Decide Local or Cloud One Task at a Time

Some AI jobs belong on your own machine and some belong in the cloud, and the wrong call costs you either your privacy or a good answer. This week: a three-question sort, what each side is bad at, and a setup that makes switching take seconds.

Glowing orange laptop and locked file box on a desk as papers drift into blue haze

Most people pick an AI the way they pick a checkout lane: whichever one is open. The cloud assistant is already sitting in a browser tab, so the client contract goes there, and so does the board memo with next quarter's numbers. Or they install a local model, hand it a 40-page report and a hard question, get mush back and decide local AI is a toy. Both mistakes have the same fix, and it happens before you open either one.

// The Lesson

Sort the task before you pick the model.

❝

Before you paste anything, ask what the job touches and what it needs, then keep confidential and routine work on your own machine and send research, long documents and hard reasoning to the cloud.

This isn't a hypothetical risk. In Cisco's 2025 Data Privacy Benchmark, a survey of 2,600 privacy and security professionals in 12 countries, nearly half admitted putting personal employee data or non-public company information into generative AI tools. Emergency rooms solved a version of this problem long ago: a triage nurse sorts every patient at the door, before a doctor sees anyone. Your AI work needs the same nurse.

Ask three questions, in this order

First, would it hurt if this left my machine? Client files, HR matters, health details, contract terms, unreleased financials and anything under NDA stay local. Second, does the job need something a small model doesn't have? Current facts, web research, a document longer than a few pages or a chain of reasoning you'd hand to your sharpest colleague goes to a cloud assistant like ChatGPT, Claude or Gemini. Third, if neither applies, is it small, repetitive and self-contained? Cleaning up meeting notes, rewriting a paragraph in your voice, pulling action items from a page, drafting a routine reply or anything you do on a flight with no connection runs fine locally. The order matters. Privacy comes first because it's the only mistake you can't undo. You can rerun a weak answer. You can't unpaste a contract.

Know what each side is bad at

Local models on a typical laptop have short working memory. One popular local runner defaults to a 4,000-token context window on machines with less than 24GB of graphics memory, about 3,000 words. Feed it a 40-page report and it quietly loses the beginning. It also knows nothing you didn't hand it. Cloud assistants have the opposite problem. Many Gemini models read a million tokens or more, roughly eight novels, but everything you paste leaves your control. On consumer plans that can include training: Anthropic's 2025 terms update lets Claude Free, Pro and Max chats train future models if you allow it, with retention stretched to five years, while work and API accounts are excluded. Check your own switches with our 60-second guide. And cloud needs a connection. When it drops, so does your assistant. More on that tomorrow.

Split the jobs that are both

The hard case is work that's sensitive and difficult at once, like making sense of a messy client dispute. Don't pick a side. Split it. Have the local model do the first pass: strip names, amounts and identifying details, or boil the material down to the underlying question. Read that clean version yourself before it goes anywhere, because local models miss things too. Then send it to the cloud for the heavy reasoning and apply the answer to the real facts on your own machine. If your company pays for a business plan with training turned off and your policy allows it, that's a middle lane. When in doubt, ask whoever owns your data policy, not the chatbot.

Set up both so switching takes seconds

The sort only sticks if switching is effortless. Pin your cloud assistant and the local app you set up in last week's lesson side by side in your dock or taskbar. Write one paragraph of standing instructions (your role, your audience, the format you want back) and save it as a text snippet, so either side gets the same briefing with one paste. Then make a folder called "Local only" for confidential working files. Nothing in that folder goes near a browser tab. The whole setup takes about 15 minutes.

Run the triage for one week

Why do people get this wrong? Two honest reasons. Habit wins, so the open tab gets the job. And they treat local versus cloud as a platform choice instead of a task choice, go all in on one side and live with its blind spots. This week, run the triage for five working days. Every time you reach for AI, jot down the task, the route you picked and whether the result was good enough. It takes two minutes a day, and by Friday you'll have a personal sort list you can follow without thinking. Skip it and the open tab keeps deciding for you.