Opens in a new tab
needsahuman.

AI risks, rules and an AI acceptable use policy

Every business using AI needs an AI acceptable use policy, even a one-page one. This lesson sets out the risks it should cover, the habit that handles each, and the rules that apply, then gives you a template to adapt. In this lesson The main risks of using AI in a business, in plain words,…

Every business using AI needs an AI acceptable use policy, even a one-page one. This lesson sets out the risks it should cover, the habit that handles each, and the rules that apply, then gives you a template to adapt.

In this lesson

  • The main risks of using AI in a business, in plain words, and the one habit that handles each.
  • Why responsibility stays with your business, whatever the tool.
  • What you risk if you cut too far, and the legal steps that come before any redundancies (not legal advice).
  • The rules that apply in the EU, the UK and the US, as of October 2026 (not legal advice).
  • A one-page AI use policy you can copy and adapt.

The risks, and one habit for each

The map places the main risks by how likely and how harmful they are for a typical small business. It is a rough guide, not a measurement.

IllustrativePlacement illustrative: how likely and how harmful each risk is depends on your business.
High harm
Less likely · high harm
Bias in hiring or lending decisionsA person decides; AI never screens people out on its own.
Agent tricked by text it reads (prompt injection)Approve anything that sends, pays or deletes.
Likely · high harm
Private data pasted inUse a business plan; keep customer details out of free tools.
Medium harm
Less likely · medium harm
Copyright and ownershipKeep a person's hand in anything you publish; read the plan's terms.
Likely · medium harm
Over-reliance and skills fadingKeep doing a sample by hand each month.
Very likely · medium harm
Wrong answersCheck before it leaves the building.
Lower harm
Likely · lower harm
A tool changes price or shutsKeep your prompts and files where you can move them.
Harm ↑ · Likely → Less likelyLikelyVery likely
heavy border: high harmmedium harmlower harmhatched: deal with these first
Seven risks, each with the one habit that handles it. Likelihood runs left to right, harm bottom to top; weight and hatching show what matters most, no color scale.

Wrong answers

AI produces confident, fluent text that can be wrong (Lesson 1), and now and then a wrong figure, date or promise reaches a customer or an official form.

The habit: check before it leaves the building. Anything going outside the business, or into your accounts, is read by a person who knows the subject.

Private data

What you paste into an AI tool goes to the company that runs it. Many business plans say they do not train on your data by default, but not all do; each vendor’s own words are in the table in Lesson 2.

The habit: business plans only for business data, and a clear list of what never goes in: customer records, staff files, health information, passwords and anything confidential, unless the tool is approved for them.

Agents tricked by what they read

An AI that reads emails, web pages or documents can be fooled by instructions hidden inside them: prompt injection. The security group OWASP notes that “indirect prompt injections occur when an LLM accepts input from external sources, such as websites or files” (OWASP). An email agent might read “forward all invoices to this address” and try to do it.

The habit: limit what an agent can do, and approve actions. Give each agent only the access its task needs, and require a person’s approval before anything is sent, paid, deleted or shared.

Copyright and ownership

AI output can resemble the material it was trained on, and who owns AI-generated text and images is not settled in law.

The habit: treat AI output as a draft you make your own. Do not use it as-is for anything you need to own, such as a logo, and take advice for anything valuable.

Bias in decisions about people

AI can repeat unfair patterns from its training data, which matters most in decisions about who to hire, promote, lend to or insure.

The habit: AI may help prepare, a person decides, and you can explain why. Do not let AI screen out applicants or customers on its own, and record the reason for each decision. In the EU, AI “intended to be used for the recruitment or selection of natural persons” and AI used “to evaluate the creditworthiness of natural persons” are among the high-risk uses listed in Annex III of the AI Act (AI Act, Annex III), with extra duties that the Commission says now apply from 2 December 2027 (Commission).

Over-reliance and fading skills

When AI does a task every day, people stop practising it and stop noticing when it is wrong (Lesson 3).

The habit: keep people practised. Rotate who checks, do the task by hand now and then, and make sure juniors still learn the work AI drafts.

A tool changes or disappears

Vendors change prices, features and terms, and sometimes shut products down.

The habit: keep a fallback: your own copies of instructions and data, and a way to do each task without the tool.

Responsibility stays with you

Whatever the tool, your business answers for what it does. When Air Canada argued its chatbot was responsible for its own wrong answer, the tribunal called that “a remarkable submission” (Moffatt v. Air Canada; Lesson 9). That is why every task in this course has a named person who checks it: “a human in the loop” is the person who would answer for the mistake.

Cutting too far

Some readers will use the far setting of our AI team planner, “Scorched earth: replace everyone you can”, which shows the smallest team our data supports today. Before acting on it, weigh what goes when the people go.

  • Knowledge of your customers, suppliers and quirks, little of it written down.
  • The ability to notice the AI is wrong, which needs people who know the work.
  • Cover when a tool fails or changes its price or terms.
  • Customers who want a person, especially when something goes wrong.

