What rules does your team need for AI?

The idea

Many people already use AI at work without shared rules. In the University of Melbourne's global study with KPMG (more than 48,000 people in 47 countries, surveyed November 2024 to January 2025), 58 percent of employees said they intentionally use AI at work, but only 40 percent said their workplace has a policy or guidance on generative AI use, and almost half admitted to using AI in ways that contravene company policies, including uploading sensitive company information into free public AI tools. These figures come from the university's media release.

Rules matter more now because assistants do work, not just answer. They draft into shared files, read inboxes, connect to your apps and run tasks on a schedule. In 2026 the agent moved next to the chat: inside ordinary Claude, behind ChatGPT's Chat and Work toggle, and beside Gemini's chat as Task. So a team rule about agents is now a rule about the assistant everyone already opens. Anthropic's safety page for Claude Cowork puts the responsibility plainly: "You remain responsible for all actions taken by Claude performed on your behalf" (as of Oct 2026).

Good team rules are short, name a person for each job, and match what the tools are actually set to do. If a rule cannot be followed on a busy Tuesday, it will not be.

An illustration, not a real case: a ten-person team agrees its page in one meeting. The rule that changes most is number 4: every client document now names the person who checked it. The rule that takes longest is number 5, because nobody knew which apps the team's assistant was connected to until someone looked.

The one page: seven rules

1. Approved tools and accounts. List the AI tools the team may use for work, on which plan, and say: work accounts only, never a personal or free account for work. Use the tool your company approves for company data: a personal preference never beats the company's data rules. New AI features inside software you already use count as new tools, and so do new connectors and skills.

2. What never goes in. Write a short never list. A common starting point: personal data, anything a client shared in confidence, passwords and security details, and financial results or prices that are not public yet. Add your own. Say what to do when unsure: leave it out, or ask first.

3. When we say AI made it. Decide when the team tells clients, colleagues or the public that AI did a large part of the work. Where the EU AI Act applies, some disclosure is required by law (see below).

4. The quality bar, and who reviews. Anything AI helped to write or make is checked by a named person before it leaves the team. "The AI did it" is never the answer to "who checked this?"

5. What agents may do without asking. Use four levels: 1 Read, 2 Draft, 3 Ask first, 4 Alone. Most team tasks start at 1 or 2. "Ask first" is set in the tool, not typed in the brief. And outside words, no power: an agent that reads outside content (the web, outside email, files others sent) may not also send, pay or delete without asking. The UK's National Cyber Security Centre says to decide "what actions it should be allowed to take" and "when it should stop and seek human approval", and adds: "do not rely on prompting alone" (20 Aug 2026).

6. Who to tell when something goes wrong. Name one person or channel, and say "the same day". Make sure someone can stop an agent at once: the NCSC says "you should always be able to 'pull the plug'". Say that nobody is blamed for reporting a slip quickly.

7. How we keep our skills up. Agree a short training for everyone who uses AI, a refresher when the tools change, and keep a simple record of what was done. Where the EU AI Act applies, the Commission says an internal record of trainings is enough and no certificate is needed (see below).

Then write two more lines: who owns the page, and the date you will review it. These tools change every month.

What the law says, in plain English

This is plain-English orientation from the official texts, read on 16 Oct 2026. It is not legal advice. Whether a law applies to you depends on where you are, what you do and how you use AI: check with your own adviser.

The EU AI Act

The AI Act (Regulation (EU) 2024/1689) applies, among others, to organisations established or located in the EU that use AI systems, and to organisations outside the EU where the output produced by the AI system is used in the EU (Article 2(1), points (b) and (c)). In the Act, an organisation using an AI system under its authority is a "deployer" (Article 3(4)).

• AI literacy (Article 4), rewritten in July 2026. Regulation (EU) 2026/1744 (signed 8 Jul 2026, published in the Official Journal on 24 Jul 2026, in force on the third day after publication) replaced the old wording, which asked providers and deployers to take measures to "ensure, to their best extent, a sufficient level of AI literacy". Providers and deployers now "shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf". The new text adds: "This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual." The Commission is to publish practical examples of how to comply. Its own Q&A says: "There is no need for a certificate. Organisations can keep an internal record of trainings and/or other guiding initiatives." That is rule 7.
• Telling people about AI (Article 50), from 2 Aug 2026. A deployer must disclose that image, audio or video it made or changed with AI is artificially generated or manipulated when it is a deep fake. A deployer whose AI generates or manipulates text "published with the purpose of informing the public on matters of public interest" must disclose that too, unless the text "has undergone a process of human review or editorial control" and a person or organisation "holds editorial responsibility for the publication". That is rule 3.
• Banned at work since 2 Feb 2025 (Article 5(1)(f)). AI systems "to infer emotions of a natural person in the areas of workplace and education institutions", except for medical or safety reasons.
• Hiring and staff decisions are high risk (Annex III, point 4). That covers AI meant for recruitment or selection, including filtering applications and evaluating candidates, and AI meant to make decisions on terms of work, promotion or termination, "to allocate tasks based on individual behaviour or personal traits or characteristics", or to monitor and evaluate performance and behaviour. The main duties for these systems now apply from 2 Dec 2027 (Regulation (EU) 2026/1744, which moved them from 2 Aug 2026).

