IA que trabaja

AI Usage Policy in Your Business: Template and Key Dates of the AI Act

An AI usage policy is the internal document that tells your team which tools are allowed and what data they can share. It helps you avoid data protection fines and leaks of confidential information.

By David García Published: 9 min read

An AI usage policy is the internal document that tells your team which tools are allowed and what data they can share. It helps you avoid data protection fines and leaks of confidential information due to carelessness. In the end, it’s about putting some basic rules in writing so people can work with peace of mind.

Why do you need an AI usage policy even if you’re a small business?

You might think that regulating artificial intelligence is just for multinationals or big corporations. The reality is quite different. Today, 75% of knowledge workers already use AI tools in their daily work.

I see this all the time when I do [AI consulting for businesses]: the business owner tells me they don’t use AI yet, and then it turns out half the staff has a chat window open in their browser to draft emails or summarize documents.

The problem isn’t that they use the technology—the problem is that they do it without any control. Naturally, an employee wants to get their work done faster. If you don’t provide an official tool, they’ll use their own. In fact, 78% of workplace AI users rely on personal accounts that aren’t managed by their company. This is called BYOAI (Bring Your Own AI), and it’s a huge risk for any business, whether you have five employees or five hundred.

What is shadow AI and how does it affect you?

Shadow AI—meaning unsanctioned, unsupervised AI usage—is exactly what I just described: when employees use generative AI tools that haven’t been authorized or overseen by IT. The data shows that 68% of employees admit to doing this.

This directly impacts your bottom line and your business’s security. Imagine a salesperson uploads your client database to a free chat tool to get a sales summary. That free chat could use those data to train its models. If there’s personal data in there, you’ve just caused a security breach.

67% of accesses to AI services from corporate devices happen through non-business user accounts. The average financial cost of a security breach linked to shadow AI incidents is $5.39 million—a figure that, to put it bluntly, would shut down any small or medium business.

That’s why it’s vital to provide [AI training for businesses] and explain the risks to your whole team. The point isn’t to ban things for the sake of it, but to teach people to use the technology responsibly.

Step-by-step AI usage policy template

To solve this, you need a written document. Here’s what your AI usage policy template should include, based on cybersecurity recommendations. These are straightforward sections to write.

Purpose and scope

Here, you need to specify who this internal guideline applies to. Make it clear whether it covers salaried employees, external providers, or subsidiaries. It’s also essential to define which devices are affected—both company-owned and personal devices (BYOD) that people use for work.

Inventory and whitelist of tools

In this section, you must list exactly which AI platforms are approved by IT and which are expressly blocked. If you pay for a corporate chat license, state it here. If you ban the use of free versions of image generation tools, put it in writing. Just saying “use common sense” isn’t enough.

Data classification matrix

This is the most important part of the template. You need to define, using a traffic light system, the types of information your company handles:

  • Green: Information that’s allowed and public for any prompt, meaning any instruction you give to the AI.
  • Amber: Restricted information that requires pseudonymization or anonymization beforehand.
  • Red: Information that’s strictly forbidden.

Mandatory human oversight

Also known as Human-in-the-Loop—meaning a human is involved in the process. This section bans automated decision-making about people. It requires critical review of all AI outputs to mitigate hallucinations and bias. A chat tool can draft a report, but a person decides whether to send it.

Transparency and duty to inform

European law requires you to explicitly warn third parties when they receive answers assisted or drafted by an intelligent agent. If a client is chatting with a bot on your website, they need to know it’s a bot.

Intellectual property and licensing

Here, you regulate the assignment to the company of results from AI-assisted work. You must also ban the use of prompts involving third-party creations that are copyrighted. You can’t ask an AI to copy your competitor’s logo design exactly.

Incident reporting channel

Set up a fast-track process so any employee can report accidental data leaks to the data protection officers (DPOs) or security teams. If someone slips up and pastes forbidden data, they need to know whom to notify quickly and without fear of immediate reprisals, so you can contain the problem.

Literacy and ongoing training plan

Finally, Article 4 of the European Regulation establishes mandatory AI literacy. You need to set up compulsory technical instruction programs on ethical use for all your staff.

What data should you never enter into a public chat?

As mentioned in the data matrix, there are things that simply can’t be processed by an AI model without legal backing. The Spanish Data Protection Agency (AEPD) and INCIBE (Spain’s National Cybersecurity Institute) are very clear about this.

Here’s what you’re strictly forbidden from entering into a public chat:

  • Directly identifying personal data: DNI (Spanish national ID), passports, full names, private phone numbers or emails of clients or employees. You could only do this if you have a data processing agreement (as per the GDPR) with the AI provider.
  • Special categories of personal data: Article 9 of the GDPR forbids uploading information related to health, biometrics, genetics, ideology, or religious beliefs.
  • Trade secrets or industrial property assets: Don’t upload proprietary source code (unless obfuscated), designs, formulations, or patents in development.
  • Corporate cybersecurity credentials: Passwords, private SSH or TLS keys, session tokens, and API keys for internal systems.
  • Financial and confidential data: Unpublished financial statements, merger details, balance forecasts before audit, or complete bank account numbers. You also can’t upload data covered by confidentiality agreements (NDAs) signed with clients or providers.

