[PowerAI]AI Office Work on Your Phone

1. Summary

This piece is done entirely on a phone — no computer touched at any point. It is the first article in the series that doesn't require a computer, so complete beginners can start reading right here.

Three awkward situations that find you when you're out and about: your boss demands meeting minutes in the group chat, a client is waiting for a firm answer, and approval requests are piling up with nobody processing them. This article walks you through building a "commute AI workbench" with one phone, so the three most common jobs — minutes, email, approvals — can each run end to end within a single subway ride. The aim is not to do heavy lifting on the phone; it is to keep delivering on the days you left the laptop at home.

2. When This Helps / Prerequisites

  • You: a commuter, a field-role worker, or a manager who is often away from the computer and runs work from the phone alone.
  • Scenario: the boss suddenly wants minutes in the group chat; a client is waiting for a reply; a stack of approvals sits in your phone expiring whether you look or not.
  • Device: one smartphone with internet access (Android or iPhone, either works), battery above half.
  • Apps: 2-3 AI assistant apps installed (choose in Step 1 below) + your company's Feishu or DingTalk + any email app.
  • Not needed: no programming tools, no paid memberships, and definitely no laptop — the only thing within reach all day is the phone in your pocket.
  • Mindset: the phone is not a shrunken computer; it is an "on-person input + review terminal." Today's goal: let it catch the urgent work on the days there's no computer.

The first 53 articles in this series all assumed you were sitting at a computer, and mobile had zero coverage until now — this piece is the makeup lesson: no computer required, and you can follow along on the subway.

3. The Steps

Step 1: Lay Out an "AI Workbench" on Your Phone — Picking and Dividing the Model Apps

Phone screens are small and typing is slow, so don't expect one app to do it all — assign roles to two or three apps first. The mainstream general-purpose AI assistants in China are as follows (this article leans on the first three):

  • Doubao: the smoothest voice conversation — hold to talk, and it talks back. Built for "mouth instead of hands" while on the move.
  • Kimi: excels at long documents — throw it a dozens-of-pages PDF or contract, and it highlights the key points first.
  • Qwen: backed by Alibaba's office ecosystem, with document and minutes features all in one place — a natural pick as your office "all-in-one."
  • Tencent Yuanbao, Zhipu Qingyan (GLM), MiniMax Hailuo: same-tier alternatives; use whichever you already prefer.

(ChatGPT / Claude / Gemini / Grok mobile apps are the international counterparts of the assistants above; the workflows in this article present the Chinese apps as they actually are, so those are the ones we follow.)

🌍 A note for readers outside mainland China: Feishu's international edition is Lark (larksuite.com), and on the assistant side the one-to-one matches are the Teams mobile app and the ChatGPT / Claude / Gemini / Grok mobile apps. Features correspond one to one, but the interfaces differ — this article follows the China editions.

Divide the labor by "each plays to its strengths." One reference setup — fine-tune it to your own habits; don't agonize:

Scenario Who to use Why
You'd rather talk than type Doubao Voice input recognizes accurately; chatting feels like a phone call
A long document needs its key points Kimi Long text is its home turf
Documents and minutes end to end Qwen Meshes most tightly with office scenarios

Next, drill the two basic skills every AI app ships with — they are the two legs phone-based work stands on:

  1. Voice input = hold the microphone and talk; the AI turns your words into text before answering. It converts "slow typing" into "fast talking" — speaking is usually more than three times faster than typing.
  2. Photo OCR = turning the text in photographed paper documents, slides, and screenshots into editable text. Paper forms, whiteboards, business cards — one snapshot and the AI can process it.

Three high-frequency photo-OCR uses out in the field — try each once today:

On site How to shoot What to say along with it
Business card Lay it flat, good light, fill the frame "Turn this business card into one contact entry: name, company, title, phone."
Paper expense form / waybill Shoot straight-on and clear, no shadows "Copy out the amounts, dates, and tracking numbers item by item, as a table."
Whiteboard / projected slides Stand square, capture it all, text sharp "Turn the whiteboard content into meeting takeaways, listed point by point."

