Sorting Work Into 'OK to Use AI' and 'Not OK to Use AI' — How to Build Company Rules for Generative AI, From the Basics
Drafting documents, summarising materials, assisting with code, preparing a first version of a translation. Generative AI has entered a wide range of office work. Convenient as it is, it also brings worries: might information entered into it leak outside? Might its output contain errors? What happens to copyright? That is why many companies have begun writing internal rules. But a document that merely lists prohibitions will not be used in practice. What is the rule meant to protect, and how should it fit the work people actually do? This article explains the thinking behind company rules for generative AI, and how to build them, from the basics.
What company rules for generative AI are — 'do not use it' is not enough
Company rules for generative AI are a written set of promises about how generative AI may be used at work: what may be entered, how output should be handled, and which tasks it may be used for, shown in a form that lets employees judge without hesitation. Without rules, employees start using it on their own judgement. Some, drawn by its convenience, may enter confidential information; others, out of fear, may refuse to use it at all. Neither is a desirable state for a company.
The role of rules is not only to prohibit. Their important function is to make clear the range in which AI can be used with confidence. If the line is drawn — 'up to here, you may use it' — employees can start using it without hesitation. Without a line, using it at all becomes a source of anxiety, and a useful tool never takes root in the workplace. The first step is to think of internal rules not as something that restricts the use of generative AI, but as the foundation for using it confidently and well.
The rules apply not only to full-time employees. It is desirable that the same rules apply to contract employees, dispatched workers, and outsourcing partners — everyone involved in the company's work. When people outside the company handle company information, the rules need to be shared with them and a mechanism put in place so that they follow them. Stating clearly at the start of the rules who they apply to prevents gaps in coverage.
First, sort the work — three categories of use
The starting point for building rules is to divide work into three categories: 'may use', 'may use with conditions', and 'must not use'. For example, summarising published materials or drafting general text falls under 'may use'. Work handling unpublished internal materials falls under 'may use with conditions', with limits set on what information may be entered. Work handling customers' personal information or information covered by a non-disclosure agreement with another company falls under 'must not use'. By deciding these categories for each type of work, employees can judge which category their own work belongs to.
- May use: work with no risk of information leaks, such as summarising published information or drafting general text.
- May use with conditions: work handling unpublished internal materials. Set conditions on the range of information entered and the services used.
- Must not use: work handling customers' personal information or information covered by non-disclosure agreements.
The key when deciding the categories is to think in units of work. Rather than reaching one company-wide conclusion — 'generative AI may be used' or 'may not be used' — draw the line for each type of work. Drafting sales materials and analysing customer information involve information of a completely different nature. Deciding the category for each type of work makes it possible to use AI boldly where it can be used, and to stop reliably where it cannot. The categories should be decided together with the people who know the work on the ground.
Even after the categories are decided, it helps to designate a contact point employees can consult when they are unsure how to judge. Legal or information-systems staff can answer questions and help with category decisions. Having a contact point lets employees check before proceeding, rather than judging alone when in doubt. Records of consultations also become material for improving the rules.
Information that must not be entered — confidential information and personal information
The most important part of internal rules is to state concretely what information must not be entered. Generative AI services often process entered text on external servers. Whether the entered information is used for training the service depends on the service and the form of contract. That is why the line on what may be entered must be shown in a form employees can understand without hesitation. Writing only 'do not enter confidential information' in the abstract leaves people on the ground unable to judge.
Concrete examples of information that must not be entered include customers' names and contact details, the details of transactions, unpublished earnings information, details of products under development, and employees' personal information. The contents of documents classified as confidential inside the company are treated the same way. By giving such examples, employees can judge that 'this must not be entered'. Conversely, published materials and information within the range of general knowledge can usually be entered without a problem. Giving examples of both what may and what may not be entered makes the rules easier to understand.
Note that, under some enterprise contracts, it is contractually stipulated that entered information will not be used for training. Using a service under such a contract can widen the range of information that may be entered. Which services to use, and under what kind of contract, should also be set out in the rules. Free services aimed at individuals and paid enterprise contracts may handle information differently. Including the choice of services in the rules reduces the burden of judgement on employees.
The range of usable work can also be widened by being creative about what is entered. For example, customer names or specific client names can be withheld, and only general conditions entered. Once information that can identify a person has been removed, the input can be used for analysis or drafting. However, even information that seems to have been withheld can sometimes identify a person when combined with other details. The thinking on anonymisation and processing should also be set out in the rules.
