The Small Business Guide to Keeping Company Data Out of Public AI

Someone in your business is almost certainly pasting sensitive company information into a public artificial intelligence tool right now. It is rarely done with bad intent. An operations manager wants to summarise an ugly dispute with a supplier. An account executive wants to polish a proposal containing detailed margin breakdowns. An HR manager wants to draft a disciplinary letter using bullet points from a confidential staff file.

In each case, the employee is simply trying to get their work done faster. However, if they are using the default, consumer-facing versions of tools like ChatGPT, Claude, or Gemini, they may be handing your intellectual property, client records, and commercially sensitive numbers over to a third party to train their future systems.

For businesses with between 10 and 250 staff, the risk is rarely an international cyberattack. The real risk is quiet, accidental data leakage driven by good intentions and poor tooling. Here is a practical guide to how data flows into public models, what the actual risks are, and how to evaluate safer alternatives.

Where Public AI Models Store Your Data (And Why It Matters)

When an employee opens a browser tab and types a prompt into a free AI tool, that prompt does not vanish after the answer is generated.

Under the default terms of service for most consumer AI platforms, the prompts and attachments submitted by users are retained on the provider's servers. More importantly, they can be reviewed by human annotators to monitor system performance, and they are routinely fed back into the training dataset to teach the model how to answer future queries.

This creates two distinct operational problems:

• Human review: Most AI vendors employ human contractors to read random samples of user prompts and system responses. If your employee pasted an NDA-protected contract or an unannounced restructuring plan, an unknown third party may read that exact text.
• Model memorisation: Large language models can occasionally regurgitate verbatim snippets of their training data when prompted in specific ways. While the odds of a competitor querying ChatGPT and receiving your exact pricing model are relatively low, the legal reality is that the data has left your control and entered a public repository.

If your staff use free consumer accounts, you have no data processing agreement, no guarantee of where the data resides geographically, and no technical mechanism to delete it once it has been absorbed into a training cycle.

The Three Real Risks for Growing Businesses

Too much commentary on AI safety focuses on sci-fi scenarios. For an established SME, the liabilities are much more grounded in daily commerce.

First is client contract violation. Most B2B service contracts, supplier frameworks, and non-disclosure agreements contain strict clauses about third-party data sharing. If you paste client-identifiable data, technical schematics, or financial statements into a consumer AI platform, you have likely breached those agreements. If a client conducts an annual data governance audit, you will struggle to explain how their data was protected.

Second is regulatory exposure, specifically under GDPR and data privacy laws. Entering personally identifiable information (PII) - such as employee medical notes, customer home addresses, or candidate CVs - into an unsecured public model is an unsanctioned data transfer. You cannot service a "right to be forgotten" request if that individual's history now sits inside a closed model training pipeline.

Third is intellectual property dilution. If your business develops proprietary methodologies, bespoke codebases, or unique manufacturing formulas, feeding those assets into public models weakens your claim to trade secrets. Once proprietary knowledge is public, keeping it defensible becomes significantly harder.

Evaluating the Options: Consumer Free vs Consumer Paid vs Enterprise Architecture

To stop staff from using consumer tools dangerously, you must provide them with safe alternatives that deliver equal or better productivity. You have three main tiers to evaluate:

### 1. Consumer Free Tiers - Examples: Standard ChatGPT, Claude Free, Google Gemini standard. - Terms: Default settings permit the vendor to train on your prompts and files. - Control: Zero administrative oversight. No central user management, audit logs, or retention policies. - Verdict: Unsuitable for any business handling client data, personnel files, or proprietary IP.

### 2. Consumer Paid Tiers ("Pro" Accounts) - Examples: ChatGPT Plus, Claude Pro billed to individual credit cards. - Terms: Often allow users to manually toggle off model training in their personal settings. - Control: Weak. The business has no centralised dashboard to verify whether staff have actually turned training off. When an employee leaves, their account and prompt history leave with them. - Verdict: A stopgap that gives a false sense of security while maintaining zero institutional governance.

### 3. Tenant-Isolated Business AI (Commercial Data Protection) - Examples: Microsoft 365 Copilot, Azure OpenAI Service, ChatGPT Enterprise. - Terms: Explicit legal agreements stating that user prompts, responses, and referenced files are never used to train the base model. - Control: Total. Access is tied to corporate identity (such as Microsoft Entra ID). The data remains within your existing security perimeter, inherits your existing permissions, and stays within your specified geographic region. - Verdict: The only defensible standard for a business with 10 to 250 staff.

If your organisation already runs on Microsoft 365, tools like Microsoft Copilot operate under Commercial Data Protection. This means that when an employee signs in with their corporate account, their data is encrypted, prompts are discarded after processing, and the system respects the access controls you already have in place for SharePoint and OneDrive.

Practical Guardrails You Can Implement This Week

Securing your business data does not require a six-month IT project. You can implement three sensible measures immediately:

• Publish a one-page Acceptable Use Policy. Keep it plain-English. Ban the entry of customer PII, internal financials, and client-confidential text into free browser tools. List the specific platforms that are approved for company work.
• Deploy an approved corporate option. Telling staff "do not use AI" does not work; they will simply hide their usage on personal phones. Provide a managed, secure tool that allows them to do their work safely.
• Audit current file permissions. Enterprise tools like Copilot inherit your internal access rights. If your internal folders are poorly organised - such as leaving director salaries or board minutes open to the whole company - an enterprise AI tool will surface that data to anyone who asks. Clean up your permissions before switching internal AI tools on.

Next Steps for Leadership

The goal is not to stop your team from using artificial intelligence. Used correctly, it eliminates administrative drag and sharpens day-to-day operations. The goal is to make sure your efficiency gains do not come at the cost of your commercial privacy.

Start by asking your team what tools they are currently using to get their work done. Take an inventory of those platforms, check their data training terms, and look at moving your organisation onto a managed, business-tier solution with commercial data protection built in.