A decision rule you can actually use

The rule

Take one piece of work. Not your whole job, not your whole week. One task you do, that you could describe in a sentence.

Ask these four questions about it, in this order.

• Sensitivity. Is there anything in this that must not leave your machine?
• Size. Does this need a model bigger than your machine can hold?
• Speed. How quickly must one answer come back, and how many do you need?
• Cost. Which way of paying suits how often you will do this?

Why the order is fixed

Each question can end the rule on its own. A later answer never overrules an earlier one.

Sensitivity comes first because it is the only mistake you cannot buy back. Get the cost question wrong and you lose money. Get the speed question wrong and you wait. Get the sensitivity question wrong and the data has already gone. Nothing you choose afterwards undoes that.

Size comes second because it is a wall, not a preference. If the work needs a model that will not fit on your machine, no amount of patience or money changes that at home.

Speed and cost come last because they are the two you can trade against each other. They matter. They are also the two you can revisit next month.

The order also protects you from yourself. Left alone, people run these questions in the order they feel them, and cost usually comes first, because cost has a number on it. Start there and you pick the lane you fancy, then find reasons afterwards.

A worked example that lands on the cloud

You are a school governor. Once a term, a public inspection report of about ninety pages arrives, and the other governors need a fair summary of it.

Sensitivity: the report is already public. There is nothing here to protect. The rule does not stop.

Size: ninety pages read closely is heavy work, and a weak summary is worse than none. A local model could have a go at it, in chunks, with you stitching the parts together. This one wants a strong model, and the strongest ones do not fit on a laptop.

Speed: three times a year, with an evening to spare. There is no pressure at all.

Cost: you would be buying a machine for three tasks a year. Paying per use fits that shape much better.

The rule says use the cloud. That is not a defeat and it is not a failure of nerve. It is the rule working.

Backed by Ollama: Cloud, Claude model pricing.

A worked example that stays at home

You run a small recruitment agency. Every week you write up notes on twenty interviews, each one about a named person who trusted you with it.

Sensitivity: stop. This must not leave your machine. The rule ends at question one.

You do not need to ask the other three. When the first question answers hard, the rest are decoration.

The other three still help you choose which local model to use, and how patient you can be with it. But the lane is settled, and it was settled in ten seconds.

Sensitive is not the same as not allowed

Ten seconds is the rule working. It is also the rule being run too fast, so read question one again.

It asks what must not leave your machine. It does not ask what feels private. People answer the second one by mistake.

Plenty of work feels private and goes out every day, lawfully and on purpose. Bookkeeping is the plainest case. Take one country's rule as an example. In the United Kingdom, from April 2026, sole traders and landlords above an income level must keep digital records of what they earn and spend. Their software sends updates to the tax office every three months. Those figures leave, and that is the law. Your own country will have its own version. The point is not the rule. The point is that money figures feel as private as anything a small business holds, and they still go out.

What makes it allowed is not that the figures are less private than they look. It is that an agreement covers where they go. The United Kingdom's data regulator puts it plainly. Hand personal data to somebody else to handle for you, and there must be a written contract. Look up the rule where you are.

So question one is really two. Is there a rule that stops this leaving at all? And if there is not, which places may it go to?

Your accounting software is a place you chose, on terms you agreed. A chat tool you signed up for last Tuesday is a different place. The first saying yes does not make the second one yes.

Backed by Making Tax Digital for Income Tax, Contracts and liabilities between controllers and processors: when is a contract needed?.

A harder example, where two answers fight the rule

You are one of six people at a small architecture practice. Every fortnight a planning application goes out. Sixty pages, built from site notes, the council's published guidance and the client's brief.

Sensitivity. Client names run all through it, so your instinct is that this must never leave. Then you check, the way the last section said to. The finished application goes to the council and onto a public register. The brief is covered by an agreement the practice already has. So the answer is not never leaves. It is leaves already, to places we chose.

Notice what that took. Nobody decided the names did not matter. Somebody looked up where they were already allowed to go.

Size. Sixty pages is a lot to hold at once. Ollama sets how much a model can hold in mind from the graphics memory you have. Under 24 GiB, that is about four thousand tokens by default, and a token is a chunk of a word. Raising it needs more memory. Its cloud models start at their largest.

Speed. The application has a date on it, and the work happens the day before. Lesson sixteen answered this. For a job needing the big model, the cloud answer comes back sooner, because the local one does not come back.

Cost. Every fortnight is often, and often is the shape that favours owning. That is the second answer pulling towards home. It arrives last, because money cannot buy you through a wall.

The rule lands on the cloud, over the objections of the two questions people feel hardest. That is what writing the order down is for.

Backed by Ollama: Context length, Ollama: Cloud.

When the rule fights itself

Sometimes question one says local and question two says cloud. The work may not go out, and it is also too big for your machine.

This is common and there is no clever trick. You have three honest options.

• Use a smaller local model and accept a rougher answer. Often this is fine, and you should test it before assuming it is not.
• Take the sensitive parts out by hand first, then send the rest. Be honest with yourself about how well that worked.
• Do not use AI for this one. This is a real answer and it is sometimes the right one.

Which of the three is yours

Those three options look equal on the page. They are not, and there is a way to tell them apart.

Try the smaller model first, on a real job rather than a made-up one. Then write down what was wrong with it. Not as good is not a finding. It invented a date is.

If you are stripping the sensitive parts out, check what is left. Could somebody who knows your field put the names back from the rest? A town, a date and a job title often manage it.

And if the answer is not to use AI here, write that down with the date, and a line on what would change your mind. Then it is a decision rather than a wall.