Key points

  • Feature entries shift between app versions — always defer to the current version on your phone: if you can't find a feature, update to the latest version first, then search in-app or dig through the "+" menu.
  • The free tier covers everything today; when an "upgrade to premium" prompt pops up, skip it — it blocks none of this article's tasks.
  • The first time you use voice input, a "microphone permission" prompt appears — tap Allow. To keep notifications from interrupting, set everything except your AI apps to do-not-disturb.
  • Do one piece of setup tonight: drag your main AI app and Feishu/DingTalk onto the same first home screen, so tomorrow morning everything is one screen away.

Step 2: Feishu / DingTalk on Mobile — Recording → Minutes, AI Speed-Read of Group Messages, Approvals

Your company most likely runs on Feishu or DingTalk — their mobile apps are an underestimated AI office platform in themselves. Put all three of these to work:

  1. Recording → minutes: Feishu calls it Feishu Minutes (妙记); DingTalk calls it DingTalk AI Notes (闪记/AI 听记) — in newer versions the entry is usually labeled "AI 听记," found under "+" → "More" in a chat window. Turn on recording during meetings, one-on-ones, or client visits, and the app transcribes automatically, splits the text by speaker, then generates a summary and a to-do list. You can record, read, and forward it all from the phone. Using Feishu Minutes as the example, five steps on the phone:

    1. Open Feishu → "Minutes" (or enter from the Calendar / group-chat entry points);
    2. Tap "New / Record" and pick the transcription language;
    3. Start recording when the meeting starts; a phone in your pocket doesn't hurt the audio (a connected Bluetooth headset mic works too);
    4. Wait a few minutes after it ends, and the app produces "full transcript + summary + to-dos";
    5. Quickly proofread proper nouns on the phone, then forward to the group or send to your boss with one tap.
  2. AI speed-read of group messages: You open the app in the morning to 999+ unread? Don't climb the thread message by message. Use the group's "AI summary / message speed-read" feature, and within seconds you'll have "who said what, what was decided, and who owes which action."

  3. AI-assisted approvals: Before approving leave or expenses, toss the request's key points to the AI for a check — do the amounts line up, is the reason clear, do any dates conflict — and only then tap "Approve."

Fallback when a group chat has no AI-summary feature: long-press messages to multi-select → combined-forward into an AI app → send one line: "Organize these group messages into 3 bullets: conclusions, to-dos, and items that @-mention me."

When an approval looks off, screenshot it for the AI with one line:

This is an expense approval form. Please check it for me: are the amounts
consistent throughout, is the reason stated clearly, do any dates conflict —
and list anything you find suspicious.

Key points

  • For a face-to-face meeting you want transcribed, tell the other party out loud before you start recording — that's both courtesy and compliance (in confidential settings, don't record at all; see A5).
  • On the phone, keep approvals to "judging": approve, reject, or write a comment are all fine. Changing the workflow, adding approval nodes, or editing the form — save that for the computer.
  • Before screenshotting anything to an AI, mask sensitive fields such as invoice numbers, company letterheads, and other people's personal details; if your company explicitly bans sending documents to external AI, verify manually only.
  • Where the Minutes / AI Notes entry sits, and how much free quota you get, varies by version and company plan — defer to the current version on your phone; if you truly can't find it, asking your company admin is the fastest way.

Step 3: Voice-Drafted Email Replies — Say the Points, AI Drafts, You Copy and Send

The slowest part of replying to email on a phone is typing. Change the flow: you only say the key points, the AI produces the draft, and your fingers do nothing but copy-paste. Three moves:

  1. Open Doubao (or your main AI app), hold the voice button, and speak the prompt below (a prompt = what you send to the AI; A3 teaches you to write them sharper) together with your points — or type it out:
Help me turn the points below into a formal reply email, polite and
professional in tone:
1. Thank the recipient for last week's visit;
2. The quotation will go out by this Friday;
3. Suggest an in-person meeting next Wednesday afternoon for details,
   and ask them to confirm the time.
  1. The AI returns a complete draft with salutation, body, and sign-off. Read it through once, and correct the salutation, names, and dates.
  2. Select all and copy, switch to your email app and paste, verify the recipients, tap send.

