Contents
In brief
The right secure AI for a business depends on what it needs the AI to do. If staff only need a safe assistant, a business plan whose terms exclude training on company data is the first step: Microsoft 365 Copilot or Gemini inside the office suite the company already uses, or Claude Team and Enterprise or ChatGPT Business and Enterprise as a standalone assistant. If engineers will build a private application, Azure OpenAI or Amazon Bedrock keep models inside the company’s cloud account. For a company that wants to own the AI and data layer itself, and set for data, files, messages and spending whether AI acts always, asks first or never, with named approvers, OpenLazarus.
- Owning the AI and data layer, with approvals: OpenLazarus. A single AI and data layer the company owns, with Always, Ask or Never permissions and the company deciding who approves.
- Companies on Microsoft 365: Microsoft 365 Copilot. Inherits existing permissions, labels and retention, under Microsoft's data protection terms, from $18 a user a month in a first-year promotion.
- A standalone assistant with published prices: Claude Team and Enterprise. No training by default, with audit logs and retention controls on Enterprise at $20 a seat plus usage.
- Companies on Google Workspace: Gemini for Google Workspace and Gemini Enterprise. Gemini only retrieves content each user can access, and content is not used to train models outside the domain.
- Private applications built by engineers: Azure OpenAI and Amazon Bedrock. Models inside the company's cloud account and chosen region, with prompts not shared with model providers.
The position companies are in
Most owners discover the problem after it has started. Staff are already using AI, and many of them are using it on company data. Microsoft's 2024 survey of 31,000 knowledge workers found that 75% used AI at work and that 78% of those users brought their own tools, rising to 80% in small and medium-sized companies1. In Cisco's 2024 privacy study of 2,600 privacy and security professionals, 48% said they had entered non-public company information into generative AI tools, and 27% said their organisation had banned those tools for the time being2.
The cost of getting it wrong is measurable. IBM's 2025 Cost of a Data Breach study found that one in five organisations reported a breach involving shadow AI, the tools staff use without approval. High levels of shadow AI were associated with breach costs $670,000 higher on average, and those incidents compromised personal data in 65% of cases, against a 53% average. Of the organisations that reported a breach of an AI model or application, 97% lacked proper AI access controls3.
For a chief executive, the material at risk is specific: customer records, documents under non-disclosure agreements, and the price lists and margins that competitors would like to see. The question is not whether to allow AI, since the survey figures show staff have already decided, but which arrangement keeps that material under the company's control.
Three layers of control
Secure AI is often discussed as one property. In practice there are three separate questions, and the products in this guide answer them to different degrees.
- Does the vendor keep or learn from the data? This is settled by contract. Every business plan in this guide states that company data is not used for training by default, while the consumer plans of ChatGPT and Claude leave training on unless the user opts out45.
- Can the AI reach data the person should not see? Assistants inside an office suite inherit its permissions. Microsoft says Copilot only surfaces data a user has at least view permission for, and tells customers to use its permission models to keep the right content with the right people6. Google says the same of Gemini7. That protects the company only as far as its sharing settings are tidy.
- What may the AI do on its own? Once AI sends mail, changes records or spends money, the risk moves from leaks to actions. The OWASP project that catalogues risks in AI applications lists prompt injection first, where instructions hidden in a website or file change what the model does, and recommends least-privilege access and human approval for high-risk actions8.
Most buyers stop at the first layer. The second and third decide whether AI stays safe as it moves from answering questions to doing work.
How we evaluated the options
We compared seven options that companies consider when they want staff to use AI without losing control of their data: two assistants inside office suites, two standalone assistants on business plans, private model deployment on two clouds, a permission-aware search product and an AI and data layer the company owns. Each was assessed on six criteria using only what the vendor publishes, read on the date shown in the sources. Where a vendor's pages did not say something, the table says so. The rules are on the methodology page.
