Get data in and out — proposed by an agent, governed by the platform
Data onboarding is where enterprise data projects actually die: mapping a source to a model, handling format drift, writing transformations, fixing the pipeline at 2 a.m. BlueHomer treats it as a first-class governed activity, not a side script.
What agent-driven exchange means
Describe the source and the outcome you want; Navi proposes the mapping, the transformation and the schedule as a set of registered, typed capabilities. The platform decides who may run it, on which data, with what approval, and records every execution. The agent proposes — governance disposes.
Import: from a file or a system to a governed model
- Connect — cloud object storage, relational and document databases, files, applications and streams, all behind one Resource Provider abstraction.
- Extract — structured sources plus documents: PDF, Word and Excel text with metadata captured alongside it.
- Propose a mapping — source-to-target field mapping and type inference suggested by the agent, with the reasoning visible.
- Review and approve — a human confirms the mapping before it becomes a running pipeline.
- Run — on a schedule, on a platform event, or when new files arrive.
- Record — who imported what, proposed by whom, approved by whom, and what changed.
Export: governed, approved, attributable
Getting data out is the operation most likely to be regretted, so it carries the same controls as any privileged action: destination policy, approval workflow where required, masking applied at the field level, and a permanent record of what left, when, and on whose authority.
When the schema drifts
Incoming files change shape without warning. The failure mode to avoid is silent success — a pipeline that keeps running and quietly loads the wrong thing.
- Detect the difference against the expected shape.
- Quarantine the affected records rather than half-loading them.
- Propose a re-mapping, with confidence signals attached.
- Approve — a person decides; the correction is recorded like any other change.
Hand-written pipeline vs. agent-proposed governed pipeline
| Hand-written script | Agent-proposed, governed | |
|---|---|---|
| Time to first load | Days to weeks | Hours — the mapping is proposed, not typed |
| Who can do it | A data engineer | A data owner, with engineering review where policy requires it |
| Permissions | Whatever the service account holds | The acting user’s own permissions, enforced field by field |
| When it breaks | Silent failure, or a 2 a.m. page | Detect, quarantine, propose, approve |
| Audit | Git history, if you are lucky | Every run recorded and attributable |
Claim discipline: BlueHomer’s workflows are designed and, increasingly, generated by AI. Approval gates are part of the feature, not a limitation of it — an agent that can load and transform your data is only safe if every step is a registered capability, checked against policy, reversible and recorded.
Frequently asked questions
Can an AI agent build a data pipeline for me?
Navi can propose one: the source-to-target mapping, the transformations and the schedule, assembled from registered, typed capabilities. It executes under your identity, and privileged steps wait for human approval. It does not receive database credentials or write free-form code against your systems.
What happens when an incoming file’s schema changes?
The difference is detected against the expected shape, affected records are quarantined rather than partially loaded, a re-mapping is proposed with confidence signals, and a person approves it. The correction is recorded like any other governed change.
Can an agent import and export data on my behalf with an audit trail?
Yes. Every import and export is an auditable event: what ran, on what data, proposed by whom, approved by whom, and what the outcome was — captured in the tamper-evident audit trail.
Which sources and formats are supported?
Cloud object storage, relational and document databases, files and streams through a growing connector library, plus document extraction for PDF, Word and Excel. Connector coverage for a specific estate is confirmed during scoping rather than promised in the abstract.