Key points

  • This "say the points → get a draft" template is the same one used on the desktop (see B29, "Tame Your Inbox with AI (Outlook Email AI in Practice)," for the reply-style presets) — write it once, use it on both ends.
  • Reading the draft through before sending is mandatory. Small screens let typos and a missing salutation slip past far more easily; once an email leaves, it's there in black and white.
  • Not happy and want changes? Revise by voice too: "make the second paragraph a bit more courteous," "add a line at the end saying you look forward to their reply." Having the AI edit beats deleting and retyping by a mile.
  • Someone else's long email you can't finish? Screenshot it to the AI: "Summarize in three sentences: what the sender wants, the deadline, and what I need to do."

Optional bonus drill — run the harder scenario below on your own:

A client pushes in the group chat: "Where are the acceptance criteria we agreed on last week?" But you have no document on hand — only scattered verbal conclusions.

Reference play: hold the voice button and tell the AI: "The client is pushing for the acceptance criteria. Let me say three points out loud: scope covers modules A and B only; issues get reported within two weeks after launch; acceptance follows the test report. Turn this into three polite bullets, and give the client a firm time first: the formal version goes out by end of day." — Giving the other side "when there will be an answer" is more professional than squeezing out an imperfect one.

Step 4: The Commute Combo, Logged Live — 25 Minutes on the Subway, Three Tasks Closed Out

Below is a run I actually did on a rush-hour subway, served exactly as it happened (names and companies are placeholders):

08:12 Squeezed onto the train, headphones on. A message from my boss pops up in the group: "Send me the minutes of yesterday's meeting with Hengda before 10." The old me would have stewed over that one line for an entire stop.

08:14 Open Feishu, find yesterday's meeting recording in Minutes, tap "Generate minutes." A 40-minute recording becomes a one-page summary: three points of consensus, two to-dos, each with an owner. One quick scan, fix one transcribed-wrong name, forward into the group. Task one, done.

08:21 Scrolling on: the Hengda client has left a message asking about the quotation. Switch to Doubao, hold and speak — 30 seconds to cover the points: "Thank them for the visit; quotation by Friday; suggest meeting next Wednesday, keep the tone courteous." Copy the draft, paste into email, read it through, send. Task two, done.

08:27 Two approvals still sitting there. Switch to DingTalk: one leave request, reason and dates all complete — approve. One expense claim — screenshot sent to Doubao for a check, and the amount doesn't match the invoice — reject with a clearly written comment, noting "details will be re-verified on the computer." Task three, done.

08:32 While I'm at it, open another project group stuck at 999+ and tap "AI speed-read" — a thread that ten minutes of scrolling couldn't finish, key points sorted in a minute or two: two things to deliver today, neither really mine. Peace of mind.

08:34 The train crosses the river and the signal drops. No panic: the PDF report I need to read was downloaded to the phone last night before bed — it reads just the same in the tunnel.

08:37 My stop. Three tasks closed out: minutes sent, email replied, approvals cleared. The close reading and rewriting of that 60-page report? That's the computer's job — it can wait until I'm at my desk.

Note the rhythm: in those 25 minutes, the heavy lifting — transcription, drafting, checking amounts — all went to the AI. I did only the three things a human should do: speak, review, and tap confirm. This isn't "grinding" on the subway; it's spending fragmented time on the steps only you can make the call on.

Play the tape back, and the three tasks map exactly onto three levers:

Time Task done Lever used
08:14 Meeting recording → minutes Feishu Minutes (transcription + summary)
08:21 Voice points → reply email Doubao voice input + drafting
08:27 Expense form check AI review + human sign-off
08:32 999+ group messages DingTalk/Feishu AI speed-read