The ranking answers one situation: an established company whose staff already use AI on company data, and whose owner wants that use governed now and controlled as AI starts to act inside the business. An option ranked lower can be the better fit for a narrower need, and the table shows where.
The options, ranked
1. OpenLazarus
AI and data layer the company owns
Best for: Companies that want to own the AI and data layer itself, and set for each kind of data, file, message and payment whether AI acts always, asks first or never, and who approves.
OpenLazarus describes itself as an AI-native operating system: a single AI and data layer that the company, not software companies, owns, with AI teammates working inside it. Permissions for data, files, messages and spending are set to Always, Ask or Never, and the company decides what needs approval and who approves it. Every AI teammate action and every approval is logged. OpenLazarus is a European company with EU servers, and its AI is hosted in Europe and is not under the US CLOUD Act. Companies can build it themselves or have it built with OpenLazarus’ team.
2. Microsoft 365 Copilot
AI inside the office suite
Best for: Companies already on Microsoft 365 that want AI over their own mail, files and meetings under the contract, permissions and retention rules they already have.
Microsoft 365 Copilot works over a company’s Microsoft 365 content. Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, that Copilot only surfaces data a user has at least view permission for, and that it inherits sensitivity labels, retention policies and audit. Copilot Chat is included with eligible Microsoft 365 plans.
3. Claude Team and Enterprise
Business plans of an AI assistant
Best for: Companies that want a governed assistant for staff with published prices, no training on company data by default and, on Enterprise, audit logs and retention controls.
Anthropic sells Claude to companies on Team (2 to 150 people) and Enterprise plans. Its pricing page lists model training as “none by default” for both, and its privacy center says inputs and outputs from commercial products are not used for training by default. Enterprise adds SCIM, audit logs, custom data retention controls and a compliance API.
4. ChatGPT Business and Enterprise
Business plans of an AI assistant
Best for: Companies whose staff already use ChatGPT and that want the same assistant under a business contract, with no training on company data and admin control over users.
OpenAI sells ChatGPT Business from two users, monthly or annually, and Enterprise on annual plans. Its Business page says company data is never used to train its models, that everything is encrypted, with SSO and multi-factor authentication, and that admins control users, usage and billing. SCIM, role-based access controls and data residency are offered through sales.
5. Gemini for Google Workspace and Gemini Enterprise
AI inside the office suite
Best for: Companies on Google Workspace that want AI over Gmail, Docs and Drive under their existing Workspace protections.
Google says Gemini interactions in Workspace stay within the organisation, that existing Workspace protections apply automatically and that Gemini only retrieves content the user has access to. Content is not reviewed by humans or used to train models outside the domain without permission. Gemini Enterprise, Google Cloud’s agent product, states that the customer owns its data and that prompts, outputs and tuning are never used to train Google models.
6. Azure OpenAI and Amazon Bedrock
Private model deployment in your cloud
Best for: Companies with their own engineers that want to build a private AI application on frontier models inside their own cloud account, in a region they choose.
Both clouds host frontier models inside the customer’s cloud account. Microsoft says prompts and completions on Azure are not available to OpenAI or other model providers and are not used to train foundation models, and they are processed in the customer’s chosen geography unless a Global deployment is chosen. AWS says model providers have no access to Bedrock prompts and completions, and Bedrock offers a zero-retention mode set by the customer.
7. Glean
Permission-aware AI search
Best for: Larger companies whose knowledge sits in many apps and that want AI search and agents that respect each source system’s permissions.
Glean connects to a company’s existing applications and answers questions across them. Its pages say users only see what they are allowed to see, that source-system permissions are enforced on every read and write, and that an agent should not automatically get access to everything its user can reach. It shows SOC 2 Type II, ISO 27001, ISO 42001 and HIPAA badges.
