Put collected information into consistent records and explicit relationships, preserve where each fact came from, and query the repository with a defined scope. Use filters and reports for exact questions, relationship traversal for questions that cross objects, and AI to translate natural-language questions into those operations and explain the results.
The repository can answer only from the facts and relationships it contains. A missing result may mean "not recorded," not "does not exist." Every useful answer should show its scope, source records, missing data and interpretation limits.
Disclosure: I’m Nick Malik. I built Useful EA, a commercial set of architecture tools, and the section on Useful EA below describes it. The method in this article works with any repository and any AI assistant.
Collection is not yet an answer
Architecture information may arrive through interviews, workshops, surveys, spreadsheets, system exports, documents or AI extraction. Before it can support reliable questions, it must be converted into repository knowledge:
- identify the object;
- classify its type;
- resolve its identity and aliases;
- record useful attributes;
- connect it to other objects with named relationships;
- preserve source and review information; and
- assign an owner and review date where appropriate.
A folder of documents contains evidence. A repository adds consistent identity, meaning and relationships so the evidence can be reused.
Manual and AI-gathered facts use the same trust process
Do not assume manually entered data is correct or AI-extracted data is wrong. Treat both according to evidence.
| Evidence state | Meaning | Appropriate use |
|---|---|---|
| Proposed | Extracted or submitted but not reviewed | Review queue; do not use as accepted truth |
| Confirmed | Checked by an accountable person against a source | Normal analysis within the confirmation scope |
| Estimated | Best available value with method or rationale recorded | Scenario and planning work with visible caveat |
| Inferred | Derived from other records or rules | Useful for investigation; distinguish from direct evidence |
| Conflicting | Credible sources disagree | Resolution work; do not silently choose one |
| Unknown | Required information is absent | Data-quality queue; never convert automatically to zero or "none" |
Keep the source, date, reviewer and evidence state wherever the decision needs that level of confidence.
Relationships make architectural questions possible
Attributes answer questions about one object: owner, cost, lifecycle, maturity or status. Relationships answer questions across objects:
- application supports capability;
- capability contributes to strategy;
- application sends data to application;
- technology enables application;
- program addresses capability gap;
- project depends on system; and
- benefit depends on business change.
The relationship name and direction must be explicit. "Connected to" is usually too vague for analysis.
Match the question to the required data
| Architectural question | Minimum repository data | Result type | Important boundary |
|---|---|---|---|
| Which applications support this capability? | Applications, capabilities and support relationships | List grouped by role or status | Only recorded support; not every operational dependency |
| Which capabilities need investment? | Current and target assessments, strategy links and calculation rules | Ranked gaps plus missing data | Ranking informs review; it does not allocate funding |
| What recorded dependencies would an application retirement touch? | Directed application links plus capability and technology relationships | Neighbors, paths and review checklist | Not a complete impact analysis or proof that retirement is safe |
| Which initiatives support this strategy? | Strategy, objective, program or project and dependency links | Recorded paths and disconnected items | Missing links may reflect incomplete data |
| Where is portfolio cost concentrated? | Annual cost, currency and scope, grouping relationships | Totals and exceptions | Estimates and mixed currencies must remain visible |
| Which records are least trustworthy? | Source, review date, owner and evidence state | Data-quality queue | Absent metadata does not prove the underlying fact is false |
Use the simplest query method that answers the question
Filters and lists
Use filters for exact questions such as "active applications with no owner" or "capabilities with no current assessment." These results should be easy to reproduce.
Aggregations and reports
Use defined calculations for counts, totals, averages and rankings. Let the database or report do the arithmetic; let AI explain the result. Always show the formula, currency, period and treatment of blanks.
Relationship traversal
Use path queries when the question crosses relationships: which applications support a capability, which systems depend on a platform, or how a project connects to strategy. Return the actual path and relationship names.
Natural-language AI questions
Use AI as an interface and explainer. A strong AI answer translates the question into repository operations, uses only the data it is allowed to see, cites the records and relationships used, and states what is missing. It should not replace absent repository facts with general knowledge.
A reusable prompt
PromptAnswer this question only from the repository data I have given you: [question]. State the repository, model or portfolio scope. List the records and relationships used. Separate confirmed, estimated, inferred, conflicting and missing information when those states are available. Show any calculation or path. If the data cannot support the question, say which data or relationship is missing. End with the next human decision or validation step. Do not fill gaps from general knowledge.
Example: from collected data to a retirement review
Suppose application information was collected from owners and supplemented by AI extraction from solution diagrams.
- Application records are matched to stable identifiers.
- Owners confirm lifecycle, cost and capability support.
- Extracted application links stay proposed until both relevant owners review them.
- Accepted facts and links enter the portfolio with source dates.
- The architect asks: "What recorded dependencies would retiring Legacy Billing touch?"
- The repository returns direct recorded links, supported capabilities, technology dependencies, open work and missing evidence.
- AI formats that result as a review checklist.
- Owners and technical teams decide whether the recorded picture is complete enough to proceed.
The AI did not decide to retire the application. It made the repository evidence easier to assemble and inspect.
Follow-up questions that expose uncertainty
- Which part of this answer relies on estimated or inferred data?
- Which records are older than the chosen review threshold?
- Which expected relationship types were unavailable?
- Which owners should validate the result?
- What source contradicts this answer?
- Which calculation rules produced this ranking?
- What changes if unresolved records are excluded?
These questions are often more valuable than asking the AI for a stronger recommendation. For questions specific to one kind of model, see the companion articles on capability models, Benefit Dependency Networks and application portfolios.
Where Useful EA fits
Useful EA’s tools are separate applications that share one sign-in. Architecture Repository keeps elements and named relationships typed by a metamodel, and can link its elements to records in the other tools and show their current names. Application Portfolio Manager can map applications to capabilities in a Capability Modeler model. The tools don’t otherwise copy or synchronize data.
Claude, ChatGPT, Claude Code, Cursor and VS Code can connect to Useful EA directly at mcp.useful-ea.com. You sign in with your usual Useful EA account, and the assistant reads only what you can already open, with your permissions; it can’t change your models. A company’s administrator can turn assistants off for the company’s members. How to connect an AI assistant.
Connected to Architecture Repository, an assistant can read the metamodel, find elements by type, viewpoint, lifecycle or text, and open an element with its properties, its relationships and its links to records in the other tools (shown as restricted when you can’t open the record). It can read Capability Modeler, Benefit Dependency Network, Application Portfolio Manager and EA Maturity Assessment the same way, so one question can draw on several tools, each under your own access. Its answers come from the live data, so ask it to cite the records it used.
You can also export instead: the whole repository as JSON or one element type as CSV, and CSV from every other tool. An assistant working from an export answers from that file, not the live tool, so record the export date with any answer you share.
Only give architecture data to an AI assistant your organization has approved for that information. An export can include names, owners, costs and plans that are confidential.
Practical rule
Collect facts once, preserve their meaning and evidence, and reuse them across questions. Let repository operations find and calculate; let AI translate, traverse and explain; let accountable people validate the evidence and make the decision.
Sources
Vendor documentation is cited as context for how the market describes these patterns, not as evidence about Useful EA.
- Essential AI — natural-language questions grounded in a connected model, with answers linked to model elements
- SAP LeanIX: MCP server for enterprise architecture — structured AI access to application records, capability maps, decisions and relationships
- Ardoq: Omnipresent AI Assistant — natural-language repository questions, document extraction and changes staged for authorized review
- Enterprise Architecture Solutions: who governs the knowledge if AI populates your repository? — provenance, conflicting sources and domain-owner review