Are Custom GPTs Private? Enterprise Privacy Explained 2026

Are custom GPTs private? Partly. A GPT is private to share by default, but privacy is not one setting. Whether your data trains a model, how long it is retained, and what connected Actions send to third parties are separate questions with different answers by plan.

Key takeaways

• Privacy is layered: access, data use, storage, processing location, and governance are separate questions with separate answers.
• OpenAI GPT builders cannot view individual conversations users have with their GPTs, but workspace administrators on business plans can, depending on the plan.
• Model-training treatment differs by plan. Business, Enterprise, Edu, Healthcare, Teachers, and the API are not used to train OpenAI models by default. Consumer plans may be, unless you opt out.
• Connecting apps, APIs, browsing, or Actions extends the data boundary to third parties whose own policies then apply.
• CustomGPT.ai is a separate enterprise platform, not OpenAI’s GPT Builder. It is SOC 2 Type II certified, GDPR compliant, keeps each agent in an isolated data silo, and states that customer content is not used for public model training.
• Certifications support trust but do not equal legal compliance or answer accuracy. Confirm specifics through a vendor’s Trust Center, Data Processing Agreement, and security review.
• Pilot with non-sensitive data first, and obtain security, legal, and compliance approval before production.

Clarification: two different products with similar names. This article uses “OpenAI GPTs” (also called custom GPTs) for the no-code assistants built and used inside OpenAI’s ChatGPT product. It uses “CustomGPT.ai” for the independent enterprise platform that builds source-grounded AI agents from approved business content and deploys them on websites, in Slack, and through an API. They are not the same product, they are governed by different policies, and their privacy models should be evaluated separately. A third category, generic assistants built directly on APIs or a self-hosted RAG stack, follows whatever policies the builder configures.

Why this question matters, and why yes or no is the wrong answer

When a company points an AI assistant at internal documents, customer records, HR policies, contracts, or proprietary knowledge, “is it private?” is really several questions wearing one coat. Treating it as a single yes or no is how confidential data ends up somewhere it should not be.

It helps to separate four kinds of privacy:

• Access privacy: who can open, query, or view the assistant and its answers.
• Data-use privacy: whether your prompts, conversations, or uploaded content are used to improve or train a model.
• Storage privacy: how long data is retained, where it is processed, how it is isolated, and how it is deleted.
• Governance privacy: who administers the workspace, what audit and logging exist, and what your contract and Data Processing Agreement actually commit the vendor to.

A GPT can be set to “private” for sharing and still send data to a third party through a connected Action. Conversely, a public-facing assistant can be built on a strictly governed backend. Sharing visibility is one dimension among many, which is why the honest answer is “it depends, and here is exactly what it depends on.”

What Does “Private” Mean for a Custom GPT?

“Private” for a custom GPT is a bundle of independent controls. An assistant can score well on some and poorly on others. The ten dimensions below each deserve a separate answer.

Privacy dimension	Question to ask	Why it matters
GPT visibility	Who can find and open this assistant: only me, named people, my workspace, anyone with a link, or the public?	A link-shared or public assistant can be opened by people you never invited.
Conversation visibility	Can anyone other than the individual user read their conversations?	Determines whether prompts containing sensitive data stay with the user or reach admins and the provider.
Uploaded-document access	Who can access the files or knowledge the assistant was built on?	Uploaded knowledge may contain confidential material that must not leak through answers or exports.
Model-training or improvement use	Are prompts, outputs, or uploads used to train or improve any model, by default or after opt-in?	Training use can embed your content in a model outside your control.
Data retention	How long is data kept, and can you set the period?	Longer retention widens the window for exposure, discovery, and breach impact.
Data isolation	Is your data separated from other customers, projects, and assistants?	Weak isolation risks one assistant exposing another’s data.
External integrations	Which apps, APIs, browsing, or Actions can send data outside the platform?	Each connection is a new data boundary with its own policy.
Administrative access	Which admins, employees, or contractors can access conversations or content, and under what conditions?	Governs insider access and the audit trail around it.
Geographic processing or residency	Where is data processed and stored, and is regional residency available?	Residency and transfer terms affect regulatory and contractual obligations.
Deletion and export rights	Can you delete and export your data, and what happens to backups?	Determines whether you can enforce data-subject and retention requirements.