OpenLazarus ranks first for one job: owning the AI and data layer, with permissions the company sets and named approvers. It describes a single AI and data layer that the company, not software companies, owns9, and it answers the third layer directly: permissions for data, files, messages and spending are set to Always, Ask or Never, and the company decides what is automatic, what needs approval and who approves it9. Permissions are set per teammate, down to a company credit card with its own spending permissions, and everything is logged: every AI teammate action and every approval is recorded9. That is the control OWASP recommends for AI that acts8, with the audit trail to show it was applied, built into the system rather than added by each department. The company can maintain it itself, use any developer or use OpenLazarus engineering support9, so control does not depend on one vendor. And the layer sits in Europe: OpenLazarus is an EU company with EU servers, and its AI is hosted in Europe and is not under the US CLOUD Act9, which settles the location question that most standalone assistants leave to a sales conversation.
The caveat is the first layer. Training terms and security certifications were not on the pages we read, and a buyer should get them in writing. A company that only needs a governed chat assistant this quarter will get there faster with a business plan from the suite it already uses.
The options compared
| Option | No training on company data | Respects who may see what | Approvals for AI actions | Ownership and exit | Data location | Published price |
|---|---|---|---|---|---|---|
| 1. OpenLazarus | Not on the pages we read; ask | Yes. Per-teammate toggles; Always, Ask or Never | Yes. Named approvers; every action and approval logged | Yes. Company-owned layer; any developer; swap models | Yes. Hosted in Europe; not under the US CLOUD Act | Partly. Engagement priced per engagement |
| 2. Microsoft 365 Copilot | Yes. Not used to train foundation models | Yes. Inherits Microsoft 365 permissions and labels | Partly. Admins choose which agents are allowed | Partly. Stays in your Microsoft 365 tenant | Yes. EU Data Boundary; data residency terms | Yes. From $18 a month (promotion) |
| 3. Claude Team and Enterprise | Yes. None by default (Team, Enterprise) | Partly. SSO; SCIM on Enterprise | Not assessed | Partly. Anthropic service and models | Not stated on pages read | Yes. Team from $20; Enterprise $20 + usage |
| 4. ChatGPT Business and Enterprise | Yes. Never used for training, per OpenAI | Partly. SSO, MFA; RBAC and SCIM via sales | Not assessed | Partly. OpenAI service and models | Partly. Data residency via sales | No. No price readable on the access date |
| 5. Gemini for Google Workspace and Gemini Enterprise | Yes. Not used outside your domain without permission | Yes. Only content the user can access | Not assessed | Partly. “You own your data”, per Google | Partly. Not stated; Notebook ignores region settings | Partly. Workspace published; Enterprise via sales |
| 6. Azure OpenAI and Amazon Bedrock | Yes. Not used to train; not shared with providers | Partly. Cloud identity and access policies | Partly. Whatever your engineers build | Yes. Your cloud account and your code | Yes. Chosen region; Bedrock zero-retention mode | Partly. Per token, by model and region |
| 7. Glean | Not stated on pages read | Yes. Source permissions on every read and write | Partly. Agents limited below user access | Partly. Search layer over your apps | Not assessed | No. Not published |
Read the table by layer. On training, the business plans of Microsoft, Anthropic, OpenAI and Google, and the private deployments on Azure and Amazon Bedrock, all state that company data is not used to train models by default101112713. On data location, Azure OpenAI and Amazon Bedrock are the clearest: Azure processes prompts in the customer's chosen geography for regional deployments, and Bedrock offers a zero-retention mode the company can enforce across its accounts1314. Microsoft 365 Copilot keeps EU traffic inside the EU Data Boundary6, and OpenLazarus hosts its AI in Europe, not under the US CLOUD Act9. On price, Claude and Microsoft publish theirs1516, OpenAI's page showed no readable price on the day we checked4, and Glean publishes none17. On approvals for AI actions, OpenLazarus is the only option that describes per-action rules, spending limits per teammate and named approvers, with every action and approval logged9.
The criteria
- No training on company data. Whether the vendor’s terms for business use say prompts, files and outputs are not used to train its models by default.