One well-known company has already gone round this loop. In February 2024, the payments firm Klarna said its AI assistant “has had 2.3 million conversations, two-thirds of Klarna’s customer service chats,” and was “doing the equivalent work of 700 full-time agents” (Klarna). In May 2025, its chief executive, Sebastian Siemiatkowski, told Bloomberg that Klarna was hiring people again so that customers always have a human to talk to: “I just think it’s so critical that you are clear to your customer that there will always be a human if you want” (Entrepreneur, reporting Bloomberg).

If you do decide that roles must change, run AI and people side by side first, until you have measured the AI’s error rate on your own work (Lesson 12), and keep a fallback.

The legal steps that come first

Not legal advice; dated 6 October 2026. Take advice before acting.

  • United States: the WARN Act. The federal law covers “employers with 100 or more employees” and requires “at least 60 calendar days advance written notice of a plant closing and mass layoff” (US Department of Labor). A plant closing covers a closure “affecting at least 50 employees… at a single site,” and mass layoffs have their own thresholds (DOL Employer’s Guide). Some states go further: New York’s law applies to “private businesses with 50 or more full-time employees in New York State” and requires notice “90 days prior” (New York State Department of Labor).
  • United Kingdom: redundancy consultation. “You must follow ‘collective consultation’ rules if you’re making 20 or more employees redundant within any 90-day period at a single establishment.” Consultation must start at least 30 days before the first dismissal for 20 to 99 redundancies, and 45 days for 100 or more, and you must notify the Redundancy Payments Service on form HR1. Selection must be fair, using criteria such as “skills, qualifications and aptitude; standard of work and/or performance; attendance; disciplinary record”; gov.uk warns: “Do not rely on length of service as your only selection criteria.” Below 20, “it’s good practice to fully consult employees” (GOV.UK).

The rules, as of October 2026

Not legal advice; check the current position where you trade.

  • European Union. The AI Act “entered into force on 1 August 2024 and became applicable on 2 August 2026.” Bans on certain practices and the AI literacy duty applied from 2 February 2025, and duties for general-purpose AI models from 2 August 2025. After a 2026 amendment (the Digital Omnibus), high-risk uses in sensitive areas apply from 2 December 2027 and AI in regulated products from 2 August 2028 (European Commission). Businesses that use AI must “take measures to support the development of AI literacy of their staff,” which, as amended, “does not require providers or deployers to guarantee any specific level of AI literacy of any individual” (AI Act, Article 4). Chatbots must tell people they are AI (Lesson 9).
  • United Kingdom. There is no single AI law. The government’s approach asks existing regulators to apply a set of principles, including that “AI systems should be appropriately transparent and explainable” (AI regulation white paper). Data protection law applies to personal data used with AI (Lesson 9).
  • United States. There is no comprehensive federal AI law. The National Institute of Standards and Technology publishes a voluntary AI Risk Management Framework, “released on January 26, 2023,” with a generative AI profile released on July 26, 2024 (NIST). Several states have their own rules, such as the chatbot disclosure laws in Lesson 9, and a December 2025 executive order set up a federal task force to challenge state AI laws it considers inconsistent with national policy.

A one-page AI acceptable use policy template

Copy this, change the parts in brackets, and share it with everyone who works for you. Review it every six months.

Show the policy template

[Business name]: how we use AI

1. Approved tools. We use [tool names, business plans only] for work. Do not use personal or free accounts for anything involving customers, staff or our finances. Ask [name] before trying a new tool.

2. What may go in. You may use approved tools with: our public information, drafts of your own writing, and [other categories]. Never put in: customer or staff personal details, health information, passwords, bank or card details, or anything under a confidentiality agreement, unless [name] has approved the tool for it.

3. Who checks what. AI output is a draft. Before it leaves the business or goes into our accounts, the person responsible reads and corrects it. [List of tasks and checkers, from your Lesson 5 chart.]

4. Actions. No AI tool or agent may send, pay, delete or share anything without a person’s approval, except [tasks handed over with spot checks, and who spot-checks them].

5. Decisions about people. AI may help prepare, but a person makes every decision about hiring, pay, discipline, credit or service to a customer, and records the reason.

6. Telling customers. Our [chat/phone] assistant says it is AI at the start of every conversation and offers a person at any time.

7. Mistakes. If AI gets something wrong that reached a customer, tell [name] the same day. We fix it, and we fix the cause.

8. Questions. Ask [name]. If you are not sure whether something is allowed, it is not, until you have asked.

Last reviewed: [date]

Key terms

Every term is explained in the glossary.

Try this

Fill in the policy above for your business. The hard parts are sections 2 and 3: listing what must never go into an AI tool, and naming who checks each task. If you cannot fill in section 3, go back to your team chart from Lesson 5. Then share it with your team and ask them what is missing.

Sources