Is Poe safe for work data in Hong Kong?

Is Poe safe for work data? The answer depends on the bot, the material and your firm's rules. Poe's privacy shield tells you something useful about who may receive a chat and whether it may be used to train a model. It does not approve a client document for you. That decision still belongs to the person responsible for the data and the organisation that owns the work.

Poe is a practical route to AI models from Hong Kong. I have written about what it gives you, and where it stops. The data question deserves its own answer because “I can open the tool” and “I can paste this file into it” are separate decisions.

Is Poe safe for work data? Start with the shield

Poe's privacy centre, updated 30 April 2026, describes two shield states on its bots and apps.

• Full shield: Poe says the content goes to the model provider, or to a service running in Poe's sandbox, and the model provider will not use the interaction for training.
• Half shield: Content may also go to the bot's developer or external services. Poe says those interactions may be used for training.

The shield can change if you grant a bot new permissions. Check the bot's information card before using it, then check the message information if you added a permission during the conversation.

A full shield does not mean your text stays inside your company or only on Poe's servers. Poe says it shares chat contents with the model providers that power its bots. Each provider sets its own retention schedule. “Not used for training” answers one question; it does not answer every question about access and retention.

A privacy shield describes the route your data takes. It does not decide whether the journey is authorised.

The part that becomes identifiable again

Poe says it does not send the name and email attached to your Poe account to model providers. That is narrower than saying a chat is anonymous. If the chat itself contains a client's name, deal terms or a distinctive description of the business, you have supplied the identifying detail in the prompt.

Consider a credit memo due on Thursday. Removing the borrower's name still leaves an identifiable transaction if you paste the exact facility size, sector, unusual covenant and closing date. Could someone work out who it describes? For the first experiment, use a synthetic memo with invented numbers.

Poe's memory setting needs a separate look. Its privacy centre says memory can carry details from earlier conversations to other model providers you use on Poe. It also says information deleted from a chat can remain in a memory summary until you delete that summary separately. If you are evaluating Poe for work, inspect that setting before making the first real upload.

The permission question Poe cannot answer

Poe's privacy page describes how its product handles a chat. Your firm's policy and client terms decide whether that chat should happen. A full shield does not establish that the vendor, provider, location and retention fit those obligations. Ask whoever owns the data decision before trying a live matter; use an invented example while the answer is pending.

Four questions before a company file goes in

1. Is this tool approved for this class of data? Your firm's permission matters even when a vendor's published settings look reassuring. If the answer is unclear, use synthetic material until someone accountable gives you the rule.
2. Which bot is handling the request? Check the shield, developer, permissions and any external access for the actual bot you will use.
3. Who may receive and retain the content? Read Poe's current terms and the relevant provider's terms. Record the answer for a team decision.
4. Can the task be done with less data? Start with an invented or properly reduced example. A whole client file is rarely needed to test whether AI can improve the structure of a memo.

Hong Kong's Privacy Commissioner framework for organisations using AI is a useful starting point when personal data is involved. It is a framework for choosing and governing an AI system, not a blanket approval of a particular app. For the wider question of company data in AI tools, I have set out the Hong Kong work-data rule set separately.

There is an honest limit here. I cannot tell from a public privacy page whether Poe meets your firm's contracts, record-keeping duties or client promises. Nor does a price tier settle it. If the work is sensitive, the useful first session is a tool and data decision: what you can use, what you must remove, and what stays out. Once that is clear, we can build the workflow around your real documents without asking everyone to guess at the boundary.

Talk to Metis AI about a private setup session.