By Rick Norberg, CEO, Vertikal6
Published November 11, 2026
Yes, your team can use ChatGPT, Copilot, Claude, and Gemini with company data when each tool runs under an enterprise account your company controls. The data going in also has to match what that account is approved to hold. Vertikal6 draws the line there for every AI rollout: company data belongs in company-governed tools, and regulated data belongs in the smaller set of tools covered by the right agreement which does not train the model.
Picture a project manager who pastes a customer contract into a personal ChatGPT account to get a quick summary. The summary is useful. The contract now sits under consumer terms between that employee and the vendor, outside your retention policy, your audit logs, and every agreement your company signed.
Personal AI accounts at work
Personal accounts are already the default for many employees. Verizon’s 2026 Data Breach Investigations Report found that 67 percent of users were accessing AI services with non-corporate accounts on their corporate devices. The same report, published in May 2026, found that 45 percent of employees are now regular users of AI on corporate devices, up from 15 percent the year before.
Those numbers describe demand. People use these tools because the tools save them time. The practical response is to give them approved tools that are just as easy to reach, and to set clear rules for what data goes where.
Enterprise accounts and consumer accounts
An enterprise account is one your company buys, administers, and signs a contract for. That contract sets how the vendor handles your prompts, files, and outputs. Your administrators control who signs in, how long conversations are kept, and who can review them.
Microsoft’s documentation for Microsoft 365 Copilot is specific on the training question. According to Microsoft Learn’s Copilot privacy documentation (last updated August 2026), prompts, responses, and data accessed through Microsoft Graph are excluded from training the foundation models behind Copilot. That commitment applies when employees sign in with their work accounts.
The same test applies to ChatGPT, Claude, and Gemini. Use the business or enterprise plan your company has contracted for, and confirm in the contract how the vendor handles training, retention, and where data is stored. A consumer account on any of these platforms carries the employee’s personal terms, and those terms govern the data the moment it is pasted in.
Permissions cleanup before a Copilot rollout
Copilot reads what your people can already open. Microsoft states in the same documentation that “Microsoft Copilot only surfaces organizational data to which individual users have at least view permissions.” Copilot follows your permissions exactly, so a salary spreadsheet shared with the whole company three years ago becomes an answer to a casual question about compensation.
Vertikal6 recommends auditing and fixing permissions before the first Copilot license is assigned. That work covers four areas:
- SharePoint sites and Teams shared with the whole organization that hold HR, finance, legal, or client records
- Sensitive files shared through links that open for anyone who has the link
- Sites and folders with no current owner
- Former employees and contractors who still hold group memberships
Sensitivity labels carry the protection forward. Microsoft’s documentation states that when content is encrypted through Microsoft Purview Information Protection, Copilot honors the usage rights granted to each user. Labeling your most sensitive files before rollout keeps those rights in force when Copilot starts pulling content into answers.
Which data can go into which AI tool?
The simplest control is a short table every employee can read in a minute. The version below is a starting point; your own table should name the specific tools your company has approved.
Data type | Approved tools | Conditions |
|---|---|---|
Public information (published web copy, press releases) | Any company-approved AI tool | Standard acceptable use |
Internal business information (meeting notes, drafts, procedures) | Microsoft 365 Copilot or Copilot Chat on a work account; other enterprise AI tools under company contract | Work account sign-in; retention set by company policy |
Confidential client data and source code | Enterprise AI tools under a signed data protection agreement | Sensitivity labels applied; data loss prevention active; access limited to the project team |
Protected health information (PHI) | AI tools from vendors with a signed business associate agreement | BAA on file before first use; minimum necessary data; logging on |
Controlled unclassified information (CUI) | AI tools inside your assessed CMMC boundary, from a cloud provider meeting FedRAMP Moderate equivalent requirements | Tool documented in your CMMC scope before first use |
Customer financial information | Enterprise AI tools from service providers under contract with written security terms | Multifactor authentication; access reviewed regularly; provider oversight documented |
Passwords, API keys, and credentials | None | Kept out of every AI tool |
How should you handle PHI, CUI, and financial data in AI tools?
Regulated data follows the same rules in an AI tool that it follows anywhere else. Each regulation already says what a vendor relationship has to include, and AI vendors fall under those rules.
