Our iPhone app, Zero Mission Guide, is in the App Store. Until this week, it was mainly a way for a firm's people to carry the firm's guide in their pocket. The latest version adds something anyone can use the day they install it: pick a folder of your own documents, and ask it questions.
Here's what it does, how to get good answers from it, and why we built it for ourselves first.
What it does
Choose a folder, such as one you already keep in iCloud Drive, and the app reads the documents in it: PDF, Word, plain text, Markdown and RTF. Then ask a question the way you'd ask a colleague. "When are expense reports due?" "How much notice do I need for vacation?"
Apple's AI model, built into the phone, answers from those documents only. Each answer names the file it came from, and you can open the passage it used and read it with your own eyes. If the documents don't cover the question, it says so rather than guessing.
A few things are true because of how it's built:
- Nothing leaves your phone. The documents and your questions stay on the device. We never see them.
- No account. You don't sign in to anything to use it.
- It's free.
- It needs an iPhone with Apple Intelligence, an iPhone 15 Pro or newer, on iOS 26 or later, with Apple Intelligence switched on.
These answers come from your files and are labeled that way. Nobody has reviewed them, which is the difference between this and a firm's own guide, more on that below.
How to get good answers from a folder
The AI on a phone is small. It does well with a small, tidy folder and struggles with a big, messy one. You'd give a new hire the same advice:
- Keep one folder on purpose. Don't point it at your whole iCloud Drive. Put in the documents you actually want answers from.
- One topic per document, with a plain title. The file name counts when the app looks for an answer, so "Expense policy" beats "Scan_0042".
- When a policy changes, say so. If a new document replaces an old one, write that at the top ("This replaces the 2024 policy"), or better, remove the old one. Two documents that disagree make for an uncertain answer.
- Use the words people use. If everyone says "PTO" and the document says "vacation," a search for PTO finds nothing. Add a line with the other word.
- Scans don't count. A photo of a page has no text to read. The app tells you which files it couldn't read.
None of this is special to our app. It's the same housekeeping that makes any shared folder useful, and it's most of what makes AI answers about your work trustworthy.
Why we use it ourselves
The honest first reason is testing. A folder anyone can make lets us, and Apple's reviewers, try the app without being invited to a firm.
The bigger reason is measurement. We built test folders, one easy and one harder, for made-up offices with expense rules, time off, closing checklists and a few traps. For each folder we wrote an answer key: the right answer to every question, which file it comes from, and which questions the folder can't answer. Then we run the questions and score every reply against the key.
That turns "seems to work" into a number, and the number has already taught us things:
- Old and new documents in one folder were the hardest case. When the 2026 expense policy replaced the 2024 one, the AI sometimes answered with the old meal limit. We now tell it to prefer the newer document, and the score on those questions went up.
- Small models invent numbers. Asked about a commission split the folder didn't cover, a model would sometimes produce a plausible-looking one anyway. Now every number in an answer has to appear in the documents or the question, or the app says the documents don't cover it.
- Search matters as much as the model. A question that shares one word with the wrong document ("dress code" and an alarm "code") pulled in the wrong passage. Now a longer question has to match on at least two words.
We also run the same questions, with the same documents, through other models that are small enough to run on a phone, and through Claude Haiku as the bar to beat. On the harder folder, Apple's built-in model got about three answers in four right. A few of the small open models did about as well as Haiku. That's how we'll decide whether a different model belongs on the phone: by the score, not by a hunch.
Every one of those lessons applies to a firm's guide too. Answers rest on how the documents are written and arranged, so we score that, fix what goes wrong, and score it again.
What it isn't
A folder of your own documents gets you useful answers. It isn't a firm's reviewed guide, and the app never labels it as one. Nobody checks that the documents are current, nobody adds the answer to the question that came up last week, and nobody notices that the new hire keeps asking about something the folder doesn't cover.
That's the work we do for firms. We organize their process into a guide, label every approved answer, watch what people ask, fix the gaps, and score the guide against questions with known answers, every month. If you try your own folder and like it, that's what it looks like when someone keeps it up for you.