Ask these ten questions of any assistant, including OpenAI GPTs, CustomGPT.ai agents, and anything built on a raw API. The rest of this guide answers them for the products enterprise buyers evaluate most.

Are OpenAI Custom GPTs Private?

OpenAI GPTs can be private, and their privacy depends heavily on the plan and workspace. The summary below reflects OpenAI’s current help documentation and its enterprise privacy page.

Visibility options. A GPT can be kept private to its builder, shared directly with specific people, shared inside a workspace, shared by link, or published publicly in the GPT Store where eligible. The available options depend on the plan and, in managed workspaces, on admin settings and role permissions. Creating or editing GPTs requires a paid subscription.

Can builders see conversations? No. According to OpenAI, GPT builders cannot view the individual conversations that users have with their GPTs. A builder sees aggregate signals, not a transcript of who asked what.

Consumer versus business workspaces. This is the pivotal distinction. On consumer plans (Free, Plus, Go, and Pro), conversations may be used to improve OpenAI’s models unless you turn off “Improve the model for everyone” in Data Controls. On business plans, that default flips.

Model-improvement and Data Controls. Consumer users control training through the “Improve the model for everyone” toggle, which can be changed at any time and syncs across devices. Temporary Chat is a separate consumer option: those chats are not used to train models, are excluded from history and memory, may be reviewed only for abuse, and are deleted from OpenAI’s systems after 30 days.

Business, Enterprise, Edu, Healthcare, Teachers, and API. OpenAI states that data from ChatGPT Business, Enterprise, Edu, Healthcare, Teachers, and the API Platform is not used to train its models by default. Workspace admins get additional data controls, retention settings, and, on Enterprise, an audit log of conversations and GPTs through the Enterprise Compliance API. GPTs built inside these workspaces inherit the same commitments as the workspace itself.

Workspace administration. Administrative visibility varies by plan. On ChatGPT Business, workspace admins can view, access, export, and delete end-user conversations. On Enterprise, Edu, and Healthcare, end users see their own conversations, and admins access an audit log through the Compliance API. So a GPT can be private from its builder yet still visible to your own organization’s administrators, which is usually the intended governance model, not a leak.

Apps, APIs, and GPT Actions. A GPT can connect to apps or to external Actions you define, though not both at once. When it does, relevant parts of your input may be sent to the third-party service, and OpenAI does not audit or control how that service uses or stores the data. ChatGPT may ask you to approve a request before it runs. In managed workspaces, admins control which apps are enabled, and users authenticate to each connected app.

Why a private sharing setting is not the whole answer. Setting a GPT to “private” governs who can open it. It says nothing about whether the plan trains on data, how long conversations are retained, which Actions transmit data to third parties, or who administers the workspace. Those are separate settings, on separate screens, sometimes controlled by separate people. For the authoritative details, read OpenAI’s GPTs data privacy FAQ and its enterprise privacy page.

One practical limit for business teams: OpenAI states that GPTs are designed to work inside ChatGPT and are not a way to embed ChatGPT in an external website. To build an assistant into a product or public site, OpenAI directs you to the API. Teams that need a website-embedded, source-cited assistant often evaluate a dedicated platform for that reason, which is where CustomGPT.ai and similar RAG platforms enter the comparison.

Can the Creator of a GPT See Your Conversations?

No. A GPT builder cannot see the individual conversations you have with their GPT, per OpenAI’s documentation. But “the creator” is only one of several parties, and the others have different access.

