Where the AI decides and where AI Power Tools decides
Some work is done by AI Power Tools and returns the same result for the same input. Some is done by the AI model, whose output varies between runs and between models. The reference labels every operation, module, skill step and scenario step so you can tell which you are relying on.
The line follows the completeness versus correctness distinction: automation can confirm that a model is complete and follows your standards, and whether the architecture is right stays a human judgment.
What "deterministic" means here. Given the same inputs and the same model state, a deterministic step produces the same result, with no AI judgment involved. For anything that reads Enterprise Architect, the model state is part of the input. If the model has changed between two runs of the same query and the answer differs, the step is still deterministic, because its input changed. What deterministic rules out is a different answer from the same inputs and the same model state, and any answer that depends on AI judgment.
| Label | Meaning | Examples |
|---|---|---|
| Deterministic | AI Power Tools computes it, with no AI judgment involved. The same inputs and the same model state give the same result | create_element writes exactly the element you specify. export_xmi exports a package and reads the file back to count the elements. compare_baseline returns a structured difference. verify_diagram reloads and renders the diagram and returns the placed geometry |
| Rule-based | AI Power Tools applies fixed rules to make a judgment, such as recommending, ranking, classifying or flagging. Repeatable, but it can be wrong | assess_mdg_situation recommends which skill to use next. find_icon ranks icon matches. The query guard suggests the column you probably meant. The diagram linter reports overlap and alignment findings by rule |
| AI | The AI model decides. Results vary between runs and between models | Deciding which skill applies. Reading a source document and choosing which elements and relationships it implies. Choosing the diagram type that fits what a reader needs to understand. Writing the content of a ruleset. Explaining a result in business language |
| AI + AI Power Tools checks | The AI plans and drives; deterministic or rule-based steps do the computing and checking | Most skills and every scenario: the AI plans and chooses, operations do the writes, and verification operations check them |
In skill and scenario steps, each step is marked AI, AI Power Tools or You. AI is a step the AI model decides. AI Power Tools is a step that AI Power Tools computes or checks, through an operation or a fixed rule. You marks a decision or approval you must make. Two further markers cover the capabilities in Capabilities and surfaces: AI with computer use; otherwise you is a step the AI does when computer use is available and hands to you when it is not, and Engine is a step a code engine computes, which needs local code execution.
Where the AI decides
- Which skill and which operations to use, and in what order.
- What the model content should be, when you give it unstructured input.
- Which of several candidates to pick: an icon, a diagram type, or a language when two claim the same stereotype.
- How to word the outcome for the reader.
Where AI Power Tools decides
- Whether a call is well formed. Unknown and missing arguments are rejected before EA is touched.
- What was actually written. Operations re-read or count after a write where EA is known to accept writes it does not apply.
- Geometry in composed diagrams. The layout grammars compute coordinates by arithmetic, not by the AI's judgment.
- Conformance. A ruleset's verdict on a model is computed. The AI does not vote.
- The shape of the reporting business layer. The technology and the inclusion choice saved in the reporting profile decide which tables exist, what they are called and what they hold. The server computes them; the AI does not design them.
- Whether a reporting build may run, and whether it is right. The reporting preflight stops a build on a defect that blocks it. The alignment and value checks measure the result against the technology and against EA, and a check that compared no cells fails rather than passes.
- Safety rails. Restoring a baseline and deleting a package are explicit operations.
collapse_parallel_connectorsandprune_legacy_skillsare dry runs unless you ask otherwise. The migration script for older installations lists what it would remove and asks first.
What this means in practice. Use the AI for planning and judgment, and treat the checks AI Power Tools runs as the arbiter. When a skill says "verify", it names the check that computes the answer. Each skill's entry rates which models suit it. The ratings are guidance, not measured benchmarks.