01

Turning metadata gapsinto an accountableremediation workflow

Company
IKEA Digital Team
Year
2026
Type
B2B SaaS · AI Workflow · Metadata Governance
Role
AI Product Intern / Product Design

IKEA's content library did not have a visibility problem anymore. It had an accountability problem. Phase 1 could detect missing metadata across thousands of assets, but managers still could not assign ownership and publishers still could not repair tags in batch. I designed a dual-role remediation workbench that turned metadata health from a dashboard metric into an operational workflow.

Metadata gaps were visible, but not actionable

A scan surfaced five critical metadata fields under 65% coverage. The dashboard made the issue visible, but it still could not answer the operational questions: who owns the gap, which assets should be fixed first, and how can publishers repair them without opening files one by one?

Phase 1 could detect metadata gaps across thousands of assets. But managers still had to track down owners by hand, and publishers still had no focused queue for repair. The system exposed the issue without making it executable.

The real gap was not visibility. It was accountability.

Critical Coverage Snapshot

41.0%Negotiated usage rights
44.2%Marketing objective
58.1%Launch timing
59.4%Channel
63.3%Project name

Figures are desensitized for portfolio presentation. The point is not the exact percentage, but that critical fields were below the threshold and no role-based repair flow existed.

Before / AfterVisibility dashboard → remediation workflowBefore: no owner assignment, no batch repair, no confirmation trail. After: assignable queue, batch tagging, exception control, audit log.

The redesign turns metadata health from a status report into an operational surface.

Another dashboard would make the problem easier to see, not easier to fix

The team already had coverage rates, gap counts, and trend lines. What it did not have was a reusable way to detect an issue, decide priority, assign an owner, batch repair the metadata, confirm high-risk fields, and leave an audit trail.

That reframing changed the whole product surface. Instead of optimizing one dashboard, I designed a remediation system: role split, queue logic, AI permission boundaries, and auditability in one connected flow.

Operations LoopDetect → Prioritize → Assign → Repair → Confirm → LogA workflow frame for moving metadata health from analytics into operations.

A batch action is only safe when the data semantics are agreed

Before batch repair could be trusted, we had to align what “missing” actually means across the DAM backend, Confluence specs, and legal interpretation. A null value, an empty string, and an outdated usage-rights value do not carry the same risk.

FieldBackend StateBusiness MeaningRiskRequired Action
Negotiated usage rightsnull / empty / outdatedUsage permission not confirmedHighHuman confirmation required
Marketing objectiveemptyCampaign context missingMediumAI suggestion + publisher review
Launch timingempty / outdatedAsset may be used in the wrong campaign windowMediumPublisher review required
ChannelemptyDistribution context missingMediumAI prefill allowed
Project nameemptyAsset cannot be grouped reliablyLow–MediumFolder inheritance allowed

Managers assign. Publishers repair. The system keeps the record.

I split the workflow into two role-specific views. Managers work from backlog, risk, and owner assignment. Publishers work from a focused repair queue, with folder-level batch tagging and per-asset exclusions.

Manager / Publisher / SystemThree columns, one shared remediation modelManager: detect risk and assign owners. Publisher: batch repair and confirm exceptions. System: update health and keep audit logs.

The same asset reads consistently in both views, but each role sees only the part of the workflow they can act on.

AI permission follows tag risk, not model capability

AI can help recommend likely tag values, but it cannot act as a universal autofill layer. For compliance-required fields, AI can suggest with confidence and source, but it cannot write back directly. For search-critical fields, it can prefill for publisher review. For descriptive fields, it can support faster batch suggestions and review by exception.

P0 / P1 / P2 MatrixCompliance-critical → search-critical → descriptiveP0 suggest only. P1 prefill allowed. P2 batch suggestion by exception.

01

Split by responsibility

Managers need ownership, risk, and assignment. Publishers need a focused repair queue.

02

Tier tags by risk

Not every missing field carries the same business and compliance impact.

03

AI suggests, humans confirm

For P0 fields, wrong write-back is a compliance event, not a cosmetic error.

04

Batch repair needs exceptions

Folder-level tagging only works when special assets can be excluded before applying.

The final workbench connects detection, assignment, repair, and review

The final prototype translated the operating model into three connected surfaces: a manager workbench, a publisher repair queue, and an audit trail. This is where the product moved from a concept deck into something a business owner, frontend engineer, and DAM backend owner could align on together.

Final Design CompositeManager workbench + publisher queue + audit logOne large composite placeholder, later replaceable by desensitized final UI captures.
ManagerHealth overview + backlogCoverage overview, ranking, and remediation status.
PublisherFocused repair queueBatch tagging with asset-level exclusions.
AuditTraceable change historyOwner, field, method, count, and status.

The reusable outcome was not another dashboard

The reusable pattern was metadata remediation as an enterprise workflow: risk-tiered tags, role-based queues, batch repair with exclusions, human confirmation for P0 fields, and audit logs for accountability.

Risk-tiered tags
Role-based queues
Batch repair
Human confirmation
Audit log

6,000+

historical missing-tag issues identified

5

critical fields prioritized

2

role-specific workflows designed

1

reusable remediation pattern

Enterprise AI becomes valuable when suggestions enter a workflow with ownership and review

01

Live DAM write-back

Connect confirmed repair actions to the production DAM with rollback and permission control.

02

Confidence calibration

Use accepted, rejected, and edited suggestions to calibrate thresholds by tag type.

03

Exception learning loop

Treat excluded assets as signals to improve inheritance rules, tag definitions, and AI boundaries.

B2B AI products do not become valuable when they generate more suggestions. They become valuable when suggestions enter a workflow with ownership, permission, confirmation, and review.

Next Project

Content Genie