Protected health information. The U.S. Department of Health and Human Services states in its HIPAA cloud computing guidance that a covered entity or business associate may use a cloud service provider for electronic PHI when it enters into a HIPAA-compliant business associate agreement with that provider. HHS also states that a provider holding encrypted PHI without the decryption key is still a business associate. For a practice or health system, that means an AI tool without a signed BAA stays off the list for patient data, however good its encryption.
Controlled unclassified information. Defense contractors work under DFARS 252.204-7012, which requires any external cloud service provider that stores, processes, or transmits covered defense information to meet security requirements equivalent to the FedRAMP Moderate baseline. A commercial AI chatbot outside that standard sits outside your boundary, and CUI pasted into it has left the environment your CMMC assessment covers.
Customer financial information. For financial institutions under the Gramm-Leach-Bliley Act, the Federal Trade Commission’s Safeguards Rule guidance requires contracts with service providers that “spell out your security expectations” and build in ways to monitor the provider’s work. The rule also requires businesses to determine who has access to customer information and to reconsider that access regularly. An AI vendor handling customer financial data is a service provider under those terms.
Vertikal6’s regulatory compliance services cover HIPAA, CMMC, DFARS, and PCI-DSS requirements, which makes AI tool approval part of the same compliance program your auditors already review.
Logging and monitoring of AI use
Logging turns an AI policy into something you can verify. Microsoft’s documentation states that administrators can use Content search or Microsoft Purview to view stored Copilot interactions and can set retention policies for them in Purview. Turn those settings on before rollout so the record exists from the first day.
Monitoring also shows where the policy is breaking down. The 2026 Verizon report found that shadow AI, meaning employees using unapproved AI tools, was the third most common non-malicious insider action in its data loss prevention dataset for 2025. Source code was the data type employees most often submitted to those external tools. Data loss prevention policies that watch for uploads to unapproved AI sites give you an early warning and a starting point for the next training session.
Training employees on AI and data
Employees follow rules they understand. Training for AI use covers three things: which tools are approved, which data classes can go into each one, and how to recognize a request that crosses the line. Short, concrete examples teach best. Show a team what a safe prompt looks like and what an unsafe one looks like, using documents from their own department. A written AI acceptable use policy gives that training its source material.
Pair AI training with ongoing security awareness training, since the same habits that stop a phishing click also stop a pasted client file. Repeat the AI portion whenever you approve a new tool or change the data table.
AI automation connected to existing systems
Custom AI automation raises the same questions at larger scale, because an integration moves data automatically and in volume. Give every automation its own service account with the narrowest permissions the workflow needs. Keep data inside systems your company already governs, and route calls to AI models through enterprise endpoints under contract. Log every call the automation makes, and review those logs on the same schedule as user activity. Teams building this kind of work can start with a workflow analysis through AI automation and optimization services.
How do you roll out Microsoft Copilot and other AI tools safely?
A safe rollout of Copilot or any enterprise AI tool follows a fixed sequence, and each step depends on the one before it.
- Inventory the AI tools employees already use, including personal accounts on company devices.
- Classify your data into the categories in the table above.
- Create a policy and the proper governance
- Select enterprise tools and sign the agreements each data class requires, including BAAs for PHI.
- Clean up permissions and apply sensitivity labels to your most sensitive content.
- Turn on logging, retention, and data loss prevention for AI activity.
- Train employees on the approved tools and the data table.
- Launch with a pilot group, review the logs, then expand.
- Review tools, permissions, and the data table every quarter.
Steps one through eight describe controls. Keeping them running requires an owner, a budget, and a way to measure results.
AI program management
elevate™ CLARITY is Vertikal6’s AI program management service that combines AI implementation with its Transformation Management Office (TMO) methodology. The CLARITY playbook integrates security, governance, quality control, and ethics into each program, starting from a discovery and current-state assessment and a strategic roadmap. It also includes enablement programs so your internal team can own the process after launch. For organizations ready to move from scattered AI use to a governed program, elevate™ CLARITY provides the structure for the controls in this post.
A client example
A quasi-governmental organization adopted an AI acceptable use policy developed with Vertikal6. The policy names the categories of tools and data that are restricted, and it defines how a team requests approval for a new tool and how a restriction is relaxed when a legitimate business need arises. Once the policy was in place, employees began submitting tickets for review and approval, a type of request the team had never received before, which shows the approval path is clear and in active use.