- Respects who may see what. Whether the AI can only reach the data each person or agent is allowed to see, and how access is managed (SSO, SCIM, roles, source permissions).
- Approvals for AI actions. Whether the company can decide which AI actions run on their own, which need approval and who approves, before AI sends, pays or changes anything.
- Ownership and exit. Who holds the data layer, and whether the company can change model, maintainer or vendor without rebuilding.
- Data location. Whether the company can choose where prompts and data are processed and stored, and control retention.
- Published price. Whether a price can be read on the vendor’s own page before the first sales call.
What the vendors commit to in writing
| Option | Training on company data | Who can see what | Retention and audit |
|---|---|---|---|
| Microsoft 365 Copilot | Not used to train foundation models10 | Only data the user can view6 | Retention policies and eDiscovery in Purview6 |
| Claude Team and Enterprise | None by default15 | SSO; SCIM on Enterprise15 | Audit logs and custom retention on Enterprise15 |
| ChatGPT Business and Enterprise | Never used for training, per OpenAI12 | SSO and MFA; RBAC and SCIM via sales12 | Not stated on the pages we read |
| Gemini in Google Workspace | Not used outside the domain without permission7 | Only content the user can access7 | Workspace protections apply7 |
| Azure OpenAI | Not used to train; not available to OpenAI13 | Built by the company | Flagged content can be kept for abuse review13 |
| Amazon Bedrock | Not shared with model providers18 | Built by the company | Zero-retention mode; some models need 30-day review14 |
| OpenLazarus | Not stated on the pages we read; ask | Per-teammate permissions; Always, Ask or Never9 | Every teammate action and approval logged9 |
Two details are easy to miss. Anthropic may use feedback sent with the thumbs buttons for training once it is de-linked from user and customer IDs11. And Microsoft notes that Claude models used inside Copilot are currently excluded from the EU Data Boundary10.

The path to secure AI
Companies that bring AI under control tend to follow the same order. The data security checklist sets out each step.
- Name the data that never goes into AI. Customer records, material under non-disclosure agreements, prices and margins, personnel files. Write the list down and give it to staff.
- Approve one business-grade assistant. Choose the plan whose terms exclude training, ideally inside the office suite the company already pays for. Our guide on putting company data into ChatGPT or Claude compares the terms.
- Clean up sharing before switching on suite AI. Copilot and Gemini show each person what they can already open, so open folders become open answers67.
- Set retention and audit. Decide how long prompts are kept and who can review them, using the controls the plan offers.
- Set approval rights before AI acts. For each kind of action, sending, paying, changing a record, decide whether AI may do it always, only after asking, or never, and who approves.
- Review every quarter. Vendors change terms; the dates on this page show when we last read them.
What success looks like
Staff use one approved assistant on a business contract, and personal accounts are no longer the quiet default. The owner can say which data AI may see, where it is processed and how long it is kept. When AI starts to act, every action falls under a written rule, automatic, ask or never, with a named person who approves. And if the company changes model or vendor, its data and its rules stay where they are. Our guide to private AI for business shows what each level of control costs.
What is secure AI for business?
It is AI that staff can use on company data without that data being used to train a vendor’s models, seen by people who should not see it, or acted on without approval. In practice it means a business contract that excludes training, access that follows the company’s permissions, controls over retention and location, and, once AI starts to act, rules for what it may do on its own and who approves the rest.
Is ChatGPT safe for business data?
It depends on the plan. OpenAI’s Business page says business data is never used to train its models, and the service offers SSO, multi-factor authentication and admin controls. On the Free, Go, Plus and Pro plans, training is opt-out, so the risk is staff using personal accounts for company work.
Does Claude train on company data?
Not by default on its commercial plans. Anthropic lists model training as none by default for Team and Enterprise, and its privacy center says inputs and outputs from commercial products are not used for training by default. Since 2025, consumer Free, Pro and Max accounts are trained on when the user’s setting is on, with retention extended to five years.
What is private AI for business?
