AI in work teams: rules, permissions and data control

Your team already uses AI, often on personal accounts. How to govern it with clear rules, business plans, permissions and training, and what the law requires.

Conceptual Creative 7 min read
Leer en español →

Your team already uses artificial intelligence, whether you’ve approved it or not. According to an OpinionWay study for Cegid published in August 2026, 82% of office workers in Spain use AI tools in their daily work, and 82% of those users admit to having used tools their company didn’t provide, without telling it. Meanwhile, according to Eurostat, only 20% of EU businesses with ten or more employees reported using AI in 2025. That gap has a name: shadow AI.

Banning it doesn’t work; it just pushes usage onto personal accounts you can’t see. What works is providing a company tool under a contract that protects your data, setting out which information can go where, assigning role-based permissions, training people and measuring results, all while keeping in mind what the EU AI Act and the GDPR already require.

What shadow AI puts at risk

When someone pastes a contract or a customer list into a personal account on an AI assistant:

  • The data leaves your control, without the data processing agreement that Article 28 of the GDPR requires when a third party processes personal data on your behalf.
  • It may be used to train models. On many consumer plans that’s the case unless the user opts out: Anthropic says so for its individual plans, and Google for what’s shared with Gemini from personal accounts.
  • There’s no trail of what was shared or by whom, and the information walks out the door with the person when they leave the company.

Spain’s data protection authority, the AEPD, summed it up in a ten-point guide published in January 2026: at work, follow your organization’s information and security policies and don’t include confidential data about the company, its staff or its customers. For that, the policy has to exist.

A usage policy that fits on two pages

You need a short document that says which tools are approved, which data can go into each one and what gets reviewed before an output is used. A tiered classification usually does the job:

Tier Examples Where it can go
Public Website, catalog, press releases Any tool
Internal Procedures, drafts, minutes without personal data Approved company tools only
Confidential or personal Customer and employee data, contracts, unpublished figures Company tool with a data processing agreement, for approved uses and with minimal data
Prohibited Passwords, API keys, health data and other special categories None, except a specific system that has been assessed

Add four rules:

  • Anything that goes to a customer, gets published or informs decisions about people is reviewed by a person first.
  • No decision about people (hiring, firing, granting credit) is made by AI alone; Article 22 of the GDPR restricts decisions based solely on automated processing.
  • Since August 2, 2026, Article 50 of the AI Act requires disclosure when you distribute deepfakes (AI-generated image, audio or video that passes as authentic) and when you publish AI-generated text on matters of public interest without editorial review.
  • Requests for new tools get an answer within days, or people will stick with their personal accounts.

Business plans: what you’re actually buying

The difference between the free version of an assistant and its business plan lies mostly in the contract and the admin controls. Based on each vendor’s terms as of October 2026:

  • OpenAI doesn’t train on data from ChatGPT Business, ChatGPT Enterprise or its API by default, and signs data processing agreements.
  • Anthropic doesn’t train on content from Claude Team and Enterprise by default; on its individual plans, it depends on the user opting out.
  • Microsoft states that Microsoft Copilot (formerly Microsoft 365 Copilot) doesn’t use prompts or responses to train foundation models, although the Anthropic models available inside it are currently excluded from its EU Data Boundary.
  • Google says Gemini content in Workspace isn’t used to train models outside your domain without permission.

Before signing, check the data processing agreement, where data is processed and what safeguards apply if it leaves the EU, how long it’s retained, which subprocessors are involved and what admin controls you get. For reference, Claude’s Team plan costs $25 per user per month, or $20 billed annually.

Roles and permissions: who can do what

A small business doesn’t need a committee, but it does need names: someone responsible for AI, even part-time, who maintains the list of approved tools and use cases; IT, for accounts, connectors and offboarding; whoever handles data protection, to assess each new use involving personal data; and department heads who approve their team’s uses.

Permissions matter more now that assistants connect to email, shared folders or the CRM, because they see whatever their user can see. If your permissions are misconfigured, AI doesn’t create the problem, but it exposes it in seconds: review who can see what before you switch on connectors.

With agents that act on their own, raise the bar. The AEPD’s February 2026 guidance on agentic AI calls for blocking automated actions without human oversight when an agent processes uncontrolled input and can reach sensitive data, and warns against pinning all the blame on the human supervisor when it’s the system that failed.

