ChatGPT Security Guidance for Business

Start with a traffic-light table. It is the part of the policy employees remember. Rule of thumb: if you would not email it to a stranger, it stays out of personal AI accounts.

"Amber" means allowed with a written condition. The workspace column assumes the workspace is configured properly, which is the next question.

The gap between a personal account and a company-managed workspace is larger than most leaders expect. Using OpenAI's own documentation as of October 2026 as the example:

Training use. OpenAI says it may use content from services for individuals to train its models, and that users can opt out by turning off "Improve the model for everyone" under Settings, Data controls. For ChatGPT Business, Enterprise and Edu, and for its API, OpenAI states it does not train on inputs or outputs by default. Other providers differ, so check each one's documentation.

Retention. In Enterprise workspaces, OpenAI says workspace admins control how long data is retained. Temporary chats on individual accounts are not used for training but may be kept for up to 30 days for safety purposes. Deleted content can still be retained where the law requires it.

Identity and visibility. OpenAI lists SAML single sign-on for its business products, so you can cut access the day someone leaves. On a personal account, a former employee keeps every conversation about your company.

Our position is blunt: if more than a handful of people use an assistant for work, pay for the business tier. IBM's 2026 Cost of a Data Breach study found that 92% of organizations that had an AI-related breach lacked proper AI access controls, and only 40% of organizations reported using access controls on AI models and data at all.

A plain chat window only knows what someone types. Once an assistant can search a shared drive, read email or query a CRM, one prompt can pull data the user forgot they could access. Custom GPTs add instructions and uploaded files that can be shared with colleagues or, if allowed, the public.

OpenAI says workspace admins can control which apps are enabled and whether GPTs can be shared publicly. Use those controls. In our reviews, the problem is rarely the assistant. It is the over-shared 2021 folder the connector can now summarize for anyone.

Connected assistants also inherit a risk the OWASP project ranks first: prompt injection. A document or web page can carry hidden instructions that steer the assistant into leaking data or acting. We cover agent permissions in how to secure AI agents.

Policies that get followed fit on one page. Here is the outline we hand clients, and it maps to the controls in our AI governance playbook.

Canada. In December 2023, the Office of the Privacy Commissioner of Canada and its provincial and territorial counterparts issued principles for responsible, trustworthy and privacy-protective generative AI. For organizations using these tools: enter personal information into prompts only where authorized, use anonymized or de-identified information where possible, and avoid prompting a system to re-identify de-identified data. Under PIPEDA, your company stays accountable for personal information it hands to a service provider. In Quebec, Law 25 requires a privacy impact assessment before personal information is communicated outside the province, which covers most cloud AI services. Our data security and privacy services cover that assessment.

United States. There is no single federal AI privacy law, but the FTC has said plainly that AI companies must uphold their privacy and confidentiality commitments, and that quietly changing terms to reuse data can be unlawful. The same logic applies to you: if your privacy notice promises customers something, your AI use has to fit it. For healthcare data, HHS guidance on cloud services and HIPAA makes a provider that stores or processes protected health information a business associate, so no PHI goes in without a signed BAA.

For a structured risk view, NIST's Generative AI Profile (NIST AI 600-1) lists data privacy and information security among generative AI risks, with actions you can map to the policy above.

If your team uses a business-tier assistant for drafting, summarizing public material and first-pass code with no secrets, the one-page policy and the plan decision are enough. You do not need an AI governance framework, a dedicated tool, or a committee.

You also do not need to block anything. Companies that block ChatGPT at the firewall usually find staff simply switched to phones.

The picture changes when you build AI into your product, connect assistants to customer data at scale, or let them take actions. Then you need threat modeling, a formal AI governance program, and possibly ISO 42001 certification if enterprise customers ask for it.

What a policy does not stop: a determined insider, a phishing email written with AI, or a vendor that changes its terms. Those need your normal controls.

Ask, do not audit. A short anonymous survey on which assistants people use, and on which accounts, gets honest answers. Check expense records for AI subscriptions.

Choose one approved assistant, buy the business tier, turn on SSO, set retention, and switch off connectors until reviewed.

Send the policy and run a 20-minute session with real examples from your teams.

Work through the connector checklist, rotate any credential ever pasted, and add AI use to your vendor register and privacy notice.

It can be, on a company-managed business plan with sensible settings and a clear policy. Most risk sits in personal accounts and over-connected workspaces.

It reduces training use, but you still have no admin control, SSO or offboarding.

It depends on your privacy notice, contracts and jurisdiction. In Canada, the OPC expects authorization and de-identification where possible.

One named owner, usually the COO, CTO or a fractional security leader. Our virtual CISO service starts with the Crawling tier from $2,000 a month for 15 to 20 hours. If you would rather talk it through first, book a consultation.