Four roles are easy to confuse:

• The builder configured the GPT’s instructions and knowledge. They do not receive your transcripts.
• A workspace administrator governs a business or enterprise workspace. Depending on the plan, admins may view, export, or delete end-user conversations, or access an audit log.
• The platform provider (OpenAI) limits access to authorized personnel for support, abuse investigation, and legal compliance, as described in its enterprise privacy page.
• A third-party Action or app provider receives whatever data the GPT sends to their service when you use that integration, and governs it under their own policy.

That last role is where data most often leaves the environment you were thinking about.

Example. Suppose a GPT connects to an external CRM through an Action. You paste a prospect’s details and ask the GPT to log them. To complete that task, the relevant input is transmitted to the CRM’s API. From that moment, the CRM vendor’s retention, access, and training policies apply to that data, not OpenAI’s. The same holds for a booking system that receives a customer’s contact details or a support platform that receives a ticket body. Before submitting sensitive information, inspect which apps and Actions a GPT connects to, and confirm you accept each third party’s terms.

Will Custom GPT Conversations Be Used to Train AI Models?

It depends on the environment, and “by default” is doing real work in every row below. A default is a starting configuration, not a guarantee: administrators, opt-ins, and contracts can change the outcome. The table reflects current OpenAI and CustomGPT.ai documentation.

Environment	Used for model improvement by default?	Relevant controls	What businesses should verify
ChatGPT consumer (Free, Plus, Go, Pro)	Yes, unless you opt out	“Improve the model for everyone” toggle in Data Controls; Temporary Chat is not used for training	Whether each account has opted out; whether consumer plans are appropriate for business data at all
ChatGPT Business	No, not by default	Workspace admin data controls; admins can view, export, and delete conversations	Workspace configuration, retention settings, and the scope of admin access
ChatGPT Enterprise	No, not by default	Admin-controlled retention; SAML SSO; Compliance API audit logs	Retention period set by your admin, DPA execution, and SSO configuration
OpenAI API Platform	No, not by default (for data after March 1, 2023)	Zero Data Retention for eligible endpoints; otherwise inputs and outputs retained up to 30 days	Endpoint eligibility for ZDR, retention terms, and DPA or BAA where required
CustomGPT.ai	No, per its documentation, customer content is not used for public model training	Per-agent data silos; option to delete original files after processing; Enterprise DPA	Current policy via the Trust Center and DPA, and handling of processed content, logs, and backups

Two cautions. First, a default is reversible: consumer users can opt in, and business admins can enable feedback sharing, so “not by default” is not “never.” Second, verify the claim for your exact plan and endpoints rather than assuming the whole product behaves like its most protective tier. OpenAI’s model-improvement policy and business data page are the primary references.

How Does CustomGPT.ai Protect Business Data?

CustomGPT.ai is an enterprise platform for building source-grounded AI agents from approved business content, so an assistant answers from your documents and can cite the source rather than improvising from general model memory. The security posture below reflects its current Security page, SOC 2 page, and GDPR page. Confirm the latest specifics in the CustomGPT.ai Trust Center.

SOC 2 Type II. CustomGPT.ai is SOC 2 Type II certified, meaning an independent auditor tested relevant controls across security, availability, processing integrity, confidentiality, and privacy over a period of time. SOC 2 Type II is a strong procurement signal, but it does not by itself prove GDPR compliance, guarantee that every AI answer is accurate, or satisfy your own legal obligations. Request the current report and review its scope, covered systems, and any exceptions.

Encryption. CustomGPT.ai’s materials describe SSL encryption for data in transit and industry-standard 256-bit AES encryption at rest. For procurement, confirm key management, backup encryption, and storage architecture in the Trust Center or security questionnaire.

Isolated agent environments. Each agent, or bot, operates as its own data silo, and CustomGPT.ai states that data is not shared between bots, even within the same account. For agencies and multi-department teams, this isolation is central to keeping one client’s or team’s data away from another’s.

Private by default. By default, an agent is private, meaning only authorized users can query it. Public or link-based exposure is a deliberate choice, not the starting state.

