Three ways to set one up.
- Personal accounts. Staff sign up on their own, free or paid. On the individual plans of ChatGPT and Claude, the provider may use conversations to train its models unless each user turns this off. The company has no admin control.
- A business plan. ChatGPT Business and Enterprise, Claude Team and Enterprise, Microsoft Copilot Chat for organisations and Gemini in Google Workspace each say they do not train models on your business data by default. The exact wording differs; see the table below. These plans add admin controls and business terms for how the provider handles your data. The provider still processes and stores the conversations.
- Your own server. An open-weight model and a chat interface run on hardware you control. Conversations stay on your network, and your team, or a partner you hire, runs and secures it.
For most small teams, moving work chats from personal accounts to a business plan is the quickest improvement. Your own server fits when some data should not leave your network, or when heavy use makes a fixed cost attractive. The two can be combined: a business plan for general writing and a local model for sensitive documents.
| Question | Personal accounts | Business plan | Your own server |
|---|---|---|---|
| Used to train the provider’s models? | Can be, unless each user opts out | Not by default | No. Nothing goes to a model provider |
| Where conversations are kept | With the provider | With the provider, under the plan’s retention settings | On your hardware |
| Company control | None. Each account belongs to the user | Admin console and company sign-in | Full control, and full responsibility |
| Models | The provider’s | The provider’s | Open-weight models that fit your hardware |
| Who runs it | The provider | The provider | You, or a partner you hire |
| Cost | Free, or a subscription per person | A fee per user per month | Hardware, power and upkeep, mostly fixed |
| If the internet goes down? | Stops working | Stops working | Can keep answering |
Five questions to ask about any option.
“Private” is not a single setting. Ask these questions about each option, and about each tool connected to it:
- Training. Can the provider use our conversations to train or improve its models? Check the feedback buttons too. OpenAI says business data shared through its opt-in feedback mechanisms may be used for training. Anthropic keeps a conversation sent with a thumbs up or down for up to five years, and owners of Claude Team and Enterprise plans can turn feedback off with the Rate chats setting.
- Retention. How long are conversations and uploaded files kept, and does deleting them remove them from the provider’s systems? On ChatGPT Business, OpenAI removes deleted conversations within 30 days, unless the law requires longer or it needs them to protect its services or others from harm.
- Access. Who can read the conversations? On ChatGPT Business, workspace admins can view, export and delete members’ conversations. OpenAI employees and contractors can access them for limited purposes, such as engineering support, abuse review and legal compliance.
- Location. Where is the data processed and stored? Microsoft says Copilot Chat sends requests to the closest data centres in the region, but can use other regions when demand is especially high. ChatGPT Enterprise offers data residency in ten regions, including Singapore.
- Connections. What else does it reach? Web search, email, file drives and plug-ins each add a data flow with its own terms. For example, Microsoft handles the short web search queries Copilot Chat sends to Bing as an independent data controller, not on your behalf.
What the business plans say.
From each provider’s own pages, checked on 10 October 2026. Plans and terms change, so check the source before you rely on a detail.
| Plan | Training on your data | Retention and access | Other terms |
|---|---|---|---|
| ChatGPT Business (OpenAI) | “No training on your business data by default.” | Admins set retention. Deleted chats are removed within 30 days. Admins can view and export members’ chats. | SGD 25 per user per month billed annually, from 2 users, or SGD 32 monthly. ChatGPT Enterprise adds custom retention and data residency, including Singapore. |
| Claude Team and Enterprise (Anthropic) | Not used for training by default, except feedback a user chooses to send. | Feedback is kept for up to 5 years; owners can turn it off. Enterprise adds custom retention and audit logs. | Team is for 2 to 150 people, with single sign-on. |
| Microsoft Copilot Chat, for organisations | Prompts and responses “aren’t used to train foundation models”. | Logged in your Microsoft 365 tenant for audit and eDiscovery. Your retention policies apply. | Enterprise data protection at no extra cost, with Microsoft as data processor under its Data Protection Addendum. Web searches sent to Bing follow separate terms. |
| Gemini in Google Workspace | “Not human reviewed or otherwise used for Generative AI model training outside your domain without permission.” | “Your interactions with Gemini stay within your organization.” | Applies with a qualifying Google Workspace edition. |
Personal plans are different. On ChatGPT Free, Go, Plus and Pro, OpenAI may use conversations to train its models unless the user turns off Improve the model for everyone under Settings, Data controls. Temporary chats are not used for training while they stay temporary. Claude’s Free, Pro and Max plans also let each user opt out of model training.
