Staff Use ChatGPT or Claude? What IT Does First

Most of the companies we talk to have not formally adopted AI. There is no policy, no approved tool and no budget line. Their staff are using ChatGPT, Claude and Gemini anyway: drafting emails, summarizing contracts, rewriting procedures, fixing spreadsheet formulas. Usually on a personal account, sometimes a paid one, occasionally on the company card.

That's not a scandal; it's how most office technology arrives. Email, the smartphone and the cloud file share all got used at work before there was a policy for them, while some people were still arguing they didn't belong there. AI is following the same path. You don't need to settle the debate between enthusiasts and skeptics. What you need to know is whose account the work runs on, and what happens to the data in it.

This post is for the person who will be asked to sort it out: the internal IT lead, and often the controller.

How common is it?

Every survey counts differently, so here is each one with its base:

• Microsoft Work Trend Index, 2024 (31,000 knowledge workers, 31 markets): "78% of AI users are bringing their own AI tools to work." Note the base: AI users, not all employees. The same report: "52% of people who use AI at work are reluctant to admit to using it for their most important tasks."
• KPMG and the University of Melbourne, 2025 (48,000+ people, 47 countries): 70% of employees using AI at work use free public tools; "Fifty-six percent say they have used AI tools at work without knowing if it is allowed"; and about half (48–49%) report uploading sensitive company information, such as financial, sales or customer data, into public AI tools.
• Cybernews, August 2025 (1,003 US employees): "59% of employees use AI tools that their employer has not approved." Of those, 75% said they had "shared possibly sensitive information" with them.

The exact number at your company doesn't matter much. If you have knowledge workers and no sanctioned tool, assume a meaningful share of them are using personal AI accounts for work today.

Why a personal account is the actual problem

The AI isn't the problem. The contract is. A personal ChatGPT or Claude account is an agreement between your employee and the vendor. Your company is not a party, has no admin console, cannot export the conversations for a legal hold, and cannot delete them when the person leaves. On top of that, consumer plans handle training differently from business plans:

• ChatGPT Free, Plus and Pro: OpenAI says data sharing for model training is enabled by default, and the user can turn it off.
• Claude Free, Pro and Max: since Anthropic's 2025 consumer-terms update, training is a user setting, with five-year retention for users who allow it and 30 days for those who don't.
• Business tiers (ChatGPT Business and Enterprise, Claude Team and Enterprise, Copilot with a work sign-in, Gemini for Workspace) keep business content out of model training by default or by contract, and give the company an admin.

Details, plan by plan, including which tiers can carry a HIPAA Business Associate Agreement: ChatGPT and Claude.

Step 1 — Find out what is actually in use

Count before you write any policy. Most firms already have the data:

• Firewall, proxy or DNS logs for chatgpt.com, claude.ai, gemini.google.com, perplexity.ai, copilot.microsoft.com and similar. Count distinct users and departments, not hits.
• Microsoft Defender for Cloud Apps, if licensed. Microsoft's catalog has a Generative AI category with more than a thousand AI apps, and you can filter discovered apps to it.
• Expense reports and card statements. Personal ChatGPT Plus or Claude Pro subscriptions claimed as business expenses are a reliable signal, and they are the controller's data, not IT's.

The point is to size the problem, not to discipline people. Staff who were using AI to do their jobs better are the ones whose habits the new tool needs to fit.

Step 2 — Pick the sanctioned tool your people will actually use

The most common mistake is choosing a tool for procurement reasons that nobody wants to use. If your team has spent a year in Claude, giving them something else and blocking Claude will send them to their phones. KPMG found that uploading sensitive data to public AI tools was most common in organizations that had banned generative AI, and concluded "outright bans may be ineffective."

So, vendor-neutral criteria (we compare the business plans side by side in Claude vs ChatGPT vs Copilot for business):

• Does the business tier keep your content out of model training, by contract rather than by user setting?
• Does it support single sign-on with your identity provider (Entra ID for most of our clients), so access ends when employment does?
• Can an admin see usage, set retention, and export or delete conversations?
• If you handle regulated data (PHI, CUI, card data), does that exact tier carry the agreement you need?
• Does it fit the work? Copilot is strongest inside Microsoft 365 documents and mail; ChatGPT and Claude are often preferred for long-form drafting and analysis. Many firms end up with two.

If you are a Microsoft 365 shop, check what you already have. Microsoft says "Using Copilot Chat doesn't require a Microsoft 365 Copilot license," and enterprise data protection applies when staff sign in with their work account. Microsoft tells admins to "instruct them to sign in with their work or school account." A personal Microsoft sign-in in the same app gets you back to the consumer problem.

Step 3 — Write a one-page rule, not a ten-page policy

Staff need to know three things: which tools are approved, what never goes into any AI tool, and who to ask. Put the long version, covering ownership, review cadence and incident handling, in the document IT and compliance keep. Every rule in it should map to a control you can actually operate (Step 4), or it's just a request. We've laid that out clause by clause in AI acceptable use policy: map every rule to a control. Have counsel and HR review the final wording.

Step 4 — Back the rule with controls you may already own

For Microsoft 365 tenants, these are the controls that exist today, with the licensing caveats Microsoft states:

• Block unsanctioned AI apps. In Defender for Cloud Apps, "When an app is marked as unsanctioned, it's automatically blocked across devices that are onboarded to Defender for Endpoint." Microsoft notes the block can take up to three hours and requires network protection in block mode. You can also warn and educate users rather than block.
• See sensitive data going to AI sites. Microsoft Purview's Data Security Posture Management can show, in Microsoft's example, when "a user pastes credit card numbers into ChatGPT." It requires devices onboarded to Purview, and on Chrome, the Purview browser extension.
• Stop the paste at the browser. Edge for Business can apply DLP to unmanaged AI apps on managed devices, and Microsoft says "You don't need to onboard the device into Microsoft Purview." Microsoft's supported list includes consumer ChatGPT and Gemini. Check it for the tools you care about; we did not find Claude on it. Microsoft lists this protection for unmanaged apps as a pay-as-you-go capability, not part of E5.
• Local AI agents. Intune added detection and blocking of local AI agents in preview this June. We covered it in Intune STIG baselines and shadow-AI control.

Step 5 — Move people over, then close the old accounts

1. Provision the sanctioned tool with single sign-on, and announce it as the better option, not a crackdown.
2. Ask staff to move ongoing work out of personal accounts, and to delete personal-account conversations that held client, patient, employee or financial data.
3. Stop reimbursing personal AI subscriptions once the company tool is live. Finance can enforce this one in a day.
4. Then switch the unsanctioned apps from "warn" to "block", and review discovery data monthly.

Where the controller comes in

AI use in finance (draft commentary, reconciliations, formula help) raises a separate question: how do you know a number in the close was checked by a person? COSO published guidance on internal control over generative AI in February 2026. The short version: AI use in finance should be visible, and its output should be reviewed before it reaches a journal entry. The controller's side is in AI in the finance team: a controller's first questions.

If you would rather have someone else do Steps 1–4, that is what our AI usage review is. We set up whichever tool your team prefers. If the harder question is who owns AI risk at all, because there is no security lead, that's what our virtual CISO engagement is for. This is technical and operational guidance, not legal advice.