Your employees are already using ChatGPT at work. The question is not whether AI is entering your organisation — it already has. The question is whether your confidential data is leaving your network every time someone asks it a question.
With ChatGPT, every prompt your team types is transmitted to OpenAI's servers in the United States. It is processed there. A response is generated there. And depending on your account tier, that conversation may be retained, reviewed, and used by OpenAI employees.
NoCloudGPT gives your teams the same productivity gains — drafting, summarising, coding, analysing — without a single byte of your data touching OpenAI's servers. It runs entirely on your own infrastructure. Your prompts stay on your machines. Your data stays in your building.
When a member of your team opens ChatGPT and types a prompt, that text is transmitted to OpenAI's servers. It is processed there. A response is generated there. And depending on your account tier, that conversation may be retained, reviewed, and used.
This is not speculation. This is how the product works.
OpenAI's own documentation confirms that conversations can be reviewed by their teams for safety monitoring, quality assurance, and model improvement. Enterprise tier customers can opt out of training data use — but opt-out does not mean your data never touches their infrastructure. It means OpenAI has contractually agreed not to use it for training. The data still transits their systems. OpenAI employees with the appropriate access level can still view it.
For most businesses, this creates three distinct problems:
Without policy enforcement at the network level, there is no reliable way to stop an employee from pasting a client contract, a financial model, a piece of proprietary source code, or a patient record into a ChatGPT prompt. It takes one person, one paste, and your confidential data is on a server you do not own, governed by a privacy policy you did not negotiate, subject to laws in a jurisdiction you may not operate in.
OpenAI's enterprise data opt-out is a contractual commitment, not a technical one. Your data still travels to their servers. It is still processed by their infrastructure. The commitment not to train on it is a promise, not a wall. If OpenAI experiences a data breach, a subpoena, a change in ownership, or a policy revision, that promise becomes complicated.
GDPR requires you to know where personal data goes and to have a legal basis for sending it there. HIPAA requires you to ensure protected health information is not disclosed to unauthorised parties. Legal professional privilege requires that confidential communications between lawyers and clients remain confidential. Every time an employee pastes regulated data into ChatGPT, you have a potential compliance event — whether or not anything bad happens as a result.
OpenAI's privacy documentation is more transparent than most, which means we can be specific. Under their standard usage policies, OpenAI states that authorised employees may access conversation data for purposes including:
"Authorised employees" is not a small group in a company of OpenAI's scale. It includes engineers, safety researchers, trust and safety teams, legal teams, and contractors. The exact number is not published. The access controls are not auditable by you.
This is not a criticism of OpenAI's intentions. It is a structural reality of any cloud service. When your data is processed on someone else's infrastructure, the people who run that infrastructure have access to it. That is the model. The question for your business is whether that is acceptable for the data your employees are working with daily.
These are not theoretical. They are the conversations happening in legal, finance, IT, and HR departments at organisations that have already started asking the right questions.
The note contains the client's name, the nature of the dispute, the legal strategy being considered, and the likely outcome assessment. All of this is transmitted to OpenAI's servers. Legal professional privilege may be waived. The client has not consented. The firm has not conducted a data processing impact assessment. Several bar associations have already issued guidance warning lawyers about exactly this scenario.
The model contains unpublished revenue figures, margin assumptions, and acquisition targets. This is material non-public information. If it reaches a third party — even inadvertently, even theoretically — the regulatory implications under MiFID II, SEC rules, or equivalent frameworks are serious. The analyst did not think of it as a data transfer. They thought of it as getting help with a formula.
The document contains the employee's name, role, salary, disciplinary history, and personal circumstances cited in mitigation. Under GDPR, this is sensitive personal data. Processing it on a third-party server without a lawful basis and appropriate safeguards is a breach — regardless of whether any harm results.
The code contains trade secrets, authentication logic, and undocumented API structures. Depending on the employment contract and the company's IP agreements, this may constitute a breach of confidentiality. The developer meant to fix a bug. They may have disclosed the product's core architecture to a US company's infrastructure.
| Feature | ChatGPT Enterprise | NoCloudGPT |
|---|---|---|
| Data leaves your network | Yes — every prompt | Never |
| Third-party employee access | Possible — OpenAI staff | Not applicable |
| Training data opt-out | Contractual only | Not relevant — no third party |
| GDPR compliance | Requires DPA with OpenAI + SCCs | Full control, data never transferred |
| HIPAA compliance | BAA available, data still transits | No third-party data transfer at all |
| Legal privilege protection | At risk on every prompt | Fully maintained |
| Audit logs under your control | No | Yes — complete |
| Air-gap / offline operation | No | Yes |
| Model selection | GPT-4o, o1, o3 only | 70+ open models |
| Custom fine-tuning on your data | Limited | Full control |
| Vendor lock-in | High | None — open source stack |
| Per-seat cost | ~$30/user/month | Infrastructure cost only |
| Government subpoena risk | Yes — US jurisdiction | Your jurisdiction only |
Article 28 of GDPR requires a Data Processing Agreement before you transfer personal data to any third-party processor. OpenAI offers a DPA for enterprise customers — but the DPA does not change the fact that data is being transferred to a US-based company, which requires either adequacy decisions or Standard Contractual Clauses to be legally valid. With NoCloudGPT, there is no third-party processor. There is no transfer. There is no DPA required. The data never leaves your infrastructure.
OpenAI offers a Business Associate Agreement for qualifying healthcare customers. Like the enterprise DPA, the BAA is a contractual framework around data that still physically transits OpenAI's infrastructure. NoCloudGPT eliminates the third party entirely. There is no covered entity relationship to manage because there is no third party involved in processing.
