Lab-to-AI: From Raw Zone to SageMaker and HealthOmics

The science team’s request arrived as four words: “AI on our data.” Behind it stood a platform team with a validated raw zone, a governance model built for audits, and no plan for what happens when a model wants to read everything.

This is the series finale: the raw zone you built and defended becomes the substrate for AI/ML—but only if governance, access, and cost controls arrive before the first model runs. This post is the readiness checklist for that day.

The staging path: raw to training-ready

Data goes AI-ward in stages, and each stage has a different contract:

• Nothing trains on raw/ directly—models read curated tables with recorded lineage, or you lose the audit thread at the worst moment.
• Consent/PHI/PII flags live in catalog metadata; extracts are filtered by policy, not by a scientist’s memory. Macie or equivalent scans for what metadata missed.
• Training sets are versioned artifacts—a model result must be reproducible from a named dataset, the way a run result traces to a manifest.

Where the work runs

• SageMaker Unified Studio as the governed front door: data prep, catalog, and ML apps in one environment with projects and access control—so “the AI environment” isn’t a snowflake account someone made in 2025.
• Catalog governance via DataZone-style domains: publish curated data with owners, expiry, and approval flows—the attribution discipline, applied to analytics.
• HealthOmics when the workload is genuinely genomic: managed sequence storage and workflow runs beat hand-rolled pipelines for cost and reproducibility—evaluate before building.
• Self-built training (your Batch estate) for everything that isn’t a fit—managed isn’t mandatory, it’s a default to justify against.

Bedrock and assistants: the guardrails that matter

• No raw PHI/consented identifiers into prompts—the filtering happens in the data layer, not in a prompt template someone edits later.
• Knowledge bases read curated, consent-cleared tables—never the raw bucket “just this once.”
• Access to Bedrock/AgentCore is a scoped role per project, reviewed like any boundary identity—not a shared API key.
• Agent actions against data platforms are read-only until a change-control story exists (validated delivery applies to agentic changes too).

FinOps for AI: the cost controls

Nearly every team now manages AI spend; the ones hurt worst are the ones that started after the bill:

• Per-project tagging enforced at role level (only a project role can launch project resources); Cost Categories roll projects up.
• Budgets with alarms per research group—surprise is the failure mode, size is negotiable.
• Model-choice discipline: start with the cheapest model that passes eval; record the eval, not the vibes.
• Weekly review of idle endpoints, notebooks, and feature groups in the same cost hygiene pass as everything else.

Before the first model runs

1. Named dataset versions with lineage, filtered by policy.
2. Scoped roles, catalog approvals, no raw-bucket shortcuts.
3. Budget alarms live, tagged at the role.
4. A written answer to “can this model’s output be traced back to runs?"—because that question is coming.

Series map

If you only do one thing: publish the rule “models read curated, humans approve the extract” before the first AI project starts—retro-fitting governance onto a running model is ten times the work, and it’s the one corner of this series where undo is not an option.