4. FAQ / Troubleshooting

Situation What to do
No signal on the subway / signal keeps dropping? Download documents ahead of time and screenshot key pages into your photo album beforehand (screenshots stand in for live reading) — work continues straight through the tunnel.
Can you trust conclusions an AI gives you on the phone? Anything touching numbers, commitments, or external messaging: verify against the original on your computer before sending. The phone screen is small, you can't lay the source side by side, and AI hallucination (the AI confidently fabricating plausible-sounding but wrong information) is even harder to spot on a phone.
The meeting content is confidential — can I still transcribe it? No recording/transcription and no photo OCR in confidential settings — sending audio to transcription is effectively feeding the meeting into a public AI, and photo OCR is the same; the red lines match A5, "AI Office Safety and Privacy Red Lines." When unsure, neither record nor shoot.
Want to change an approval workflow or add a node? On mobile, stick to "judging" only; workflow and form changes always go back to the computer.
Battery / data won't last until end of day? Lower the brightness and keep only one AI app open; if data is tight, switch to downloaded content and don't push large files online.
Can't find the Minutes / speed-read entry in the app? Feature entries shift between versions — defer to the current version on your phone: update to the latest first, then search the feature name in settings.
Talking on the subway with people all around — awkward? Headphones on, short sentences, low volume; if it's truly awkward, type the points instead, or do the voice part after you get off — don't dictate a long email into a packed carriage just to save effort.
The boss wants the minutes as a Word file? Create a doc in mobile WPS/Tencent Docs, paste in the AI minutes, and send it — good enough in a pinch. A properly formatted formal version waits for the computer (exactly the "short output" vs. "production" line in the boundary list below).
Too many voice-transcription errors? In noisy surroundings, lean in and slow down; replace proper nouns by hand afterward — don't expect one-pass perfection.
Afraid of mis-sending or mis-tapping "Approve"? One extra confirmation step on mobile never hurts: read before sending, and re-check the amounts and dates before approving.

5. Recap & Going Further

The principle, recapped: the three levers of mobile AI

  1. Voice input: turns "slow typing" into "fast talking" — the first lever of phone-based work; this article's email and approval comments both run on it.
  2. Photo OCR: turns paper and screenshots into text the AI can process — paper forms, whiteboards, someone else's screen; photograph it and the AI can get to work.
  3. The system share sheet: the bare-minimum "workflow" — the share sheet is the row of app icons that pops up when you tap "Share" in any app. Fling a screenshot or a passage of text straight to another app for processing, no copy-paste needed — that's apps being "wired into one flow."

One line to remember the division of labor: the phone is bad at "production" and built for "input + review + short output." From that, draw a boundary list:

Keep on the phone (input · review · short output) Keep on the computer (production · heavy lifting)
Recording → minutes, forwarding summaries Close reading and rewriting of long reports
Voice-drafted email replies, group-message replies Long-form writing, formal document layout
Approving requests, writing short comments Changing approval workflows, editing forms
Photo OCR of cards / forms / screenshots Building spreadsheets, data processing
AI speed-read of group messages, reading summaries Final verification of important conclusions

Week-one practice checklist — add just one drill per commute; after a week you'll work the phone like a pro:

  • Monday: hold one conversation with an AI using voice input only, and feel "talking instead of typing."
  • Tuesday: photograph one paper form or business card and run a full photo-OCR pass.
  • Wednesday: record one meeting with Minutes / AI Notes and check the generated result right on the phone.
  • Thursday: use the "say the points → get a draft" template to answer one non-urgent email fast.
  • Friday: replay this article's Step 4 combo in full — even if all three tasks are small ones, closing the loop means you've graduated.

Going further (take what you need)

  • B29 quick email-reply presets: the email template in this piece is the same one as on the desktop — build out the full presets over there and both ends share them.
  • B30 Tencent Docs / Feishu AI: the next level up from this piece — mobile is just the entry point; cloud collaboration in Tencent Docs / Feishu AI is the full form.
  • Phone Shortcuts / automation: turn "photograph → recognize → file" into a one-tap flow — the mobile version of the T5 automation idea; no dedicated article in the series yet, so build it yourself with iOS Shortcuts or Android automation tools.
  • Revisit A7, Day One: if you haven't yet run your very first AI task end to end, start there, then come back.

6. CTA

Tonight before bed, spend 5 minutes: drag your two AI apps and Feishu/DingTalk onto the phone's first home screen and confirm you can log in to both — tomorrow morning's subway ride is the first test run of your "commute AI workbench."

Don't overreach: tomorrow, get just one thing running end to end (we recommend starting with "AI speed-read of group messages" — zero risk), and add one more the day after. A week from now, look back at the three awkward situations at the top of this piece, and you'll find they no longer pose a threat.