Where AI Fits — and Where It Doesn't

The transcript service is "AI-assisted," and the word assisted is doing legal, ethical, and architectural work. This lesson draws the line precisely, because every design choice in Module 11 follows from it.

The division of labor

AI assists. Humans determine. In this system the AI does three things well:

• Extraction — turning a scanned transcript into structured coursework rows (Lesson 2).
• Matching suggestions — proposing which courses provide evidence toward which SMR domains, with a rationale (Lesson 3).
• Confidence — attaching a score to every extracted field and every suggested alignment, so the system knows what it doesn't know.

What the AI never does: issue a determination. Whether a candidate has met subject matter requirements is decided by a Commission analyst, every time, for every submission. The AI's output is a proposal pre-populated into the analyst's workbench — never a decision that takes effect on its own.

Fail-closed by design

The governing rule is fail-closed: when the system is unsure, it escalates to a human rather than guessing.

• An extracted field below its confidence threshold is flagged for human correction, not passed along as fact.
• A suggested alignment below threshold is shown as "needs review," not pre-accepted.
• If the AI service is down or errors out, the submission flows to fully manual review — the queue keeps moving, just without pre-population. AI unavailability degrades speed, never correctness.
• And structurally: there is simply no code path from model output to a Determination row. The only writer of determinations is the analyst-facing service method, which requires an authenticated analyst identity. Make the safe behavior the only possible behavior.

Why this split satisfies both quality and governance

It's tempting to read human-in-the-loop as bureaucratic caution. It's actually the correct engineering answer to a probabilistic component:

• Quality. LLMs and OCR are confidently wrong at a nonzero rate. Course titles are ambiguous ("Foundations of Science" — which SMR domain?). A trained analyst catches exactly the errors that matter, and the override telemetry in Lesson 4 measures the AI so it can improve.
• Governance. A state commission must defend every determination — to a candidate who disagrees, to an EPP, potentially in litigation. "A named analyst reviewed the evidence and decided, following methodology version 3.2" is defensible. "The model said so" is not.
• Consistency. The AI applies the versioned methodology identically to every transcript; the human layer catches its failures. Together they beat either alone: pure-manual review drifts between reviewers, pure-AI review fails silently.

The practical consequence for you as the developer: confidence scores, review flags, and override reasons aren't nice-to-haves bolted on later. They're first-class columns in the domain model, present from the first migration. The next three lessons build the pipeline that fills them.