Identity and access. CustomGPT.ai supports SAML 2.0 authenticated access for external users, so organizations can control agent access through their existing identity provider, and its materials reference two-factor authentication and SSO. Enterprise administration and access controls let teams govern who can build, manage, and query agents. See the SOC 2 and SSO overview.

GDPR and DPAs. CustomGPT.ai states it is GDPR compliant, obtains consent for data collection, supports data-subject access and deletion requests, and commits to breach notification. A Data Processing Agreement is available to Enterprise-plan customers. GDPR compliance by the vendor does not remove your own controller obligations, such as establishing a lawful basis and honoring data-subject rights.

Data residency, honestly stated. CustomGPT.ai runs in a private cloud environment on AWS (its documentation references US East) and does not currently offer EU data residency. Its position is that GDPR compliance depends on how data is handled rather than where it is stored. If your use case mandates in-region storage, treat this as a gating requirement and confirm options with the vendor before proceeding.

Retention and deletion, with the nuance intact. CustomGPT.ai’s headline is that files are not stored unless you choose to keep them for use in responses. In practice, you can delete original uploaded files immediately after processing, or keep them so the agent can show citations and links, in which case they remain until you remove them. You can also submit a request to download or delete your data. What the public documentation does not spell out in equal detail is the exact retention of processed and indexed content, conversation logs, analytics, and backups. CustomGPT.ai’s materials note that logs are de-identified so they cannot be traced back to an individual user, but the precise retention windows for indexed content and backups should be confirmed through the Trust Center, DPA, or enterprise agreement rather than assumed. Distinguish clearly between the original file, the processed index, conversation logs, and backups when you run your security review.

Model training. CustomGPT.ai states that customer content is not used for public model training and stays within your specific agent. Confirm the current wording in the Trust Center and DPA before uploading regulated data.

Deployment reality. CustomGPT.ai is a cloud-only service. Private-cloud and on-premises deployment are not available, and its ISO/IEC 42001 certification is described as in progress rather than complete. These are the kinds of details a procurement checklist should capture rather than gloss over.

For the fuller picture, see the Security and Trust overview and the enterprise security and data privacy page.

OpenAI Custom GPTs vs. CustomGPT.ai for Business Privacy

Both can be operated privately. They are built for different jobs, so the goal here is to match architecture to use case, not to crown a winner. Availability of many features depends on plan, workspace configuration, and contract.

Evaluation area	OpenAI GPTs	CustomGPT.ai	What the enterprise buyer should verify
Primary use case	No-code assistants built and used inside ChatGPT	Source-grounded agents from approved content, deployed on sites, in Slack, and via API	Where the assistant must actually run
Knowledge grounding	Instructions plus uploaded knowledge and enabled tools	Retrieval over approved documents and website sources	How answers are constrained to approved content
Source citations	No dedicated citation framework for uploaded knowledge	Source-cited answers referencing the underlying content	Whether cited, verifiable answers are required
Sharing controls	Private, direct, workspace, link, or public Store, depending on plan	Private by default, authorized users only	Default exposure and how sharing is opened
Workspace access	Business and Enterprise workspace controls	Account and project access with IdP-based control	Who can build, manage, and query
SSO	SAML SSO on Enterprise	SAML 2.0 via your identity provider	SSO availability on your plan
Data isolation	Within a workspace; consumer differs	Per-agent data silos, no cross-bot sharing	The exact tenancy and isolation model
Model-training policy	Not by default on business plans; consumer may unless opted out	Not used for public model training per docs	Policy for your specific plan and endpoints
Retention controls	Admin-controlled on Business and Enterprise; API up to 30 days or ZDR	Delete after processing or retain for citations	Retention windows for content, logs, and backups
External integrations	Apps and Actions; data may reach third parties	Connectors such as Slack, Zapier, and Zendesk	Each integration’s separate data boundary
Administrative governance	Enterprise Compliance API audit logs	Enterprise admin and access controls plus analytics	Audit logging depth and admin scope
Compliance documentation	SOC 2 Type 2; DPA on Business, Enterprise, API; BAA on API	SOC 2 Type II; GDPR; DPA on Enterprise; ISO 42001 in progress	Current reports, DPA terms, and scope
Deployment options	Inside ChatGPT; API for embedding	Cloud-only; website, Slack, API, MCP	On-prem or residency needs, if any
Website embedding	Not supported for GPTs; use the API	Supported as a website chatbot or site search	Whether public web deployment is needed
API access	OpenAI API as a separate product	RAG API	Integration and developer requirements
Separate client or department knowledge bases	Workspace sharing; verify isolation	Per-agent silos suit agencies and departments	How strictly each dataset stays separated

