in a business system, some familiar questions need answers: which data applies, which rule governs the decision, who is responsible, and who may authorize the action?
My whitepaper “Designing Agentic AI for the Enterprise” proposes an architecture that connects enterprise knowledge, decisions and controlled execution. It follows a request from the information an agent uses to the point where a business system confirms the change.
Knowledge needs sources, and actions need authority
A glossary explains terms. An ontology describes business objects and their relationships. A knowledge graph connects concrete objects to what is known about them. None of these makes a source current or grants an agent access to it. Ownership, validity, provenance and permissions remain part of the architecture.
The same care applies to decisions. Code can calculate a defined result, rules can evaluate explicit conditions, and a language model can interpret a document or propose a next step. The surrounding system must decide when review is needed and enforce whether a specific action may proceed. A convincing explanation is useful; it does not grant permission.
The whitepaper develops these relationships into an outline of an agent runtime. It explains how a request retains its state across steps, how an action contract binds its target and inputs, and what must happen when approval arrives later or a service times out after receiving a write request. An unknown outcome needs reconciliation before dependent work continues.
Start with one decision
You do not need a complete model of the enterprise to begin. Choose one decision with a manageable set of sources and a clearly defined action. Work through it with the business and application teams: where does the information come from, who can approve the change, and what must the target system confirm?
Use that work to agree a small knowledge model, an action contract and acceptance cases for the implementation. These give architecture and engineering a concrete basis for deciding what to build and how to check it.
Read the full whitepaper online · Download PDF · Subscribe to the newsletter