Reconciling duplicate records

Task boundary

Reconciling duplicate records is assessed as a specific task and workflow, not as a general endorsement of AI for the surrounding job. A bounded agent may compare a preselected queue but merges require preview and approval.

Assumptions

• The agent receives a preselected queue and cannot search or merge outside it.
• A deterministic comparison identifies exact duplicates before AI handles ambiguous pairs.
• Every merge has a complete preview, human approval, and tested rollback.

Decision

• Disposition: Proceed with controls
• Pattern: Bounded agent
• Inherent risk: high

Rules that drive the decision

• Ambiguous comparison may require adaptive steps, but write access makes permission and stop boundaries mandatory.
• Irreversible or unreviewed merging would require redesign; bounded reversible merging may proceed with controls.

Why alternatives were rejected

• Fully autonomous merging can irreversibly lose information.
• A direct prompt cannot inspect or modify records safely.

Controls

• Limit records, fields, merge direction, run duration, and total merges outside the model.
• Preserve both source records and show every retained, overwritten, and conflicting value.
• Stop on uncertain identity, concurrent edits, missing history, permission expansion, or rollback failure.

Acceptance threshold

These are example starting thresholds for a bounded pilot of this workflow, not universal benchmarks. For reconciling duplicate records, the accountable owner should make them stricter when the task, consequence, or policy requires it.

• No record outside the queue is read or changed, and no ambiguous pair merges without exact approval.
• Rollback restores both records for every failure injection, and repeated execution cannot duplicate a completed merge.

Reassess when

• The queue source, identity rules, writable fields, merge volume, or rollback mechanism changes.
• The workflow gains deletion rights or permission to resolve identity without approval.