Migration without
assumption
Designing a verification workflow for a real archive migration: establish the facts first, then use AI where interpretation adds value.
Moving a substantial professional archive between cloud repositories creates a deceptively simple operational question: how do I know everything actually arrived?
Missing directories, incomplete transfers, duplicate files, naming collisions, and silent corruption can remain unnoticed when verification depends on manually comparing thousands of assets.
Rather than create an agent for the sake of demonstrating agentic AI, I began with work I actually needed to accomplish.
A real problem first
This workflow is being developed against an active migration, not a simulated portfolio exercise.
The implementation is intentionally documented at the architectural level. Private directory structures, credentials, asset names, access details, and other security-sensitive information are excluded from the public case study.
As testing progresses, this page can be updated with measured outcomes, discrepancy classes, refinements, and what the evidence taught me.
Deterministic Facts.
Interpreted Exceptions.
The Migration Verifier
The workflow is being designed to compare the source archive in Google Drive with its migrated destination in Dropbox using read-only access.
- Compare directory structures
- Reconcile file inventories and counts
- Compare individual and aggregate byte sizes
- Identify zero-byte or incomplete files
- Detect likely duplicates or naming collisions
- Use cryptographic hashes when definitive integrity verification is appropriate
- Classify and prioritize exceptions for human review
Read-only by design.
Verification does not require authority to delete, move, rename, share, or reorganize source or destination assets.
The system's job is to inspect, compare, identify discrepancies, and report. Remediation remains a separate decision.
That separation reduces the consequence of a bad inference: the verifier can surface a problem without silently “fixing” the archive.
Use code for facts.
Use AI for interpretation.
Not every problem needs AI
Deterministic work
File counts, byte sizes, directory paths, and cryptographic hashes are questions with factual answers. Conventional code is the appropriate tool for establishing them.
Interpretive work
Once discrepancies exist, AI can help organize them, recognize patterns, summarize likely causes, prioritize investigation, and make the exception report easier for a human to act upon.
Not every difference
means failure
A discrepancy is evidence to investigate, not a conclusion.
An exception might represent a missing file, an incomplete transfer, a duplicate, a naming collision, or another difference that requires context before action is taken.
The workflow therefore separates detection from interpretation and interpretation from remediation. AI may help explain what a discrepancy could mean, but the underlying comparison remains independently verifiable.
Find the difference first.
Decide what it means second.
Verification without
destructive authority
Capability does not confer authority. The verifier receives only the access necessary to establish what is present, identify discrepancies, and prepare findings for review.
The Archive
after the move
Reconciliation solves only the first problem. Once the archive can be trusted, the next opportunity is retrieval.
A future DAM/Librarian capability could help locate assets, identify versions, relate files to projects, and preserve the history and context behind the work.
That capability is deliberately separate from migration verification. Knowing where something is and deciding what it means are different responsibilities.
What I'm testing
Insight
AI is strongest here after deterministic methods establish a trustworthy factual baseline.
Impact
Verification can reduce the human burden of manually inspecting a large archive while keeping destructive permissions out of the workflow.
Excellence
Success is not “the agent ran.” Success is a reconciled, explainable result with exceptions a human can understand and resolve.
In Development
This is a living case study.
The workflow is being developed and tested against the migration it was designed to support.
Future updates will distinguish planned capability from implemented capability and add measured findings only after they have been observed.
