An internal AI use policy should state which categories of data staff may enter into AI tools, which tools are approved for each category, and who decides when a case is unclear. Sort company information into a few tiers, allow low-risk material under clear conditions, and require approval or a block for personal data, client material and confidential business information.
Key takeaways
- Sort company information into a small number of data tiers before you name any AI tool.
- Approve tools tier by tier, based on the written terms that govern retention and data use.
- Treat personal data and client confidential information as requiring explicit approval.
- Review the policy on a fixed schedule, because tools, contracts and regulations change.
What does an internal AI use policy need to cover?
It needs to cover four things: the data staff may enter, the tools they may use, the tasks AI may support, and the person accountable for exceptions. A policy that only names approved tools leaves the data question open, and that is where most of the risk sits. The same chat window can be harmless for a public product description and harmful for a client contract.
Keep the document short enough to read in one sitting. Staff follow rules they remember, so a long document that is filed once and never reopened offers little protection. A one-page summary for daily use, backed by a fuller reference for managers, works better than a single long text.
How do you classify data before staff use AI?
Classify data by how much harm a leak or a misuse would cause, then attach one rule to each class. Three or four tiers are enough for most organisations, because more tiers are hard for staff to apply while they work. Each tier should have examples taken from your own systems so that people recognise their files.
A practical tier structure
- Public: published marketing copy, public product documentation and openly available articles.
- Internal: meeting notes without personal data, internal process descriptions and draft plans that will not leave the company.
- Confidential: contracts, pricing models, source code, unreleased financial figures and supplier negotiations.
- Restricted: personal data about employees or customers, login credentials, health-related information and any data covered by a client's non-disclosure terms.
The restricted tier needs the most careful wording. The main reason is that personal data is protected by law, and the regulation section below explains the basis for that. In practice, most policies allow restricted data only in tools with written business terms and a documented approval route.
Why the boundary between tiers matters
Staff rarely work with one tier at a time. A support ticket can contain a customer name, a product defect and a contract reference in the same paragraph. Give staff a rule for mixed content, such as 'classify by the most sensitive element and remove what the task does not need', so that they do not have to guess.
Which AI tools should staff be allowed to use?
Approve only the tools that your organisation has reviewed for the tiers they will receive. Two products can share a similar chat interface and still differ in how they handle retention, access and the use of inputs, so the decision depends on the contract and documentation behind each tool, not on its appearance. Read the vendor's own documentation and terms, and record the answers you rely on.
| Option | Typical use | Main risk | Fit for tiers |
|---|---|---|---|
| Personal consumer account | Quick drafting and general questions | Terms and data handling are hard to confirm for work use | Public material only, unless the policy says otherwise |
| Business account with written terms | Drafting, summarising and analysis of company material | Staff may still paste more than the task needs | Internal and confidential, with training |
| Private or self-hosted deployment | Work on sensitive data sets | Higher operating effort and ownership of security controls | Restricted and confidential, under documented controls |
| AI features inside an approved system | Tasks inside existing workflows | Data flows depend on that system's configuration | Depends on the host system's classification |
When you approve a tool, record five items: the tiers it may receive, the account type required, the contract or documentation reviewed, the named owner and the review date. Without this record, the approval becomes a verbal agreement that nobody can check later. Where tasks need repeatable AI steps rather than ad hoc chat, the design is closer to AI automation work, where the data path is designed and logged rather than left to individual choices.
Hypothetical examples
Here are two hypothetical scenarios: the first shows how a tier rule resolves a common request, and the second shows how a mixed-content case is handled.
In the first scenario, an operations assistant wants to draft a reply to a supplier. The email thread includes a contact's personal phone number and the supplier's negotiated pricing. The pricing is confidential and the contact details are restricted, so the rule points to the approved business account, with the phone number removed before pasting. The free consumer tool is not an option for this task.
In the second scenario, an IT lead wants a summary of incident logs. The logs contain user identifiers and internal host names. The identifiers are restricted, so the policy requires replacing them with placeholders before the text leaves the internal environment. Host names are confidential and may go only to the approved tool. The lead records the decision, so the next person facing a similar log does not need to ask again.