What the law requires as of October 2026

If you use AI but don’t build it, the EU AI Act treats you as a “deployer.” This is general information, not legal advice:

  • AI literacy (Article 4). Applicable since February 2, 2025. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since July 27, 2026, softened the wording: deployers must now “take measures to support the development of AI literacy” of their staff, and the text clarifies that this doesn’t require guaranteeing any specific level. The European Commission says no certificate is needed and suggests keeping an internal record of training.
  • Prohibited practices (Article 5). Since February 2025, using AI to infer a person’s emotions in the workplace has been banned, except for medical or safety reasons. Fines for prohibited practices reach €35 million or 7% of worldwide annual turnover, whichever is higher; for SMEs, whichever is lower.
  • Transparency (Article 50). Beyond the disclosures in the policy above, if you build your own chatbot you’re probably its provider and must tell users they’re talking to an AI.
  • High-risk uses at work (Annex III). Using AI to screen candidates, filter résumés, decide on promotions or dismissals, or evaluate performance is high-risk. The Omnibus pushed those obligations from August 2, 2026 to December 2, 2027; they include informing workers’ representatives and affected staff beforehand.

In Spain, the draft organic law that was meant to designate national authorities and set penalties lapsed on October 6, 2026, when parliament was dissolved ahead of the general election. The EU regulation applies regardless.

The GDPR still weighs more day to day: a legal basis, a clear purpose, minimal data and, if you use AI to evaluate employees or candidates, most likely a data protection impact assessment. On top of that, Article 64.4.d of Spain’s Workers’ Statute has, since 2021, given works councils the right to be informed of the parameters, rules and instructions behind algorithms or AI systems that affect working conditions and employment.

Training that counts toward Article 4

According to the Cegid study, 56% of Spanish employees haven’t received specific AI training from their company. Training that works is short, practical and tailored to the role:

  • Which tools are approved and which data never gets pasted in, with examples from your sector.
  • How to check an output: made-up figures, citations that don’t exist, errors stated with total confidence.
  • For anyone using agents or connectors, what prompt injection is and why an external document can end up giving the AI instructions.

Keep the materials, dates and attendance lists: it’s the simplest way to show you’ve taken measures.

How to measure adoption and value

Without measurement, you’ll end up paying for licenses nobody uses or missing widespread unofficial use. Five indicators are enough to start:

  1. Real license usage. Weekly active users against paid licenses.
  2. Unapproved usage. A quarterly anonymous survey. If shadow AI doesn’t drop, your official offering isn’t meeting a need.
  3. Use cases with an owner. Which task, which data it touches and who approved it.
  4. Time saved, measured rather than perceived. Before and after on two or three specific tasks. Perception is misleading: in a 2025 METR trial, experienced developers believed AI had made them 20% faster when they had actually taken 19% longer.
  5. Quality and incidents. Errors caught in review, customer complaints and AI-related data incidents.

Review these numbers every quarter: renew what gets used, drop what doesn’t and expand what proves its value.


Bringing order to AI use in a team is mostly an organizational job, but there’s a technical side: choosing and configuring tools, connecting them to your CRM or processes without opening permission gaps, and logging what they do. At Conceptual Creative we help companies apply artificial intelligence with that control built in from day one, and automate processes where it makes sense.

Sources

  • artificial intelligence
  • governance
  • data protection
  • eu ai act
  • training

FAQ

Frequently asked questions

Is AI training for employees mandatory?

Yes. Article 4 of the EU AI Act has applied since February 2, 2025 to companies that use AI systems. Since July 27, 2026, after the Digital Omnibus, it requires measures to support staff's AI literacy, without mandating a specific level or any certificate. It's wise to keep a record of the training you provide.

Can an employee use a personal ChatGPT or Claude account for work?

For public information, yes. For customer or employee data or confidential information, they shouldn't: a personal account comes with no data processing agreement, and on many consumer plans conversations can be used to train models unless the user opts out.

Which AI plans don't train on my company's data?

According to their terms as of October 2026, ChatGPT Business and Enterprise, Claude Team and Enterprise, the OpenAI and Anthropic APIs, Microsoft Copilot and Gemini in Google Workspace don't use customer content to train models by default. Still, check the data processing agreement, retention and where data is processed.

Has the EU AI Act been delayed?

Only partly. Regulation (EU) 2026/1744, the Digital Omnibus on AI, pushed high-risk obligations for Annex III systems to December 2, 2027 and for Annex I systems to August 2, 2028. Prohibited practices and AI literacy have applied since February 2025, and most transparency obligations since August 2, 2026.

Keep reading

All articles

Can we help with something similar?

Tell us about your case. We reply within 24 hours.

Our products

Our products

Besides building for our clients, we create and run our own products.

Factulum

Factulum

Online invoicing software for freelancers, SMEs and accounting firms. It applies IGIC in the Canary Islands or VAT on the mainland automatically and is ready for VERI*FACTU.

  • Automatic IGIC and VAT
  • VERI*FACTU
  • API with A3, Sol and CEGID
Arcania

Arcania

Our video game: an MMO RPG for PC, an online multiplayer role-playing game coming in 2027 to Steam and the Epic Games Store. ALPHA sign-ups are open.

  • MMO RPG
  • PC
  • Steam and Epic Games Store
  • 2027