Do AI companies use my business data?

It depends on the plan more than the vendor - across the major providers, consumer, team and enterprise terms differ on content use, administrative controls and what an administrator can configure. Two people using the same tool can get entirely different answers because they are on different plans, so read what applies to your own account in the official terms.

If I delete a conversation, is the data gone?

Deleting a conversation is not necessarily immediate deletion everywhere, and it is certainly not a remedy for a leaked secret. If a password, access key or connection string was pasted, rotate it - change the password, revoke and reissue the key - rather than deleting the message. That is the standard response to any credential leaving a controlled place.

What is the single most effective privacy step?

Send less. Delete columns the question does not need, replace names with identifiers where names are not needed, take fifty rows instead of five thousand, and upload the relevant page rather than the whole file. It reduces exposure under every possible answer to every other question and costs nothing - which is why it is the first step rather than the last.

Do the vendor's terms cover my obligations to clients?

No - those stand entirely on their own. A confidentiality agreement or a contract clause applies to what you paste exactly as it applies to what you email, and some contracts require prior approval before passing information to an external provider. Check the contract beforehand rather than after, and for anything touching personal data get the position from a lawyer.

Should a business ban AI tools to be safe?

A blanket ban produces the worst outcome rather than zero usage - people use personal accounts instead, giving the same exposure with no control and no knowledge of it. What works is a written three-line rule saying what is allowed, what is not and who to ask when unsure, plus a plan that permits the allowed uses. A team plan is control as well as a price.

Do connectors and add-ons change the privacy picture?

Yes - they introduce another party, so not everything happening in a tool happens at that tool's vendor. A connector someone sent as a link is not the same decision as an official one, and if who is behind it is unclear that is the answer. Some products are also a layer over another provider's model, meaning two organisations rather than one.

AI Tools Privacy: Your Data - Yehonatan Saadia

The answer differs by tool and by plan, so the useful thing is not a verdict but a checklist. Four questions cover it - and one habit reduces exposure under every possible answer, at no cost.

There is no single answer to "what happens to my data", because it differs by tool and above all by plan. So what helps is not a verdict but a checklist - four questions, and one habit that reduces exposure under every possible answer.

Question 1: which plan are you on

The most skipped question, and usually more important than which vendor.

Across all the major providers, terms differ between consumer, team and enterprise plans - on content use, on administrative controls, and on what an administrator can configure.

The practical meaning: two people can use the same tool and get entirely different answers to this question, because they are on different plans.

What to do: read what applies to the account you are using. In the official terms, not in an article - including this one. Terms are updated, and any quotation dates.

Question 2: what is retained and for how long

Two distinctions worth holding:

Whether content is used to improve the product - and this is usually the point business plans treat differently from consumer ones.
How long history is retained, and what happens when you delete a conversation.

What is worth knowing without reading anything: deleting a conversation is not necessarily immediate deletion everywhere, and it is certainly not a remedy for a leaked secret. If a password or access key was pasted, rotate it rather than deleting the message. That is the standard response to any credential leaving a controlled place.

Question 3: who else is in the picture

A point that gets forgotten: not everything happening in a tool happens at that tool's vendor.

Connectors and add-ons introduce another party. One someone sent as a link is not the same decision as an official one - and if who is behind it is unclear, that is the answer.
Tools that wrap another tool. Some products are a layer over another provider's model - meaning two organisations rather than one.

What to ask: who else sees this besides the provider I signed up with?

Question 4: what you owe

And this question is entirely separate from the vendor's terms.

Even where the terms are excellent, your undertakings to your own clients stand on their own:

A confidentiality agreement or a contract clause applies to what you paste exactly as it applies to what you email. Some contracts require prior approval before passing information to an external provider. Check the contract beforehand, not after.
Personal data about people - names, phone numbers, identity numbers, medical or financial information - carries obligations, and the position for your case has to come from a lawyer. Israeli law in this area has been updated in recent years, and how it applies depends on the type of information, the scale and what you undertook. This text describes questions and is not legal advice.

The habit that works under every answer

This is the practical part, and the thing to take from here.

Regardless of how the four questions are answered - data minimisation reduces exposure, and it costs nothing:

Instead of	Do this
Uploading the whole export	Delete columns the question does not need - phones, addresses, ID numbers
Real names	Identifiers, if names are not needed for the analysis
5,000 rows	50 to build the structure, then run it on everything
A whole file	The page the question is about
The data	The logic only, when building a tool

The test before every upload: if this content reached an outside party, would that be a problem? If so - what can be removed without harming the question?

And remember: once the tool is inside your environment - built into the apps or connected to the file store - there is no moment of "I am uploading a file" to prompt the decision. So the decision has to be made in advance.

What almost never gets sent

Passwords, access keys, connection strings. And if one was pasted, rotate it.
Scans of identity documents.
Medical or financial information about identified individuals, unless explicitly established as permitted.
A client's information you undertook confidentiality over, without checking the agreement.

And in a business with employees

Two points that change the picture:

A written rule beats an expectation. Three lines - what is allowed, what is not, who to ask when unsure. Assuming everyone exercises the right judgement is an assumption.
A blanket ban produces the worst outcome. Not zero usage - usage on personal accounts, the same exposure with no control and no knowledge of it. A team plan is control as well as a price.

The checklist

Read the terms for your plan - not for the tool in general.
Establish internal policy if it is a company account.
Reduce before uploading. Every tool, every time.
Check who else is in the picture - connectors, add-ons, layers.
Check client contracts before passing on their information.
Personal data - a lawyer on the specific case.
A leaked secret gets rotated, not deleted.

The detailed version of these questions applied to a specific tool shows what this looks like in practice - but the framework is the same for every tool.

AI Tools and Your Data: The Four Questions to Ask About Any of Them

Key takeaways