Privilege applies to confidential communications between lawyer and client made for the purpose of legal advice. Once that communication is shared with a third party — including an AI service provider — privilege may be lost. Several bar associations have issued guidance warning lawyers about the privilege implications of using cloud-based AI tools. A locally deployed model does not receive the communication. There is no third-party disclosure.
Enterprise security frameworks require you to know where your data goes and to have controls over it. A self-hosted AI deployment on your own infrastructure is auditable end-to-end. A cloud AI service introduces a third-party dependency that your auditors will ask about — and that requires vendor risk assessment, ongoing monitoring, and incident response planning for a system you do not control.
Switching from ChatGPT to a private deployment does not require replacing every workflow overnight. Most organisations run both in parallel during transition, with clear policies about which tool is appropriate for which data classification level.
# Install Ollama on Linux
curl -fsSL https://ollama.com/install.sh | sh
# Pull a ChatGPT-quality model
# Llama 3.3 70B or Mistral Large give the closest experience
ollama pull llama3.3:70b
# Or for lighter hardware requirements
ollama pull mistral:7b
# Start serving
ollama serve# Install Open WebUI — browser-based interface your team will recognise immediately
docker run -d \
--name open-webui \
-p 3000:8080 \
-v open-webui:/app/backend/data \
--add-host=host.docker.internal:host-gateway \
ghcr.io/open-webui/open-webui:main
# Access at http://localhost:3000
# Your team sees a familiar chat interface — nothing leaves the server# Ensure Ollama only listens on localhost — no external access
# In /etc/systemd/system/ollama.service add:
# Environment="OLLAMA_HOST=127.0.0.1"
# Block outbound connections from the inference server
# using your firewall — ufw example:
ufw default deny outgoing
ufw allow out 53 # DNS only
ufw allow out 443 # Only if you need to pull model updates
ufw enable
# Verify no outbound connections during inference
ss -tunp | grep ollama# Open WebUI supports SSO via OAuth/OIDC
# Connect to your existing identity provider
# Set in Open WebUI admin panel:
# OAUTH_CLIENT_ID=your_client_id
# OAUTH_CLIENT_SECRET=your_client_secret
# OPENID_PROVIDER_URL=https://your-idp.company.com
# Enable audit logging — all prompts logged locally
# Open WebUI logs conversation history to its local database
# Back up with your standard backup tooling — no external dependency| Use Case | Recommended Model | VRAM Required | Why |
|---|---|---|---|
| General office assistant | Llama 3.3 70B | 40 GB | Closest ChatGPT-quality experience |
| Legal document analysis | Mistral Large 2 | 70 GB | Long context, EU origin, privilege-safe |
| Financial modelling help | DeepSeek R1 70B | 40 GB | Strong quantitative reasoning |
| Code review and generation | DeepSeek Coder V2 | 40 GB | Purpose-built for code tasks |
| Lightweight / low hardware | Mistral 7B | 8 GB | Runs on a single consumer GPU |
| Multilingual EU workflows | Mistral Large 2 | 70 GB | Native EU language depth |
| Document Q&A / RAG | Command R+ | 60 GB | Purpose-built for retrieval tasks |
No. Once a model is downloaded, NoCloudGPT operates entirely offline. It can run in fully air-gapped environments with no network connectivity whatsoever. This makes it suitable for classified environments, secure research facilities, and organisations with strict network isolation policies. The initial model download requires internet access — after that, nothing does.
For the tasks most business users perform daily — drafting documents, summarising text, answering questions about uploaded documents, writing and reviewing code, analysing data — models like Llama 3.3 70B and Mistral Large 2 are practically indistinguishable from GPT-4o in output quality. The gap is real for highly specialised tasks and the very latest capabilities. For general enterprise use, it is not a meaningful limitation.
ChatGPT Enterprise removes your data from training and provides a DPA. It does not change the fundamental architecture: your data is still processed on OpenAI's infrastructure, is still accessible to OpenAI personnel under certain conditions, is still subject to US law and potential government access, and is still outside your audit perimeter. For many use cases this is acceptable. For legal documents, MNPI, health records, and source code containing trade secrets, the contractual protections do not address the structural exposure.
Model updates are weight files, not data. Downloading a new model version does not send any of your prompts or documents anywhere — it simply downloads a new set of neural network parameters, exactly like downloading a software update. You can schedule model updates during maintenance windows, verify checksums, and choose never to update if your current model meets your requirements. Your data and your model updates are entirely separate concerns.
Yes — and most organisations do exactly this during transition. The practical approach is a clear data classification policy: anything classified as internal, confidential, or restricted uses the private deployment. General research, public-facing content drafting, and non-sensitive productivity tasks can use whatever tools your policy permits. The private deployment does not require you to eliminate other tools — it requires you to have a clear policy about which data goes where.
ChatGPT is a remarkable product. OpenAI has invested billions in safety research, alignment work, and enterprise data handling. This is not an argument that OpenAI is careless with your data.
It is an argument that no matter how careful a third party is with your data, they still have your data. And for legal documents, financial models, medical records, source code, and HR files, that is the problem.
NoCloudGPT does not ask you to trust a third party. It runs on hardware you own, in a location you control, with audit logs only you can read. Your employees get the AI productivity tools they want. Your compliance team gets the data residency guarantees they need. Your lawyers keep their privilege.
The data never leaves. That is the only guarantee that actually means something.
No OpenAI. No cloud. No compliance risk. Just AI that works on your infrastructure.
→ Deployment Guide | NoCloudGPT vs Claude → | NoCloudGPT vs Copilot →
Related: NoCloudGPT vs Claude · NoCloudGPT vs Copilot · Legal Sector AI Guide · Finance Sector AI Guide · Quickstart Deployment Guide