Help us Map
Help us Map

Sprout Labs › Studies › Which AI can handle your kid’s school apps?

Study · School and activity apps

Which AI can handle your kid’s school apps?

We gave AI assistants the errands a parent sends into ClassDojo, Seesaw, Google Classroom, TeamSnap and two private school portals, and checked every answer. Here is what happened.

Han LeeCo-founder, Sprout · · LinkedIn ↗

TL;DR

  • The leading setup got most of the school errands right. Most free setups did not get in at all.
  • The best setups give a parent back real time: the leader saves about several minutes per round of errands against doing it by hand.
  • Paying helps less than you would think. A free setup sits near the top; a $200 a month one does not.

Why we ran this

The school errands never stop. Is there anything to sign on ClassDojo. What is due on Google Classroom. What time is the field trip bus. Not once a week: a parent checks a couple of times a day, every day, or does not check and forgets, and then it is chaos. Each check is simple. The problem is that they are scattered across every app the kids’ schools and teams happen to use, each with its own login.

So the question we wanted answered is simple. Can an AI agent get the answer out of each of these apps on its own, and get it right, reliably enough that we can build automations on top of it later? Not in a demo. On the real apps, with the real logins, with nobody helping.

What we asked

We wrote down the errands a parent sends, word for word, and gave each one to each AI setup with no link and no hint. The assistant has to find the app, get in, read the right page and come back with the answer. A few examples:

Everything is read only for now. No replies, no RSVPs, no payments. If the assistant changes anything, that errand fails. Benchmarks where the assistant has to act, sign, pay, reply, come next.

How we tested

We started with the apps one family actually has: four real apps seeded with a made-up second grader’s class (ClassDojo, Seesaw, Google Classroom, TeamSnap) and two real private school portals. More join as parents run the bench on theirs. The parent is played by a person who answers the assistant’s questions from a fixed sheet and never hints at the answer.

Each setup is one product on one surface at one price, like Claude on the web on the Pro plan. We tested the price tiers a family actually buys, not the strongest model money can rent, because the result has to reflect what a real consumer gets, not a bespoke setup. That includes the on-device apps, the Mac apps, set up the way a parent would set them up.

Every assistant gets a fair start. Memory is on, because a real parent does not want to repeat the same facts every time. Each one is told who is in the family, which school apps the family uses and the details a parent would share. Then it is on its own.

A big edge for the winners was the harness, not the intelligence. Secure storage of the family’s logins that the assistant reaches for every time, on every surface, is the clearest example: Muse’s fired consistently. ChatGPT had it on some surfaces and not others, and the numbers follow.

Every try is graded by hand against what is actually in the app. A try counts as right only if the answer is correct and came from inside the app. An answer the assistant made up, even a lucky one, does not count.

The board ranks setups by the cautious end of their 95% interval, so one lucky try cannot top it.

What we found

The leading setup leads, with most of its errands right. The next setups follow close behind.

Who got the errands rightShare of graded tries answered correctly. The line is the 95% interval; rows are ordered by its low end, the way the board ranks.

The gap is not between companies. It is between setups. The same product on a free plan and on a paid plan can sit at opposite ends of the board, usually because the free plan cannot sign in to anything.

Getting in is most of the battle. Where a setup got in, it usually read the page correctly. Where it did not, it often said it could not help, or answered from memory. The second one is the dangerous one for a parent.

App by appEvery ranked setup pooled per app, and the best single setup on it.

Time is the number to watch. A round of all the errands takes a parent about a dozen minutes by hand. The leader gives back about several of those minutes. A few setups cost more time than they save, because the parent has to step in.

Minutes given back per roundA round of every errand, against the parent's own time by hand. Solid is measured, lighter is projected for errands not yet run, the line is the 95% interval.

What this means if you are a parent

What is next

This is the first step of the benchmark: get in and read. Next come think (put together what it read across apps), change (sign the form, pay the fee) and coordinate (update two parents’ calendars without a mistake). More apps join as parents run the bench on the ones they use.

Data and method