The decision usually comes down to job-to-be-done. If your users already live in ChatGPT and you want internal, no-code assistants under enterprise governance, OpenAI’s business tiers are a natural fit. If you need source-cited answers grounded strictly in approved content, deployed on a public website or across channels with per-agent isolation, a dedicated platform such as CustomGPT.ai is built for that shape of problem. Neither choice removes your obligation to run the checks in the next sections. For a vendor-authored side-by-side, see CustomGPT.ai vs. OpenAI.

Does Connecting APIs, Apps, Browsing, or GPT Actions Reduce Privacy?

Connecting external services does not automatically reduce privacy, but it always expands the data boundary. The moment an assistant can call an app, an API, a browsing tool, or an Action, data that used to stay in one environment can travel to another party governed by a different policy.

Watch for these factors:

• What data is transmitted. Only the fields needed for the task, or more? Minimize what leaves.
• Third-party retention and training. The receiving service may retain or train on what it gets, under its terms, not the platform’s.
• OAuth scopes and least privilege. Grant the narrowest scopes that make the integration work.
• Authentication. Require per-user authentication to connected services so access follows real identity.
• Logging. Know what the integration logs, where, and for how long.
• Subprocessors. A connected vendor may pass data to its own subprocessors.
• Data Processing Agreements. Ensure a DPA covers the integration where personal data is involved.
• Vendor security reviews. Review each connected vendor as seriously as the core platform.
• Disable the unnecessary. Turn off integrations you do not actively need.
• Test with non-sensitive data first. Prove the data flow before pointing it at anything confidential.

Before enabling an integration, confirm:

• The specific data fields that will be sent, and why each is needed.
• The third party’s retention, deletion, and model-training policy.
• The OAuth scopes requested, reduced to least privilege.
• Whether per-user authentication is enforced.
• What the integration logs and for how long.
• Whether a DPA and subprocessor list cover the vendor.
• That the integration passed a vendor security review.
• That you have tested the flow with non-sensitive data.

How Can a Business Keep a Custom GPT Private?

Use this implementation checklist to move from “probably fine” to “verified and approved.”

1. Select the correct business or enterprise plan rather than a consumer plan for any confidential use.
2. Confirm whether business content is excluded from model training by default on your exact plan.
3. Review retention and deletion terms, and set the shortest retention that meets your needs.
4. Enable SSO and identity-provider authentication, such as SAML, for all users.
5. Apply role-based access controls so people see only what their role requires.
6. Keep public and link-only sharing disabled for any confidential assistant.
7. Separate client, department, or regulated datasets into isolated assistants or tenants.
8. Review which administrators, employees, and contractors can access conversations and content.
9. Audit every connected Action, app, API, and subprocessor, and document each data flow.
10. Use least-privilege permissions and the narrowest OAuth scopes for integrations.
11. Establish content-classification rules for what may and may not be uploaded.
12. Avoid uploading categories of regulated data the platform is not contracted to handle.
13. Define and publish an employee usage policy for the assistant.
14. Test deletion and access-revocation workflows before relying on them.
15. Review logs and access periodically, and obtain security, legal, and compliance approval before production deployment.

Enterprise Privacy Evaluation Checklist

A procurement-ready scorecard. Ask every AI vendor the same questions and compare the written answers, not the sales narrative.