Let’s look at a practical example of how a simple mistake can be costly. Imagine Carlos, from HR, needs to screen 50 résumés. To speed things up, he uploads the full PDFs—with photos, DNI, addresses, and phone numbers—to a free AI tool to extract the top three profiles.

Carlos has just committed a serious violation. He’s uploaded directly identifying personal data to an external server without legal grounds and without informing the candidates. If you had your AI usage policy in place, Carlos would know that résumés are red-light data and that, before automating the process, he’d need to remove names and DNI numbers—in other words, anonymize the document.

When we want to [automate processes with AI], the first step is to clean the data and make sure you know where that information will travel.

Is it safe to use free AI tools in your business?

It depends on what you tell the tool. Many people search online for [free AI tools] and start dumping business information into them without reading the fine print.

For example, with the most well-known model, in general consumer accounts—whether free or standard—OpenAI may, by default, use the data and content entered in conversations to train future AI models. In other words, what you tell it today could appear in another user’s response tomorrow.

However, specific corporate plans like ChatGPT Business, ChatGPT Enterprise, and the API Platform, by default, exclude model training on client-processed data.

So, the free version is fine for generic brainstorming, drafting template emails, or translating public texts. But if you’re handling internal business information, you obviously need corporate plans or to use the API.

Key dates for the AI Act: adaptation timeline

The European Union has already published its Regulation on Artificial Intelligence, known as the AI Act. It’s the world’s first legal framework on this topic. It doesn’t all come into effect at once—it’s phased in.

Here’s the official timeline with the key dates of the AI Act so you can see the deadlines.

DateKey milestone of the European AI Act
July 12, 2024Publication in the Official Journal of the European Union.
August 1, 2024General entry into force of the AI Act.
February 2, 2025Start of the ban on unacceptable risk systems and mandatory AI literacy (Article 4).
August 2, 2025Obligations for general-purpose AI model providers (GPAI) and designation of authorities.
August 2, 2026Transparency obligations (Article 50). Notice of AI interaction and content labeling.
December 2, 2026Deadline for synthetic content systems created before August 2026 to comply with labeling.
December 2, 2027Application of obligations for high-risk AI systems in Annex III.
August 2, 2028Application of requirements for high-risk systems integrated into products in Annex I.

In my opinion, the date that most affects you in the short term if you’re a small business is February 2, 2025. That’s when the obligation to provide AI literacy for staff begins. You’ll need to train your workers.

That same day, unacceptable risk practices are also banned. These are very specific things like social scoring, emotion recognition in the workplace, manipulation, harmful deception, or indiscriminate scraping of the internet for facial recognition databases. These are systems that a normal business shouldn’t be using anyway, but now it’s clear they’re illegal.

Another important date is August 2, 2026. From then on, if you have a bot serving customers, you’ll need to clearly inform them that they’re interacting with an AI.

Penalties for non-compliance with the AI Act and who enforces it

If you ignore this law, the consequences can be serious. Penalties for using prohibited AI systems can reach up to 35 million euros or 7% of total worldwide annual turnover, according to Article 99 of the regulation. That’s a huge amount of money.

In Spain, the Agencia Española de Protección de Datos (AEPD, the Spanish Data Protection Agency) has the authority to sanction and intervene against prohibited AI systems that process personal data without legal grounds. They’re the ones who’ll be watching to make sure you don’t put client data where you shouldn’t.

Can you legally automate everything with AI?

I believe you can automate almost everything, but always with human oversight. As I mentioned earlier in the template section, the final decision must remain with a human.

In the [automations with n8n and Make] that I set up, we often include intermediate steps. For example, an AI can classify incoming emails, extract invoice data, and prepare a draft response. Up to that point, everything is fine and legal if you use the right tools and servers. But the final “send” or “approve payment” button must be pressed by someone on your team. That way, you comply with the human-in-the-loop rule, cut out boring work hours, and sleep easy.

In short

Implementing an AI usage policy in your business is no longer optional—it’s a basic legal and security requirement. Use the template, define your rules, and above all, train your team so they stop using personal accounts in secret.

If you want to start automating things in your business sensibly, step by step, and without breaking the rules, just join [El Atajo] and sign up. See you there.

Keep reading

Guides

AI advisors: how intelligent assistants work for businesses

An AI advisor uses data and context to recommend actions. Learn how to choose one and put it to work in your business.

Read · 10 min ·
Guides

Claude Opus, Sonnet or Haiku: practical differences and when to use each

Compare Claude Haiku, Sonnet, Opus and Fable by speed, quality and cost, with a task-by-task guide and common mistakes to avoid.

Read · 5 min ·
Guides

Multimodal AI: What It Is, Examples and How to Use It in Your Business

Everyday AI assistants can read photos, listen to audio and understand PDFs. Learn how businesses use them and what to consider before adding them to your processes.

Read · 6 min ·