Does the GDPR apply when staff enter personal data?
Generally yes: when the data relates to an identified or identifiable person, entering it into an AI tool is a form of processing, and the General Data Protection Regulation sets the obligations that apply to that processing. The regulation text is available on EUR-Lex, the official EU law database, and it should be read directly rather than through summaries.
The EU AI Act, Regulation (EU) 2024/1689, sets obligations for providers and deployers of AI systems, and its obligations apply on the dates stated in the regulation itself. Check the official text on EUR-Lex for which obligations apply to your organisation and when. This article is general information and not legal or tax advice.
Technical support from Rui Codex can help implement the controls a policy describes, such as access settings, data removal steps and logging. The legal wording of the policy and its fit with your obligations should be reviewed by qualified counsel.
Where do AI use policies fail in practice?
Policies fail when the rule is unclear at the moment of use, when no approved alternative exists, or when nobody owns the exceptions. Each failure has a predictable mechanism, and each can be designed out before the policy is published.
- Ambiguous categories: staff cannot tell whether a document is internal or confidential, so they guess. Give them examples and a default rule for doubt.
- Blocking without alternatives: a ban with no approved tool pushes people toward personal devices, which removes all visibility. A restricted but usable option is usually safer than a ban that people ignore.
- Unowned exceptions: requests are granted informally and accumulate. After a while the written policy no longer describes what happens.
- Static documents: tools and their terms change, and a policy written once drifts away from the tools in use.
Each failure has a trade-off. Stricter rules reduce exposure but increase the workarounds people find, while looser rules are easier to follow but harder to audit. Decide deliberately where your organisation sits on that line, and state the reason in the policy so that the decision can be revisited.
How do you write and roll out the policy?
Draft the tier list with the people who handle the data, approve tools with the owners of each tier, then publish a one-page summary and a clear route for questions. Rolling out a policy is a process of testing it against real work, not only of publishing a document.
- Map the main data sets and assign each one a tier with its data owner.
- List the tools in use, including AI features inside existing systems, and check their written terms.
- Decide which tool may receive each tier and record the account type required.
- Write the one-page summary with three examples per tier and a default rule for doubtful cases.
- Run a short briefing for each team and collect the questions that the summary does not answer.
- Name an owner for exceptions and set a review date for the full document.
Organisations that need help turning this into working controls can use an AI consulting engagement to review tools and data flows. For background on the data protection side, the security and GDPR articles on the Rui Codex blog cover related engineering topics.
Frequently Asked Questions
Can staff use free AI chat tools for work?
Only for material classed as public, unless your policy explicitly allows more. Free consumer tools usually lack the contract terms and administrative controls that a business account provides, so you cannot easily confirm how inputs are retained or used. Check the vendor documentation for each tool and record the decision, so that teams do not answer the same question differently.
Should the policy ban personal data entirely?
No, a total ban is hard to enforce and tends to push use onto personal devices. A more workable approach restricts personal data to approved business tools with written terms, requires removal of identifiers where the task allows it, and requires approval for anything beyond that. Keep the rule simple enough that staff can apply it while working.
Who should own AI policy exceptions?
A named role should own exceptions, usually the IT lead working with the data owner for the affected tier. Each decision should record the reason, the tool, the data involved and an expiry date. Without an owner, exceptions accumulate quietly and the policy stops reflecting actual practice.
How often should the policy be reviewed?
Review it whenever a tool, a vendor contract or a relevant regulation changes, and at least once a year as a routine check. Ask the people who use the tools what they do differently from the written rules, because gaps appear there first. Update the one-page summary at the same time as the full document.
Conclusion
A workable AI use policy starts from data, not from a list of tools. Sort what staff handle into tiers, decide which tools may receive each tier, and give people one place to ask when a case is unclear. Review the document when tools or contracts change, and record exceptions so the next reviewer can see why a decision was made. If you want help turning the tiers into working controls, request a project consultation at the Rui Codex contact page or email info@ruicodex.com to discuss your project.
Sources
- Regulation (EU) 2016/679 (General Data Protection Regulation) - EUR-Lex, Publications Office of the European Union
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) - EUR-Lex, Publications Office of the European Union