• Is customer content used to train or improve any model, by default or after opt-in?
• Who can access uploaded documents and knowledge?
• Who can access prompts and conversations?
• Are separate agents, projects, or tenants isolated from one another?
• What is the exact data-retention period?
• Can retention be customized to our policy?
• Can we delete and export our data on request?
• What happens to deleted information in backups, and on what timeline?
• Where is data processed and stored, and is regional residency available?
• Which subprocessors are used, and is the list published?
• Is encryption applied in transit and at rest, and how are keys managed?
• Is SAML SSO available on our plan?
• Are role-based permissions supported?
• Are audit logs available, and how detailed are they?
• Is a Data Processing Agreement available to us?
• Is there a current SOC 2 Type II report, and what is its scope?
• Does the vendor support GDPR-related contractual requirements?
• What information is transmitted to model providers or other third parties?
• Are external integrations governed by separate agreements?
• What security documentation is available in the Trust Center?
• What breach-notification commitments are in the contract?

When Should You Use a Private Enterprise AI Platform?

A dedicated enterprise platform tends to fit better than a consumer assistant when the work involves confidential content, source-cited answers, controlled deployment, or multi-team governance. Common cases include:

• Internal knowledge search across policies and procedures.
• Customer-support knowledge bases that must cite approved articles.
• Employee policy and onboarding assistants.
• Legal and compliance research that needs traceable sources.
• Proprietary product documentation for support or field teams.
• Member-only content for associations and membership organizations.
• Agency client knowledge kept strictly separate per client.
• Government information services with public-record accuracy.
• Regulated or confidential workflows in finance, healthcare, and the public sector.
• Answers that must be source-cited and verifiable.
• Controlled website or in-app deployments.
• Multi-team governance with role-based access and analytics.

Choosing an enterprise platform does not by itself make a regulated use case compliant. It gives you the controls and documentation to build toward compliance, which your own review still has to confirm.

Real-World Enterprise AI Examples

These examples illustrate enterprise deployment, controlled knowledge access, and source-grounded assistance. They are not, on their own, proof of security or compliance, and one customer’s configuration does not transfer automatically to another.

• Ontop, a global payroll and Employee of Record company, built a CustomGPT.ai agent named “Barry” inside Slack to answer sales-team questions from internal compliance and payroll documentation, with a citation on every answer. Ontop reports 130 legal-team hours saved per month, response time cut from 20 minutes to 20 seconds, more than 400 complex queries handled monthly, and a 60% acceptance rate in a compliance-sensitive setting. See the Ontop case study.
• Bernalillo County (BernCo), a US county assessor’s office, deployed multiple CustomGPT.ai assistants across its website and support channels, grounded in county documentation and public records. BernCo reports roughly $108,000 in net savings over 18 months, an 80% lower cost per interaction, and a 4.81x return on its spend. See the BernCo case study.
• GEMA, a European rights-management organization, used CustomGPT.ai to scale member support without adding headcount and reports more than 6,000 working hours saved. It is a useful example of a membership organization deploying source-grounded assistance across a large member base.

Browse more in customer stories and testimonials.

Frequently Asked Questions

Conclusion

Custom GPTs can be private, but enterprise privacy depends on far more than whether the assistant is hidden from public view. Whether the answer is genuinely “private enough” comes down to training use, retention, data isolation, authentication, connected integrations, administrative access, your contract and DPA, and deletion rights. Evaluate each layer on its own terms, verify the specifics for your exact plan and endpoints, and get security, legal, and compliance approval before production.

For organizations that need source-grounded AI assistants built strictly from approved business content, with cited answers, private-by-default agents, per-agent data isolation, SAML-based access, and SOC 2 Type II and GDPR documentation, CustomGPT.ai is built for that shape of problem. Confirm the current details for your use case in the documentation and the Trust Center.

Start a free trial to build an agent on your own content, or talk to enterprise sales and review the Security and Trust Center to walk through data handling with your security and compliance teams.