# How do you keep an enterprise architecture repository up to date with AI curation?

> Canonical: https://decisionpen.com/blog/how-to-keep-an-enterprise-architecture-repository-up-to-date-with-ai  
> Published: 2026-10-27 · DecisionPen · Living EA series (D8d)  
> Related docs: /docs/incremental-modelling, /docs/architecture-decision-workspace, /docs/enterprise-ask, /docs/beyond-traditional-ea

## Summary

Keep an enterprise architecture repository current by running it as a continuous loop rather than an annual modelling project. AI proposes entities, relationships, and descriptions whenever new evidence arrives; an accountable architect curates them in a short review queue; accepted changes update the map; freshness metrics show drift early. Builds on human review and curation (D7c), keeping factsheets current (D7f), and the Discovery Tax.

## Why repositories go stale

Upkeep is manual, pays back late, and rarely benefits the person doing it. It is deferred, batched into an annual project, then distrusted. A stale repository also forces each new programme to rediscover the landscape: the Discovery Tax paid again.

## Annual project versus continuous refresh

- Annual project: large push, high accuracy on day one, steady decay afterwards.
- Continuous refresh: many small updates tied to real events; accuracy is maintained rather than restored; each update is cheap because AI has drafted it.

Same rigour, applied in smaller pieces, closer to the evidence. This is incremental modelling in practice.

## AI proposes, humans curate

- AI proposes: new entities, typed relationships, updated descriptions, flagged contradictions, each with evidence.
- Humans curate: accept, edit, reject, or defer. Rejections are remembered.
- The map records the outcome in the Context of Record, with provenance.

## What triggers a refresh

New or revised design documents and decisions; releases, migrations, and decommissioning; incident reviews that expose undocumented dependencies; refreshed spreadsheet exports from an inventory or CMDB; new programmes needing a slice of the landscape.

## Running the review queue

Prioritise high-impact entities; time-box a regular short slot (for example weekly); review claims rather than every field; give use a role, because the map is needed for real work (see the incremental modelling flywheel, D8e).

## Measuring freshness

Description and relationship coverage; age of evidence behind key entities; accepted-versus-rejected rate of proposals; review-queue age and unresolved contradictions; questions the map could not answer. Trends matter more than one score.

## How DecisionPen solves this

Architecture Decision Workspace: incremental modelling from documents and spreadsheets; AI proposals with evidence; human curation; a living map that feeds Enterprise Ask and generated ADRs, blueprints, and diagrams. Trial: https://decisionpen.com/trial

## When is DecisionPen the wrong tool?

Unattended, auto-published updates with no steward; live infrastructure scanning; a CMDB sync. DecisionPen proposes from the documents and spreadsheets you give it, for human curation.

## FAQ

**How do you keep an enterprise architecture repository up to date?**  
Replace periodic modelling projects with a continuous loop: events trigger AI proposals, a steward reviews them in a queue, accepted changes update the map, and metrics show where it is going stale.

**Why do enterprise architecture repositories go stale?**  
Upkeep is manual data entry with no immediate payoff for the person doing it, so the repository is rebuilt annually, drifts, and is distrusted — the Discovery Tax paid again.

**What does AI curation actually mean?**  
AI proposes; humans curate. Nothing becomes authority without review.

**How often should the repository be refreshed?**  
Continuously, in small steps tied to real events, with a short weekly review slot for the highest-impact proposals.

**How do you measure whether the repository is current?**  
Coverage, evidence age, accept-versus-reject rate, review-queue age, and unresolved contradictions.

**When is DecisionPen the wrong tool?**  
When you expect unattended publication, live scanning, or a CMDB sync.

## Related posts

- https://decisionpen.com/blog/how-should-humans-review-and-curate-ai-proposed-architecture-models
- https://decisionpen.com/blog/how-to-keep-ea-factsheets-current-with-ai-descriptions-and-relation-recommendations
- https://decisionpen.com/blog/what-is-the-discovery-tax-in-it-transformations
- https://decisionpen.com/blog/why-architects-hate-manual-ea-modelling
- https://decisionpen.com/blog/how-to-assess-an-it-landscape-without-months-of-workshops
- https://decisionpen.com/blog/how-does-incremental-modelling-create-an-architecture-knowledge-flywheel

## Series (Conversion how-tos)

- D8a: https://decisionpen.com/blog/how-to-fast-track-ma-it-due-diligence-with-automated-landscape-modelling
- D8b: https://decisionpen.com/blog/how-to-build-application-dependency-maps-from-existing-documentation
- D8c: https://decisionpen.com/blog/how-to-assess-an-it-landscape-without-months-of-workshops
- D8d: this post
- D8e: https://decisionpen.com/blog/how-does-incremental-modelling-create-an-architecture-knowledge-flywheel