Your team already uses AI. Write the policy they’ll follow.

Your team is already using AI. Developers use coding assistants. Sales teams draft emails with chatbots. Someone has a meeting-notes bot quietly joining client calls. Much of this happens on personal accounts that nobody in the business approved, a pattern often called ‘shadow AI’.

Banning AI tools rarely works. People switch to their phones, and you lose any visibility you had. A short, specific policy does better. It tells people what they can do, protects client data and intellectual property, and keeps the productivity gains.

This guide walks through how to write that policy for a small firm or startup, with example rules you can adapt. It draws on requirements in the EU, UK, US and Asia, but the core is the same wherever you operate.

Start with an honest inventory

Before writing a rule, find out what people actually do. Send a short, no-blame survey. Ask which tools people use, on which accounts, and for which tasks: coding, drafting, research, meeting notes, image generation or data analysis.

Check what your existing software has switched on, too. Many office suites, CRMs and design tools now include AI features by default. Those count as AI use, and their settings matter.

The answers usually surprise founders. They also tell you where the real risks are, which is where the policy should focus.

What a good policy contains

A useful policy fits on one or two pages. Long policies go unread. Use plain language and concrete examples, and link to longer guidance only where needed.

Most good policies have the same core sections. Keep each one short.

Purpose and scope: who the policy covers, including contractors, and that it applies on any device or account used for work.
Approved tools: a named list, with what each may be used for.
Data rules: what may and may not go into which tools.
Use rules: what people may do, what they must not do and what needs sign-off.
Output rules: review, accuracy, copyright and disclosure.
Reporting: how to raise a mistake or incident, without fear of blame.
Ownership: who maintains the policy and how often it is reviewed.

Classify the data first

The biggest practical risk is information leaking into a tool that stores it, uses it for training or lets the vendor’s staff see it. So classify your data before you classify tools. Three tiers are enough for most small businesses.

‘Personal data’ means any information about an identifiable person, such as names, emails, IDs or support tickets. Privacy laws such as the GDPR in the EU, the UK GDPR, India’s DPDP Act and US state laws all apply to it, whatever tool processes it. Confidentiality clauses in client contracts often go further.

Red, never in any AI tool: passwords, API keys and other credentials, health or financial data about individuals, and anything a client contract forbids sharing.
Amber, approved enterprise tools only: customer personal data, client confidential material, source code, unreleased financials and product plans.
Green, any approved tool: public information, your own published content and generic questions with no confidential detail.

Approve tools, not brands

The same AI product can be safe or unsafe depending on the plan. Consumer accounts may use your inputs for training and keep them for long periods. Business plans usually offer a contract, no training on your data and admin controls.

Approve each tool against a short set of criteria and record the result. Keep the list in one place everyone can find. Pay special attention to browser extensions and meeting bots, which can read far more than people realise.

Inputs are not used to train the vendor’s models, or you can switch training off.
Retention is limited, and you can delete data.
A data processing agreement is available, covering your privacy law obligations.
Single sign-on, admin controls and audit logs exist.
You know where data is stored and which sub-processors see it.

Example rules you can adapt

Rules work best when they are specific and short. The examples below suit a typical startup or professional firm. Adjust them to your clients and sector.

Use only tools on the approved list, through your work account.
Never paste red-tier data into any AI tool.
Remove names and identifying details before using amber-tier data, unless the tool is approved for it.
Ask before recording or transcribing a call with an AI tool, and tell external participants.
Do not use AI to make or recommend decisions about hiring, firing, pay or credit without approval from the policy owner.
Check every fact, figure, citation and link in AI output before it leaves the business.
Tell clients when AI has played a material role in work you deliver, if your contract or their policy requires it.

Own the output

The person who sends or ships something is responsible for it, whatever tool helped. AI systems can ‘hallucinate’, meaning they produce confident, plausible and wrong answers. Review is not optional.

Copyright needs care. The US Copyright Office said in January 2025 that prompts alone generally do not make you the author of AI output, so purely AI-generated material may not be protectable in the US. Other countries take different views. If a deliverable’s value lies in owning it, make sure there is meaningful human creative input.

Generated code raises a separate issue. It can resemble existing open-source code with licence conditions attached. Run your usual licence and security scanning on AI-written code, just as you would on code copied from the internet.

Customer-facing and high-stakes uses

Internal drafting is low risk. Putting AI in front of customers, or using it to decide things about people, is not. Your policy should send these uses to the owner for review before launch.

Since 2 August 2026, the EU AI Act has required that people be told when they are interacting with an AI system, unless that is obvious. If you add a chatbot to your website, label it. If you publish AI-generated images or video that look real, disclose that too.

AI used in hiring, credit and similar decisions attracts the most regulation. In the EU, many of these uses are classed as ‘high-risk’, with full obligations deferred to 2 December 2027. New York City already requires bias audits and notices for automated hiring tools. The UK requires safeguards for significant automated decisions about people.

Keep it alive

Name one owner. In a small company that may be the founder, the COO or the head of engineering. Review the approved tool list every quarter, because vendors change their terms often.

Train people briefly when they join and when the policy changes. The EU AI Act expects providers and deployers of AI to take measures supporting their staff’s AI literacy. After the 2026 Omnibus amendments, the duty focuses on taking those supporting measures rather than guaranteeing a set level of skill. Keep a simple record of what you did.

Log incidents and near misses, and learn from them. As you grow, voluntary frameworks such as the NIST AI Risk Management Framework, or certification to ISO/IEC 42001, the international standard for AI management systems, can help you scale. Your policy will also feed straight into enterprise customers’ security questionnaires.

Checklist

Survey your team on the AI tools and accounts they actually use.
Classify your data into red, amber and green tiers.
Approve tools against written criteria, preferring business plans with no training on your data.
Publish a one-page policy with approved tools, data rules, use rules and an owner.
Require human review of AI output before it reaches a client or customer.
Route customer-facing and people-decision uses to the owner before launch.
Label chatbots and realistic AI-generated media shown to the public.
Train staff on the policy and record that you did.
Review the approved tool list and the policy every quarter.

FAQs

Should we just ban public AI tools?

A blanket ban usually drives use underground. A better approach is to approve a small number of business-grade tools, make them easy to access, and ban only specific data or uses.

Do we need a policy if we only have ten people?

Yes, but a short one. A single page covering approved tools, forbidden data, review rules and an owner is enough to start. Enterprise customers will ask for it in due diligence.

Is it a data breach if someone pastes customer data into a chatbot?

It can be. If personal data is disclosed to an unapproved third party, privacy laws such as the GDPR may treat it as a personal data breach with reporting deadlines. Make sure people know to report such slips quickly, so you can assess them.

Who owns content our staff create with AI?

Between you and your employees, normal employment terms usually apply. Whether anyone owns copyright in purely AI-generated material is a separate question that varies by country. Meaningful human input strengthens your position.

How does the EU AI Act affect a small firm that only uses AI tools?

As a deployer, you should take measures to support staff AI literacy, disclose chatbots and realistic synthetic media, and avoid prohibited practices. Heavier duties apply only if you use AI in high-risk areas such as hiring or credit.

Primary sources

NIST · AI Risk Management Framework European Commission · EU AI Act

General information, not legal advice.