# How should humans review and curate AI-proposed architecture models?

> Canonical: https://decisionpen.com/blog/how-should-humans-review-and-curate-ai-proposed-architecture-models  
> Published: 2026-10-27 · DecisionPen · Living EA series (D7c)  
> Related docs: /docs/incremental-modelling, /docs/architecture-decision-workspace, /docs/enterprise-ask, /docs/why-not-copilot-or-rag

## Summary

Treat every AI proposal as a draft with evidence. Review in a queue where each item can be accepted, edited/enriched, rejected, or deferred before joining the map. Prioritise by blast radius and evidence strength; record rejections; measure trust. The goal is a trusted slice you would defend — not unsupervised 100% coverage.

## Why curation is required

Failure modes are quiet: hallucinated integrations, merged lookalike systems, deprecated designs read as current, roadmap aspiration read as fact. Because the map feeds impact analysis, Ask, artefacts and delivery agents, unreviewed errors become Context Tax later.

## The review queue

Each item shows claim, source evidence, and effect of acceptance.

- **Accept** — right and evidenced; joins the living map.  
- **Edit / enrich** — correct name, relationship or attribute, or add missing context, then accept.  
- **Reject** — wrong, out of scope, superseded; record why.  
- **Defer** — plausible but unproven; keep as an open question.

## What to check first

1. Relationships over attributes.  
2. Weakly evidenced claims.  
3. Conflicts between sources.  
4. New entities resembling existing ones (merge early).  
5. Lifecycle and ownership changes (decisions, not inferences).

## Trust and governance

- Named stewardship per domain.  
- Evidence stays attached for audit and challenge.  
- Measure accept/edit/reject ratio, time-to-first-trusted-slice, Ask citation usage; a ~100% accept rate with no edits may signal rubber-stamping.

## Curate-Don’t-Create

Architect time goes to judgement (right? current? material?), not retyping what documents already say. AI drafts; humans ratify.

## Should AI auto-approve?

No. Unsupervised near-100% coverage recreates a hollow, untrusted repository faster. Automation helps around the gate (pre-sorting, grouping, surfacing conflicts and gaps); the decision to make something canonical stays human.

## How DecisionPen supports curation

AI output does not write straight to the Context of Record. Proposals from documents, dialogue and spreadsheet imports are queued with bound evidence; architects accept, edit, reject or defer; Enterprise Ask and artefacts then draw on the curated map, with gaps kept visible. Trial: https://decisionpen.com/trial

## When is DecisionPen the wrong tool?

If you want fully automated population with no human review; or Visio/Lucid drawing, BPMN, Copilot-only text drafting, or operational CMDB-only tracking is the primary need. If nobody can own curation of a domain, fix that first.

## FAQ

**How should humans review AI-proposed architecture models?**  
In a queue of evidence-backed proposals: accept, edit/enrich, reject or defer, prioritising relationships, weak evidence, conflicts and new lookalike entities.

**Why can’t AI auto-approve its own proposals?**  
Errors are quiet and propagate into impact analysis, Ask and agent context. Coverage without review is untrusted coverage.

**What should reviewers check first?**  
Relationships, weakly evidenced claims, source conflicts, possible duplicates, and lifecycle/ownership changes.

**How do you measure whether curation is working?**  
Accept/edit/reject ratio, time-to-first-trusted-slice, Ask citation use, and fewer conflicting answers.

**How does DecisionPen help?**  
It queues evidence-linked proposals for architect ratification before they become the Context of Record, and answers from the curated map.

**When is DecisionPen the wrong tool?**  
For unsupervised auto-population, or when drawing, BPMN, Copilot-only drafting or CMDB-only tracking is the sole need.

## Series index (D7)

- https://decisionpen.com/blog/what-does-ai-do-inside-an-architecture-decision-workspace  
- https://decisionpen.com/blog/how-to-model-enterprise-architecture-from-structured-and-unstructured-sources-with-ai  
- https://decisionpen.com/blog/how-should-humans-review-and-curate-ai-proposed-architecture-models  

## Related posts

- https://decisionpen.com/blog/how-ai-accelerates-automated-enterprise-discovery  
- https://decisionpen.com/blog/from-automated-discovery-to-durable-architecture-context  
- https://decisionpen.com/blog/why-ea-must-own-enterprise-context-in-the-agentic-era  
- https://decisionpen.com/blog/what-is-the-context-tax-in-enterprise-it  
- https://decisionpen.com/blog/why-architects-hate-manual-ea-modelling  
- https://decisionpen.com/blog/ai-enterprise-architecture-copilot-vs-metamodel-agent