Illinois, United States

• Public Act 103-0804, in effect since 1 Jan 2026. In recruitment, hiring, promotion, discipline, discharge and the other employment decisions it lists, an employer may not "use artificial intelligence that has the effect of subjecting employees to discrimination on the basis of protected classes", or use zip codes as a proxy for them. It must also give notice when it uses AI for those purposes. The state's Department of Human Rights is to set the rules for that notice; its rules page listed none for AI on 16 Oct 2026.

Left out on purpose: Colorado's AI law, because its current start date could not be confirmed on 16 Oct 2026. Other countries' and states' rules are not covered here.

Common mistakes

• A twenty-page policy. Nobody reads it. Start with one page and link to the detail.
• Rules nobody can follow. A rule that bans what people need to do gets worked around.
• Agent rules only in the brief. Set "Ask first" and the blocked actions in the tool.
• No names. "Someone checks it" means nobody does. Name the role.
• Forgetting the AI already inside your software, and meeting note-takers.
• Blaming the person who reports a slip. Then the next one is hidden.
• Treating the law section as legal advice, or copying another country's rules as if they were yours.
• No review date. The tools change every month, and so do their defaults.

What to check before you trust it

Whether the page came from a template, a colleague or an AI tool:

• Every law line was read from the official text, with the date, and your adviser has looked at it.
• If an AI tool drafted any of it, every legal claim has been checked at its source. Quotes and links can be invented too.
• The tool settings match the page: open the admin settings and look.
• Every rule names a role.
• Someone new to the team can read it in two minutes and know what to do.
• It has an owner and a review date.

Check it. Rule 4: a named person checks anything AI helped to make before it leaves the team.

Keep it safe. Rule 5 in the tool, not on paper: outside words, no power.

As of: product steps as they were on the vendors' own pages on 16 Oct 2026.

Every product name, setting and step above comes from Neosaipien's platform reference, checked on the vendor pages below on 16 Oct 2026; every law line was read on its official site that day.

• Regulation (EU) 2024/1689, the AI Act (Articles 2, 3(4), 5(1)(f), 50(4), 113; Annex III, point 4), on EUR-Lex
• Regulation (EU) 2026/1744 (8 Jul 2026, OJ 24 Jul 2026): Article 1, point (5), new Article 4; point (40), new dates in Article 113; Article 4, entry into force. On EUR-Lex
• European Commission, AI Literacy: Questions and Answers (last update 27 Jul 2026)
• Illinois General Assembly, Public Act 103-0804 (775 ILCS 5/2-102(L); Section 99, effective 1 Jan 2026)
• Illinois Department of Human Rights, Rules and Regulations (no AI notice rules listed on 16 Oct 2026)
• UK NCSC, Managing the cyber risk of agentic AI (20 Aug 2026)
• University of Melbourne, media release on the Trust, attitudes and use of AI global study 2025 (with KPMG; more than 48,000 people, 47 countries, surveyed Nov 2024 to Jan 2025)
• Anthropic, Use Claude Cowork safely (the responsibility quote)
• Anthropic, Claude Cowork and chat are now one Claude (16 Sep 2026)
• Anthropic, Get started with connectors
• Anthropic, Use the connectors directory
• Anthropic, Use Claude's chat search and memory
• Anthropic, What are projects
• OpenAI, Connected apps in ChatGPT
• OpenAI, Projects in ChatGPT
• Google, Use Gemini Apps with a work or school Google Account
• Google, Use apps connected to Gemini with a work or school account
• Google, About Gemini models (work accounts depend on the licence)
• Google, Manage your tasks and workflows in Gemini Apps
• Google, Connect and manage custom apps for Gemini Apps
• Google, Create and manage skills for Gemini Apps
• The shared rules (the four levels, "Ask first" set in the tool, outside words no power, "The AI did it" is never the answer) are Neosaipien's own words, used the same way in every kit.

Not legal advice. This lesson gives plain-English orientation from the official texts only; check your own situation with your adviser. The illustration in this lesson is made up. Neosaipien is not affiliated with Anthropic, OpenAI or Google.