Banning AI is a decision too, and usually the wrong one.
The instinct after discovering staff pasting documents into a free chatbot is to forbid it. The IW data points the other way: unapproved AI use was most common where employers actively build AI skills. Among employees who said their employer strongly supports AI skills, 44.5% used tools the company had not provided, against 21.3% of those who saw no such support. The researchers read this as a supply gap, not a discipline problem: people who learn what AI can do want to use it, and if the company offers no approved tool, they bring their own. A ban mainly hits your most engaged people and pushes the usage out of sight. The data is self-reported and a single snapshot, so it shows a link rather than a cause, but it is a strong reason to decide what to provide before deciding what to forbid.
Five decisions a usable AI policy has to make.
Many templates run to ten pages of definitions and principles. For a company of 20 to 500 people, one page that answers these five questions does more:
- Tools: Which AI tools are approved, under which license, and how does someone request a new one? Name an owner and a turnaround time, or people will route around the process.
- Data: What may go in? A simple traffic light works: public information is green; internal documents go only into approved tools covered by a data processing agreement; personal data about customers or employees, trade secrets and confidential contracts never go into unapproved tools.
- Use cases: Where may AI draft, summarize and analyze, and where may it not decide? Decisions about individual people, such as hiring, performance reviews or terminations, stay with humans.
- Sign-off: Which AI-assisted outputs need a named person to check them before they reach a customer, the board or a decision paper?
- Ownership: Who owns the policy, how are incidents reported, and when is it reviewed? Quarterly is a sensible starting rhythm while tools change this fast.
Write a separate rule for AI-assisted business decisions.
Most policies are written for everyday tasks such as drafting an email or summarizing a meeting. The bigger exposure is when AI analysis quietly feeds a strategic call on pricing, a market entry or a senior hire. A 2025 global study by the University of Melbourne and KPMG of more than 48,000 people in 47 countries found that many employees rely on AI output without evaluating its accuracy (66%). For decisions that are expensive or hard to reverse, require the decision paper to state which parts were AI-generated, which sources and assumptions a person has checked, and who approved the final call. That one rule turns AI from an invisible author into a documented input.
More on AI Governance — every week.
Join 2,400+ leaders getting the decision intelligence briefing.
Know what the law already expects.
Under Article 4 of the EU AI Act, companies that deploy AI systems have had to take measures to support their staff's AI literacy since 2 February 2025. A clear policy plus a short training session is a sensible baseline. If AI touches decisions about employees, such as recruitment or performance monitoring, stricter high-risk rules may apply later, and in Germany a works council may have co-determination rights over tools that can monitor employees. This article is general information, not legal advice; involve an employment or data protection lawyer before you finalise your policy.
Hypothetical example: an 85-person engineering consultancy.
Imagine a firm in Hanover whose managing director learns that the sales team summarizes customer tenders in free chatbots. Instead of an outright ban, the leadership team makes the five decisions in one meeting: one business-tier AI tool with a data processing agreement; a traffic-light data rule that keeps customer drawings and pricing out of anything else; AI allowed for first drafts of bids, but every bid and every staffing decision signed off by a named person who notes where AI was used; and the head of operations as policy owner, with a review after 90 days. A one-hour training session follows the rollout. The policy fits on one page, and every rule in it answers a question employees were already facing.
Roll it out in weeks, not quarters.
Draft the page in a week, test it with two or three teams, fix what is unclear, then publish it with a short training session and a visible place to ask questions. At each review, look at two things: requests for new tools, and incidents or near-misses. Many requests mean your approved toolset is too thin; zero requests often means people have stopped asking. If your leadership team is split between a ban, an approved tool or a staged rollout, you can run that choice through the NeuroAgent council for a structured second opinion before you commit.
The bottom line.
Your employees are already making AI decisions every day: which tool, which data, how much to trust the answer. An AI policy moves those decisions back to leadership, where they belong. Keep it short, decide what to provide before what to forbid, and treat AI input to important business decisions as something to document, not to hide.
Research context
IBM found that 97% of AI-related security breaches in its 2025 research involved AI systems without proper access controls.
IBM: Cost of a Data Breach Report 2025 (2025)This independently published statistic provides context for the topic; it does not substantiate every claim in this essay.
One decision framework, every week.
The week's best articles on making and executing high-stakes decisions, backed by sourced research.