DEDUPLICATION AGENT AI-assisted duplicate resolution for enterprise asset records - [In-progress]

Overview

Large enterprise organizations manage hundreds of thousands, even millions, of physical assets across factories, warehouses, plants, and field operations.

Over time, duplicate asset records become unavoidable.

A pump might be registered twice because different technicians entered it separately. An imported spreadsheet may recreate assets that already exist. Different business units may use different naming conventions for the same equipment.

Although duplicates seem harmless at first, they slowly corrupt the entire asset ecosystem.

Instead of one reliable asset history, organizations end up with fragmented maintenance records, duplicate work orders, conflicting inspections, and disconnected integrations with external ERP systems.

Cleaning these duplicates manually is extremely expensive.

Our goal was to design an AI-powered Deduplication Agent that could detect duplicate assets, explain why they were duplicates, and help users merge them safely, without losing trust or damaging enterprise data.

The Challenge

Most customers already knew duplicate assets existed. The real problem wasn't detection.

It was resolution. Users had very little confidence merging records because every merge carried significant operational risk.

A single incorrect merge could:

• Remove maintenance history

• Disconnect work orders

• Break preventive maintenance schedules

• Lose external ERP mappings

• Create compliance issues

  
  

Because of this fear, duplicate records accumulated for years. Instead of fixing the problem, teams simply learned to work around it.

The challenge became:

How might we design an AI system that users trust enough to perform irreversible actions on critical enterprise data?

Understanding Users

We collaborated with industry experts and professionals with deep operational experience to better understand how enterprises handle duplicate asset records in day-to-day workflows.

Despite differences in organizations and processes, the challenges and workflows were remarkably consistent.

Existing workflow

Teams typically:

1. Exported assets into Excel.

2. Compared records manually.

3. Investigated linked work orders.

4. Contacted technicians for confirmation.

5. Deleted one record.

6. Updated references manually.
   

A single duplicate could take several minutes to validate. Across thousands of records, cleanup became a massive operational burden.

Major Pain Points

No confidence in duplicate detection

Even when records looked identical, users wanted evidence before accepting they referred to the same physical asset.

High fear of accidental deletion

Deleting the wrong record could permanently disconnect years of maintenance history.

Relationships mattered more than attributes

Users cared less about which description survived.

They cared about:

• Work Orders

• Maintenance Plans

• Spare Parts

• Inspections

• ERP IDs

• Sensor connections

Those relationships represented years of operational history.

Manual comparison was exhausting

Users constantly switched between multiple screens comparing fields line by line.

This created significant cognitive load.

Design Principles

From our research we established four design principles.

AI should explain, not decide

Recommendations needed to feel transparent.

Users shouldn't wonder why something was flagged.

Humans own irreversible decisions

The AI could recommend.

The user always confirmed.

Relationships are more valuable than fields

Field conflicts are easy.

Broken relationships are expensive.

The interface should prioritize protecting linked records.

Reduce comparison effort

Users shouldn't compare twenty fields manually.

The system should surface meaningful differences automatically.

Designing the Experience

Step 1

Configure the Scan

Instead of hiding configuration inside settings, users begin by selecting:

• Business Object

• Matching Attributes

• Similarity Rules

  
  

This gives complete visibility into how the AI will evaluate duplicates. The goal was making the scan predictable before it even begins.

Step 2

AI Scan

Once configured, the AI scans records asynchronously. Instead of displaying a generic loading spinner, we visualized scan progress. This gave users confidence that the system was actively analyzing records rather than simply waiting.

The scan also summarized:

• Total records analyzed

• Duplicate groups identified

• Confidence distribution

Step 3

Review AI Recommendations

This became the heart of the experience. Rather than presenting a confidence score alone, every duplicate group included explainable reasoning.

For example:

These assets match on: • Serial Number• Manufacturer• Installation Location Differ only in: • Installation Date

Users could immediately understand why the AI made its recommendation. This dramatically reduced the need to inspect every attribute manually.

Each group clearly identified:

• Primary Record

• Duplicate Record

• Matching Evidence

• Differing Fields

Possible actions remained intentionally simple:

• Ignore

• Delete

• Merge

  
  

Nothing happened automatically.

Step 4

Guided Merge Wizard

The merge flow became a multi-step wizard. Instead of overwhelming users with every decision simultaneously, we separated the process into manageable steps.

Resolve Field Conflicts

Whenever records contained conflicting values, users selected which value should survive.

Preserve Relationships

This was the most critical step.

The wizard automatically reattached:

• Work Orders

• Maintenance Schedules

• Attachments

• Asset Relationships

• ERP IDs

• External References

  
  

Instead of hiding this behavior, we surfaced it clearly so users understood exactly what would happen.

Step 5

Final Preview

Before executing the merge, users reviewed the final surviving asset.

This preview showed:

• Final attribute values

• Relationships retained

• Associations transferred

• Duplicate record removal

  
  

Only after reviewing this screen could users confirm the merge. This made the destructive action intentional rather than accidental.