What the AI can and cannot see

Sunny is built into Clear Day. To help with your day it reads some of your school's records. Here is what it uses, what it can't reach, and where the edges actually are.

It reads through a fixed set of skills

Sunny doesn't have a general key to your database. It has a list of named skills, each one a specific read or a specific proposed action, and it can only ever use one of those. If a skill doesn't exist, there is no answer it can give you: it will say so rather than guess.

The skills that read cover:

• Your roster: children, families, classrooms, enrollment status
• Today: attendance, who's expected, who's out, staffing and ratios
• Classroom records: daily reports, observations, lesson plans, curriculum coverage
• Paperwork: which policies and forms are outstanding, immunization compliance status, inspection readiness
• Money at the summary level: billing health, billing anomalies. Not card numbers, not bank details
• Enrollment: pipeline, funnel, occupancy forecast, churn, tours (Grow plan)
• Filed paper documents: the fields read off forms you've scanned and approved

Ask about your fall roster and it reads the fall roster. That's what the receipt shows you afterwards. See work receipts.

What about messages?

This one is worth being exact about, because "the AI reads our messages" is the fear.

When you ask Sunny a question in the tray, it has no skill that opens a message thread. The closest thing it can see is activity: which families haven't been in touch in the last few weeks, and how many days since the last touch. Names and dates, not what anyone wrote.

Two features do read message text, and both are ones you asked for:

• The morning brief triages overnight messages so it can tell you which thread needs you. It carries a short excerpt of an urgent message into the brief.
• A reply draft reads the family's message you're replying to, which is the only way to draft a reply to it.

In a staff channel, @sunny reads the thread it was mentioned in so it can answer the question. It is never mentionable in a thread a family can see: see below.

Your school, and only your school

Every skill is scoped to the school you're signed in to. A school-scoped skill given no school, or the wrong one, refuses to run rather than quietly returning nothing: that refusal is enforced in the same place for every skill, not re-implemented per feature. Underneath, the database enforces the same boundary on its own rows.

There is no skill on the staff surface that reads across schools. If you run more than one site, each one answers for itself.

Your role, and the room you're in

Every skill carries a role floor. The school-wide operational reads (today's attendance across the school, staffing, billing health, paperwork completeness, the enrollment pipeline) are director and admin only. A teacher asking for one gets a refusal, not a quietly smaller answer.

In a staff channel there's a second cap: the answer is limited by the least-privileged person in the room. A director asking in a channel that includes teachers gets a teacher-level answer, because the channel is the audience, not just the asker.

What it cannot reach

• Other schools' records. Not by asking nicely, not by phrasing it differently.
• Anything outside Clear Day: your personal email, your phone, your bank login, files on your device.
• Full card or bank numbers. Payment details live with our payment processor and never enter Clear Day's AI path.
• Money movement. Charging a card, issuing a credit, changing a rate: these are classified in the product and switched off for the AI entirely. There is no draft you could approve that moves a family's money.
• Family-facing threads. @sunny cannot be summoned into a conversation with a parent in it. If a parent is a participant, the request is refused before Sunny reads a single word of the thread: it isn't a rule in a prompt, it's the absence of a path. See ask @sunny in staff messaging.

If you connect Clear Day to your own Claude account

Some schools add Clear Day as a connector inside their own Claude account. That sends results into an account whose data handling we neither control nor can check, so a narrower set of skills is published there.

Any skill whose output can carry health or medical detail, injury or illness reports, or individually identifiable detail about a child is refused on that connector: the product will not start if one is added by mistake. That covers a child's profile, a day's daily report, incident summaries and detail, immunization compliance, the roster listing, family memory notes, and anything read off a scanned form.

It isn't a blanket ban on the underlying data. Those skills stay available inside Clear Day itself, on our own inference path, under our own agreement. The line: individual-level health or incident data is always off the connector; an aggregate that can't be traced back to a child is fine. A per-classroom immunization breakdown in a room of eight is identifying, so it stays off.

Some of it runs on your phone

On recent iPhones with Apple Intelligence, the first pass of five features runs entirely on the device: dictating a classroom event, the recap in voice intake, the quick-fill that turns a typed note into the fields of an incident report or a planned-absence form, and the caption suggestions when you log a photo for a room. The app carries its own copy of those instructions, so the text is handled where you typed it. It falls back to our server only when the device can't run it, and the caption suggestions don't appear at all on a phone that can't.

Voice recordings never leave your phone at all: the app transcribes them on the device and sends only the resulting words. There is no recording anywhere for us to hold.

When the on-device path isn't available, the request goes to our server and on to Claude. Every one of those calls goes through a single piece of code with no setting for handing your data over to train a model (there is no field for it to pass), and a build check stops any other part of Clear Day from reaching Claude directly.

We do not use your school's data to train models for anyone else.

Images

If you attach a photo to an ask, or photograph a form, that image goes to the model: that's the whole point of both features. Photos your teachers post to a child's day are a different path: the family-facing summaries are written from your logged text, and the photo stream itself is never sent to a model.

Photo captions are the clearest case. When a teacher logs a photo for a whole room, the app can offer three caption suggestions (written from the selected children's first names and the teacher's own note, never from the picture). Clear Day does not analyse the image. On the phones that can run it, the suggestion is generated on the device, so neither the photo nor the note leaves it; on the phones that can't, the suggestion simply isn't offered.

Where answers cite their sources

Answers carry tappable chips naming what they drew on (a child, a classroom, a scanned document), and tapping one opens that record.

The honest limit: a chip only appears where the answer names a real record we can link you to. Most operational answers are aggregates ("26 confirmed, 5 unopened") with no single record behind them, so they carry no chip rather than a caption dressed up to look like one. A citation you can't tap isn't a citation.

When our support team is in your account

If Clear Day support is impersonating a staff account to reproduce a problem, they can drive Sunny, but every entry it writes records that it happened under impersonation, and they cannot approve anything. Approving, rejecting, and redrafting are all blocked on an impersonated session. An operator can propose; only your team can decide.

It drafts. It doesn't send.

This is the part to hold onto. Sunny never sends a message, files a document, or changes a record on its own. Everything it proposes is a draft, a person on your team approves it, and the approval is recorded with their name. See drafts wait for your team.

Two things elsewhere in Clear Day are exceptions worth knowing about, because they're scheduled communications rather than actions an assistant decided to take. Only one of the two is something anyone switches on:

• The end-of-day report email goes to a child's parents at the time your school sets, with a written recap of the day built from what your teachers logged.
• The weekly family digest goes out midday Sunday, again written from the week's logged records.

Neither asks for a per-message approval (if that matters to you, they're the two to look at). They don't have the same controls, so it's worth being exact:

• The end-of-day report is a school setting, and it also honours each parent's own notification preferences, right down to muting it for one child. A parent who turns it off stops receiving it.
• The weekly digest has no such switch today. It goes on Sunday to every parent linked to a family with an enrolled child, and there is no per-parent or per-school way to opt out of it. The only thing that stops it is a very quiet week: when a family's daily reports, moments and photos for the week (added up across every enrolled child, not any one child alone) come to fewer than three between them, Clear Day files a short "quiet week" note in the digest history and sends neither the push nor the email.

See daily reports and the weekly digest.

Have a privacy or compliance question we didn't cover? Email us and we'll walk you through it.