Can You Trust AI With Your School Schedule? How We Designed AI Scan to Earn It

“AI puts your timetable in automatically” is an easy sentence to write and a dangerous feature to build. A planner is a promise you make to yourself: if it’s in here, it’s true. One hallucinated exam date — one Physics class on the wrong weekday — and the whole app becomes something you double-check instead of something you rely on.

So when we built AI Scan, the design question was never “how much can the AI do?” It was: what does the AI have to prove before it’s allowed to touch your planner? Here’s what we landed on.

Principle 1: The model proposes, you dispose

No scanned item — not one — is written to your planner without passing through a review screen you can edit. That’s not a settings toggle; it’s the architecture. The AI’s output lands in a staging area, and the only path from there to your timetable runs through your thumb.

This costs us a little magic. The demo where a photo instantly becomes a planner looks better on stage. But “instant” is only impressive until the first wrong room number, and we’d rather be trustworthy than theatrical.

Principle 2: Uncertainty must be visible

Models are confident about clean print and less confident about handwriting, glare and cramped grids. AI Scan keeps that confidence — every row below a threshold is flagged “needs review” and starts unchecked. It’s still there, still one tap to include — but it can’t ride into your planner on autopilot.

The flip side: we never hide uncertain rows entirely. Deleting the model’s guess is your call, not ours.

Principle 3: Fail loudly, retry safely

Real networks fail mid-import. When that happens, AI Scan tells you exactly what happened — “9 imported · 2 need attention” — and keeps the failed rows in a draft you can fix and retry.

Two guarantees under the hood make retrying boring (the good kind of boring):

• Nothing imports twice. Every row carries a stable identity from the moment it’s scanned, so a retry updates rather than duplicates — even if the first attempt half-succeeded.
• Nothing conflicts silently. A scanned class that overlaps one already on your schedule is flagged as a conflict, not stacked on top of it.

Principle 4: The document is the truth

AI Scan’s job is to read what the paper says — not to improve it. Two examples of what that means in practice:

Merged periods stay splittable. Back-to-back identical periods merge into one class (that’s almost always what schools mean), but the original boundaries are preserved — one tap splits them back exactly as printed. The scan never destroys information you might need.

Ambiguous dates never guess. “Due Friday” resolves against your real academic calendar. A date the model couldn’t read shows a loud pick a date chip instead of a silent best guess. A wrong deadline is worse than no deadline.

Principle 5: Your photo is not a product

The economics are boring by design: your photo is downscaled on-device, processed to extract the schedule, and not stored after extraction. There’s no training on your homework sheets, no photo library building up server-side. Scans are rate-limited and capped per account — which is also why the feature can stay generous without ads or data games.

Why this matters beyond one feature

Every planner app will bolt on AI this year. The differentiating question won’t be who has it — it’ll be whose AI you can leave alone with your schedule.

Our answer: the one that shows its work, flags its doubts, fails in the open, and asks before it writes. AI Scan is available now in the latest StudyWillow update on iPhone and Mac — the feature page has the full tour, and Siri support follows the same confirm-first rules.