What running your own involves.
A self-hosted assistant has three parts:
- A model. Open-weight models can be downloaded and run on your own hardware. They range from a few billion to hundreds of billions of parameters, and what fits depends on your hardware’s memory. Test candidates on your real tasks before you choose.
- A runtime that loads the model and answers requests, such as llama.cpp, Ollama or vLLM.
- A chat interface for staff. Open WebUI, for example, connects to Ollama or any OpenAI-compatible server. It supports roles, groups and per-model permissions, and knowledge bases built from uploaded files.
Conversations stay private only if every part stays under your control: the interface, the document index, connected tools, remote support and backups. See “Local” is a data-flow question.
You also take on work a provider would do for you: user accounts and sign-in, security updates, backups, log retention, monitoring and checking answers. Open-weight models can do a lot of everyday work, but they are not the largest cloud models, so test them on your own tasks. For memory sizing and costs, see Local AI vs cloud AI.
How to choose.
- If staff use personal accounts for work today, move that work to a business plan or an internal tool, and tell them what they may paste in.
- If the work involves customer records, NRIC numbers, health or financial details, or client documents under confidentiality terms, keep those tasks on a system you control, or first check the business plan’s retention, access and location against your obligations.
- If many people use it all day, compare the per-user fees with a fixed-cost server. The cost section of Local AI vs cloud AI shows how to estimate both.
- Whatever you pick, test it on your own tasks before you roll it out.
The PDPA applies either way.
Choosing a provider does not move your obligations to it. PDPC’s 2026 guidelines on generative AI say businesses that deploy AI systems “bear primary responsibility” for making sure the systems they choose can meet their obligations under the PDPA. No set-up is compliant on its own. PDPA and AI explains the obligations and includes a checklist.
Common questions.
Is ChatGPT private?
It depends on the plan. On personal plans, OpenAI may use conversations to train its models unless you opt out. ChatGPT Business and Enterprise do not use business data for training by default. On every plan, OpenAI processes and stores the conversations.
Does ChatGPT Business train on our data?
Not by default. OpenAI’s enterprise privacy page, updated 8 January 2026, says it does not train its models on business data unless you opt in, for example by sharing feedback.
Can staff use free AI accounts for work?
That is a policy decision for your company. Free accounts give the company no admin control, and on ChatGPT and Claude they may be used for training unless each person opts out. Keep personal data and confidential client information out of them.
Can a private AI assistant work without the internet?
A local model with a local chat interface can keep answering when the internet is down. Web search, cloud connectors and software updates still need a connection. Cloud plans stop working without one.
How much does it cost?
Business plans charge per user per month. ChatGPT Business, for example, is listed at SGD 25 per user per month billed annually, or SGD 32 monthly, on ChatGPT’s pricing page as shown in Singapore on 10 October 2026. Your own server is mostly a fixed cost: hardware, power and upkeep.
Next: list what your team pastes today.
Write down the AI tools in use and the data that goes into them, then pick a set-up for each kind of work. The checklist in PDPA and AI is a good start. If you want a local assistant set up for you, see our deployment process.
Sources.
- OpenAI: Enterprise privacy (updated 8 January 2026)
- ChatGPT pricing (checked 10 October 2026, Singapore prices)
- OpenAI Help Center: How your data is used to improve model performance
- Anthropic Privacy Center: Is my data used for model training? (commercial products)
- Claude plans and pricing (checked 10 October 2026)
- Microsoft Learn: Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat
- Microsoft Learn: Microsoft Copilot Chat privacy and protections
- Google Workspace: Generative AI in Google Workspace Privacy Hub (updated 6 October 2026)
- Open WebUI: Features
- PDPC: Advisory Guidelines on Use of Personal Data in Generative AI (issued 20 July 2026)
ChatGPT, Claude, Copilot and Gemini are trademarks of their owners, and this guide is independent of them. Plans and terms change, so check the source before relying on a detail. This guide is general information, not legal advice.