Kosher AI at work: using it for parnassah without compromise

Almost everything written about kosher AI is written about children. That is understandable, but it leaves out the person who will use it most: the adult doing eight hours a day of real work with it — drafting proposals, summarising documents, writing emails, working through code.

The concerns at that scale are genuinely different, and worth setting out on their own.

The volume changes the maths

A child might send twenty prompts in an evening. A professional user sends that before lunch. At a few hundred exchanges a week, anything with a small per-exchange failure rate becomes a near-certainty over a year, and any friction in the tool compounds into something you stop using.

This cuts both ways. It means filtering has to be reliable rather than approximately right. It also means a filter that constantly refuses ordinary business requests is not a safe tool, it is an unused one — and an unused tool sends the person straight back to the unfiltered alternative.

The refusals that drive people away

A badly built filter for this audience refuses things that are plainly fine, and the pattern is predictable:

• A marketing email mentioning wine or a bar, refused for alcohol.
• A medical practice's patient letter, refused for anatomy.
• A legal summary involving a criminal matter, refused for violence.
• A song or poem for a simcha, refused because the model was told to be careful about creative writing.
• A debt-collection letter, refused for aggressive tone.

Each one teaches the user that the tool cannot be relied on for real work. Three or four and they have opened a different tab. A filter has to distinguish between a business context and a problematic one, and that is harder than blocking a keyword list — which is precisely why keyword lists are the wrong approach.

What is actually at stake at work

Two things, and they are different from the home concerns.

What goes in

Business use means pasting in things that are not yours to paste: client lists, financial figures, contracts, someone's medical or personal details. Before adopting any AI tool for work, find out whether your prompts are used for training, how long they are retained, and who can read them. This is a bigger practical exposure for most businesses than anything the model might say back.

What comes back

A general model will produce images of people without any concept of tznius, and will write marketing copy in a register that may be wrong for your audience. For anyone producing material for the frum market, a model with no sense of the community's norms is a quiet liability.

Questions worth asking a vendor

1. Is my business data used to train models? Get this in writing, not in a sales call.
2. How long are prompts retained, and can I delete them?
3. Who inside the vendor can read my conversations?
4. Does the filter distinguish professional context from problematic content, or is it a word list?
5. Can I set policy for a team, and see usage across it?
6. What is the refusal rate on ordinary business writing?

The realistic standard

The goal at work is not a tool that refuses a great deal. It is a tool that refuses the right things and gets out of the way for everything else, because a tool that fails the second test will not be used — and an unused kosher tool protects nobody.