Don’t paste anything into a consumer chatbot that you wouldn’t email to an outside vendor, because that is effectively what you’re doing. Check your employer’s AI policy first and use whatever tool the company has approved, since business AI plans generally commit not to train on your data and give admins control over retention. For your own drafts and notes, a model that runs on your device never sends the text anywhere, though it doesn’t override your company’s rules about what can sit on a personal phone.
What counts as confidential at work? #
More than most people assume. If it isn’t already public, treat it as confidential until the policy says otherwise.
| Usually confidential | Usually fine with any AI |
|---|---|
| Customer names, emails, account details | A public press release you’re summarizing |
| Source code and internal tools | Generic coding questions with no company code |
| Financials, forecasts, pricing | Published annual report figures |
| Contracts, legal matters, HR issues | A question about a law in general |
| Unreleased product plans | Brainstorming with no real names or numbers |
| Anything under an NDA | Rewriting your own public bio |
Regulated data carries extra rules on top. If you work with client files, patient records or tax returns, the obligations are specific to your profession; using AI with confidential client data covers what lawyers, clinicians, accountants and HR teams have to check.
Why is pasting into a consumer chatbot risky? #
Once you hit send, the text sits on the provider’s servers under its retention policy. It may be reviewed by staff for safety, and on many consumer plans it can be used to train future models unless you opt out. Your company has no contract with that provider covering your data, and no way to get it back.
This has already caused trouble. In 2023, Samsung restricted staff use of generative AI tools after engineers pasted internal source code into ChatGPT. No hack was involved. People used a helpful tool the obvious way.
Does your employer’s AI policy allow it? #
Most organizations now have something written down. Look for four things specifically, because a policy that only says “be careful with AI” leaves you guessing:
- Which tools are approved, and whether that means a specific account or just the product name.
- What categories of data are banned from any AI tool, approved or not.
- Whether personal devices are allowed to hold work material at all. This is the one people miss, and a local AI app doesn’t get around it.
- Who to ask. Security, legal or IT usually owns the exception process.
If no policy exists, that isn’t permission. It usually means nobody has written one yet, and the confidentiality terms in your contract still apply.
Shadow AI, meaning staff quietly using unapproved tools because the approved one is slow or doesn’t exist, is the common failure mode. It’s also the reason many companies would rather hear the question than discover the answer later.
What changes with a business account? #
A fair amount. Business and enterprise tiers such as ChatGPT Business or Enterprise, Gemini in Google Workspace and Claude for Work generally exclude your content from training by default, come with a contract covering data handling, and let administrators set retention. That’s the difference your security team cares about.
The flip side is worth knowing: on a company-managed account, administrators may be able to see what you’ve pasted. A business account protects the company, not necessarily your privacy from the company.
Safer options, ranked #
| Option | Where your text goes | When it fits |
|---|---|---|
| Your company’s approved AI tool | A provider under a business contract | Most work tasks, when one exists |
| Local model on a company-managed computer | Stays on that machine | Teams with IT support for it |
| Cloud chatbot with details removed | The provider, minus the sensitive parts | Quick help with structure and wording |
| Local model on your own phone | Stays on your phone | Your own drafts and notes, if policy allows |
| Consumer chatbot with raw confidential data | An outside company, possibly training on it | Don’t |
Most of AI’s value for writing is in structure and tone rather than the specifics, so removing names, numbers and identifiers usually costs you nothing. “Rewrite this note declining a request from Client A about extending the deadline” works exactly as well as the version with the real name. The step-by-step techniques are in what not to tell an AI chatbot. Redaction stops working when the details are the substance, which is when the next option matters.
Where an offline AI on your phone helps #
Some of the most useful work tasks involve your own text: drafting a tricky email, making a message sound calmer, turning meeting notes into action items, working out what a long document actually says. A local model does all of that without sending anything out, and it works on a plane or a train with no signal.
Personal LLM runs open models such as Qwen 3.5 4B and Gemma 4 E4B on your phone’s own chip. You can attach a PDF, text or Markdown file and ask about it: the text is extracted and searched on the phone, the model answers from the matching passages and tells you which ones it used, and nothing is uploaded. A per-chat system prompt lets you set up a “work email editor” chat once and reuse it. There’s no account and no server of ours, and after the one-time model download the only network traffic is ads in the free version while you’re online.
Be realistic about quality. A phone-sized model is good at rewriting, summarizing and explaining what’s in front of it, and weaker on facts it would have to remember. Check anything that matters.
What local AI doesn’t fix #
- Your company’s policy. Some employers ban work documents on personal devices outright. A local AI keeps data off AI servers but still puts it on your phone. Ask before you do it.
- Device security. A lost phone without a passcode exposes everything on it.
- Backups. Phone backups to iCloud or Google may include app data, and exported chat backups are plain readable files.
- Accuracy. Local or cloud, AI makes mistakes, and you’re still responsible for what you send.
For the specific case of forwarding a message you received, see is it safe to paste emails into ChatGPT. To confirm a tool is really offline before you trust it with anything, use the checks in how to tell if an AI app is really private.
Frequently asked questions #
Is it safe to use ChatGPT for work? #
It’s fine for tasks with no confidential information, such as general questions or rewriting public text. For company data, use your employer’s approved tool or plan, which typically comes with business terms on data use. Pasting confidential material into a personal consumer account is the main risk.
Can my employer see what I type into ChatGPT? #
On a company-managed business account, administrators may have access to usage and conversations depending on how it’s configured. On a personal account your employer can’t see your chats through OpenAI, though work devices and networks are often monitored, which can log the activity even if not the content.
Does ChatGPT use my data for training? #
On consumer plans, conversations may be used to improve models unless you turn that off in data controls. Business and enterprise plans are generally excluded from training by default. Check the terms for the account you’re actually signed into.
Is local AI allowed under my company’s AI policy? #
It depends on the policy. Running a model locally avoids sending data to an AI provider, which addresses the most common concern, but some policies also restrict storing work data on personal devices. Ask your IT or security team rather than assuming.