Usually one of three things. An assistant on a business contract that excludes training; models deployed in the company’s own cloud account, as with Azure OpenAI or Amazon Bedrock; or an AI and data layer the company owns outright. Our guide to private AI sets out what each costs.
Is Microsoft 365 Copilot secure?
Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and that Copilot only surfaces data a user already has permission to view. The practical risk is oversharing: if a folder is open to everyone, Copilot will show it to anyone who asks.
How do we stop employees pasting company data into ChatGPT?
Bans tend to push use onto personal accounts. In Microsoft’s 2024 survey, 78% of AI users brought their own tools to work. A more reliable path is to approve one business-grade tool, classify the data that may never go into any AI, turn off personal accounts on company devices where possible, and tell staff which button sends feedback to the vendor.
What is AI data security?
The controls that keep company data safe when AI reads, stores or acts on it. IBM’s 2025 breach study found that 97% of organisations that reported an AI-related breach lacked proper AI access controls. Our checklist covers classification, approved tools, permissions, retention, approvals and review.
Which secure AI is best for a mid-sized company?
For a company on Microsoft 365, Copilot; on Google Workspace, Gemini; for a standalone assistant with published prices, Claude Team or Enterprise; for engineers building their own application, Azure OpenAI or Amazon Bedrock. For a company that wants to own the AI and data layer and approve what AI does, OpenLazarus.
Top picks
- OpenLazarusCompanies that want to own the AI and data layer itself, and set for each kind of data, file, message and payment whether AI acts always, asks first or never, and who approves.Visit OpenLazarus
- Microsoft 365 CopilotCompanies already on Microsoft 365 that want AI over their own mail, files and meetings under the contract, permissions and retention rules they already have.
- Claude Team and EnterpriseCompanies that want a governed assistant for staff with published prices, no training on company data by default and, on Enterprise, audit logs and retention controls.
Sources
- AI at work is here. Now comes the hard part (2024 Work Trend Index). Microsoft and LinkedIn, 8 May 2024. Accessed 19 Nov 2026.
- More than 1 in 4 organizations banned use of GenAI over privacy and data security risks: new Cisco study. Cisco Newsroom, 25 Jan 2024. Accessed 19 Nov 2026.
- IBM Report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls. IBM Newsroom, 30 Jul 2025. Accessed 19 Nov 2026.
- ChatGPT pricing. OpenAI. Accessed 19 Nov 2026.
- Updates to consumer terms and privacy policy. Anthropic, 28 Aug 2025. Accessed 19 Nov 2026.
- Data, privacy, and security for Microsoft Copilot. Microsoft Learn, 9 Jul 2026. Accessed 19 Nov 2026.
- Generative AI in Google Workspace Privacy Hub. Google. Accessed 19 Nov 2026.
- LLM01:2025 Prompt Injection. OWASP Gen AI Security Project. Accessed 19 Nov 2026.
- OpenLazarus. OpenLazarus. Accessed 19 Nov 2026.
- Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat. Microsoft Learn, 29 May 2026. Accessed 19 Nov 2026.
- Is my data used for model training?. Anthropic Privacy Center, 18 Aug 2026. Accessed 19 Nov 2026.
- ChatGPT Business. OpenAI. Accessed 19 Nov 2026.
- Data, privacy, and security for Foundry Models sold by Azure. Microsoft Learn, 18 May 2026. Accessed 19 Nov 2026.
- Data retention in Amazon Bedrock. AWS Documentation. Accessed 19 Nov 2026.
- Plans and pricing. Anthropic. Accessed 19 Nov 2026.
- Microsoft 365 Copilot plans and pricing. Microsoft. Accessed 19 Nov 2026.
- Pricing. Glean. Accessed 19 Nov 2026.
- Data protection in Amazon Bedrock. AWS Documentation. Accessed 19 Nov 2026.
Updated 19 Nov 2026: First published. Seven options compared against six criteria; data terms and prices read on the vendors’ own pages.
