ChatGPT at Work: What Employees Can Do — and What Companies Should Put in Policy

If you walk into a mid-sized company in Linz today and ask how many employees regularly use ChatGPT, Claude, or Microsoft Copilot, you usually get two answers: management estimates 10 to 20 percent — and the reality is 60 to 80 percent. We see this in almost every audit we run for clients at psquared. The workforce is far ahead of the company on AI. What's missing is clear rules.

The problem isn't that employees use AI. The problem is that they use it without guardrails. When someone in sales pastes a customer record into ChatGPT to draft a personalised email, that's a GDPR violation in many cases. When someone in engineering uploads source code from a confidential project to a public AI, that's a trade-secret exposure. When HR uses ChatGPT to pre-screen CVs, that may fall under the EU AI Act's high-risk provisions once they activate in August 2026. Three real risks, three separate legal frames, and most Austrian SMBs have no written answer to what is and isn't allowed.

This piece is a practical guide — not a list of banned tools, but an honest description of what employees can usually do safely, where the red lines actually are, and what a policy looks like that people will actually follow.

The Five Most Common Mistakes in Practice

In the workshops we run with teams in Upper Austria, the same patterns keep appearing. They're not malicious — they're almost always the result of missing information.

Pasting personal data into public AI. The classic: an employee copies a list of customers, suppliers, or colleagues into ChatGPT to generate something — an analysis, an email template, a summary. As soon as that data goes to OpenAI and lands in their standard processing, it's a GDPR transfer to a third-country processor without documented legal basis. Statutory penalty ceiling: up to 20 million euros or 4 percent of group revenue. For an SMB, realistically more in the five- to six-figure range — but that's usually enough.

Uploading trade secrets. Source code, design data, internal strategy documents, financial data. Once that content lands in model training, confidentiality is effectively gone. Even with OpenAI's opt-out mechanisms, you cannot reliably prove that your data didn't end up shaping the model. In a dispute, that's an evidentiary problem you don't want.

Using AI output without checking it. Second classic: someone has ChatGPT write a customer email and sends it without review. If the AI inserted a wrong figure or invented a detail (which it reliably does), the company is still liable for the statement. There's no clause that gets you out of liability because "the AI wrote it."

Letting AI make decisions without human oversight. When HR uses ChatGPT to pre-sort applications without a human making the final call, that falls under Article 6 of the EU AI Act starting August 2026 — high-risk systems in employment context. That requires documented risk analyses, transparency to applicants, and human oversight. Most SMBs are not aware of this threshold.

Pasting credentials and access keys into chat. Sounds absurd; happens constantly. Someone wants to debug a script and pastes the whole file including an API key into the chat. Those keys land in logs, in training data, in chat histories. Treat anything you put into a public AI like an email to a stranger — you can't pull it back.

What Employees Can Usually Do Safely

Just as important as the red lines: there's a wide range where workplace AI use is productive and uncritical. A policy that bans everything gets ignored — and in practice is counterproductive.

Anonymised or general tasks are unproblematic. An employee who asks ChatGPT to draft a professional rejection letter to a candidate — without the name, without the company, without identifiable details — is operating in a space that breaches neither GDPR nor trade-secret protection.

Public research is allowed. "What are the most important changes in Austrian building regulations in 2026?" is a question anyone could ask on the internet. It doesn't become a privacy issue because it goes through an AI tool.

Text and language refinement is low-risk, as long as the content itself isn't confidential. You can have ChatGPT proofread a technical documentation for a product that's already publicly marketed. You can't do that with an internal strategy paper.

Idea generation, brainstorming, and template creation are harmless — as long as you don't write company data into the prompt. "What are ten arguments for a four-day work week in a trades business?" works. "Please read our employee review notes and tell me how to terminate Mr. Huber" does not.

Programming help for non-confidential code — for an open-source project, a tutorial example, or generic snippets — is industry-standard practice. The difference is not in the tool, but in the data that goes into it.

What a Sensible Policy Actually Looks Like

Most policies we see in practice fail for one of two reasons: either they're a 20-page legal document nobody reads, or they're a ban without justification that gets ignored. A workable AI policy for an Austrian SMB fits on two pages and answers four questions.

Which tools are approved? List the AI tools the company explicitly allows. Ideally with a clear default — say, Microsoft Copilot under your M365 subscription, or a GDPR-compliant provider with EU hosting. Most problems happen because employees use tools that aren't approved, because no alternative was provided.

What data is allowed in? A simple three-tier classification is enough. "Public information": anywhere. "Internal non-personal information": approved tools, not public models. "Personal data and trade secrets": nowhere except explicitly approved enterprise solutions with a data processing agreement in place. This classification is trivial to remember and covers 95 percent of cases.

Who is responsible when AI output is used? The answer must be: the human who sends or implements the output. Period. If a sales rep sends an AI-generated email to a customer, they're responsible for the content — not the tool. This clarification is legally important and simultaneously reinforces the duty to review.

What must be reported? When someone notices that customer data accidentally ended up in a public tool, it must be clear who they report it to and within what timeframe. For GDPR breaches you have 72 hours to notify the data protection authority — the clock starts the moment someone inside the company learns of it. A low-friction internal reporting channel saves you six-figure penalties when something does happen.

Which Tools Actually Make Sense for Austrian SMBs

Most companies we work with end up in one of three setups — depending on size, data sensitivity, and existing contracts.

For Microsoft 365 customers, Copilot is usually the simplest answer. The data stays in the M365 tenant, the data processing agreement is already part of the Microsoft contract, and the tool integrates seamlessly into Word, Excel, Outlook, and Teams. The licensing cost (around 30 euros per user per month on the business tier) is real, but for the compliance relief it usually pencils out.

For companies with higher data sovereignty requirements — healthcare, legal, public administration — EU-hosted models are a serious alternative. Providers like Aleph Alpha (Germany), Mistral (France), or specialised Austrian hosters offer LLMs with documented EU data residency. Model quality lags behind US frontier models, but it's more than sufficient for most business use cases.

For very small businesses without an IT department, ChatGPT Team with active training opt-out and a signed Data Processing Addendum can be a pragmatic solution — provided the internal data classification rules are followed. The legal cleanliness is thinner than Copilot, but better than the status quo (everyone uses their own account without oversight).

General principle: the tool that's actually used beats the theoretically perfect tool nobody touches. If your workforce is already using ChatGPT and you bring it into a contractually clean frame, that's a bigger win than searching for the legally flawless system.

What Happens If You Do Nothing

The honest answer: probably nothing, at first. Most Austrian SMBs still have no ChatGPT policy and nothing concrete has happened to them yet. But three things change over the next 12 to 24 months.

First, the first data protection authority rulings will land that explicitly sanction ChatGPT use without a data processing agreement. Initial cases already exist in Italy and Germany — Austria typically follows with a 12 to 18 month delay.

Second, the operational provisions of the EU AI Act activate from August 2026. Anyone who hasn't at least documented which AI systems are in use and which of them are risk-bearing runs the risk of standing empty-handed at the first authority enquiry.

Third — and this is probably the most expensive effect — you lose control of your own data and processes. When employees use AI without guidance, shadow workflows emerge that aren't documented, whose quality no one reviews, and that create gaps in any compliance audit. A policy isn't a ban — it's the prerequisite for AI to be used productively and traceably in the company.

Practical Next Step

If you're reading this as the managing director of an SMB in Upper Austria and you have no ChatGPT policy: block 90 minutes on your calendar and write the first version yourself. Four pages is enough. Data classification, approved tools, responsibilities, reporting line. Then have a lawyer or data protection officer review it — but don't start with them, that's the most common reason these policies never get finished.

If you need support — the ki-linz.at community hub runs regular workshops on this topic, and psquared works with SMBs on rolling out workable AI policies. But the most important work happens internally: the workforce wants clarity. Most uncertainty isn't created by the tool itself but by the absence of any rule.

AI at work isn't a theoretical risk anymore. It's the reality in every office, every shop floor, and every back office in Linz. The question isn't whether your employees use ChatGPT — it's whether they use it in a way that won't land on your desk a year from now. A proper policy costs 90 minutes of work. Not having one can run into six figures.