Handling output — the principle of never using it as it is
The output of generative AI can contain errors. Even plausible-sounding text may get numbers, proper nouns, or dates wrong. Explanations of laws or systems may be based on outdated information. Such 'plausible errors' are a characteristic of generative AI output. That is why it is important to include in the rules the principle that output must always be checked by a person, never used as it is.
How to check depends on the work. Numbers and proper nouns should be verified against the original materials or official information. Explanations of laws or systems should be checked against the latest official materials. Documents to be sent outside the company should be finished by the person in charge, taking responsibility, based on the output as a draft. Writing such checking procedures into the rules prevents errors in output from going outside the company as they are. The rules should make it clear that checking output is the responsibility of the person who uses generative AI.
For documents to be sent outside the company, a procedure can also be set out in which, after the person in charge checks the output of generative AI, approval is obtained from a manager or a responsible person. The steps of approval vary with the importance of the information being sent out. Everyday internal documents and proposals handed to customers demand different levels of checking. Deciding the flow of checking and approval according to the importance of the work reduces the risk of incorrect information leaving the company.
Copyright and ownership of what is produced — whose is it?
How copyright in the output of generative AI is treated is still a field where debate continues. Questions such as whether output may resemble existing copyrighted works, and to whom the rights in output belong, have no clear answers in some respects. When writing rules, rather than stating conclusions where none have been reached, it is important to show a way of thinking that avoids risk. For example, when output is to be used as it is in a product or an advertisement, set out a procedure to check that it does not resemble an existing copyrighted work.
It is also desirable to set out in the rules how to treat results produced using generative AI. Materials or text produced by employees using generative AI are generally treated as the results of the company's work. But by stating this clearly in the rules, later arguments about 'whose is it' can be prevented. As for copyright, it is important to decide the company's policy in consultation with the legal department or external specialists.
Attention is also needed on the copyright of what is entered into generative AI. Other companies' copyrighted works, or text whose use is restricted by contract, must not be entered as they are. Even internal materials must not be entered if providing them externally is prohibited. The rules should also set out the thinking on the copyright of information entered, and on contractual restrictions. It is important to consider copyright risks from both the input side and the output side.
Records and accountability — leaving a trace of use
In work that uses generative AI, it is also important to have a mechanism for recording that it was used: which task, which service, and what output was obtained. If that record remains, the origin of output can be verified later. For documents sent outside the company, or materials used as the basis for important decisions, recording that generative AI was used is also desirable from the point of view of accountability.
How much to record should be decided according to the importance of the work. For everyday drafting, records may be unnecessary. For reports submitted outside the company or materials handed to customers, the fact that generative AI was used, and that it was checked, should be recorded. If the burden of recording is too heavy, people on the ground will stop following the rules. Deciding the scope of recording according to the importance of the work is the knack for making rules last.
The way records are kept should also be made as light a burden as possible for people on the ground. For example, when preparing a document to be submitted outside the company, adding a single line to the document management ledger — that generative AI was used and that it was checked — is enough. If the recording mechanism is too heavy, people stop keeping records. Decide the scope of recording according to the importance of the work, and in a form people can keep up.
How to build the rules — deciding them together with the front line
When building internal rules, it is important not to decide them in the management or legal departments alone, but together with employees on the front line. Front-line employees know best what work they want to use generative AI for, and what information would be entered. Rules made without listening to the front line do not fit the work on the ground and become rules that are not followed.
As a way forward, first collect from the front line the work they want to use generative AI for, and decide the 'may use', 'with conditions', and 'must not use' categories for each type of work. Next, write out concrete examples of information that must not be entered, tied to the work on the ground. Then decide the procedures for checking output and the scope of records. Following these steps produces rules that fit the reality of the front line. Once a draft of the rules is ready, it is also important to have front-line employees read it and check whether they can actually follow it.
- Collect the work on the ground: list the work people want to use generative AI for, and the information it involves.
- Decide categories for each type of work: 'may use', 'with conditions', or 'must not use', decided work by work.
- Make the prohibited inputs concrete: show them with examples tied to the actual materials on the ground.
- Decide the checking and recording procedures: set the method for checking output and the scope of records according to the importance of the work.
Once the rules are made, informing employees and providing training are also essential. Distributing the document alone does not convey its contents. Training using concrete examples from actual work, or in-house briefings, helps the meaning of the rules reach the front line. In training, showing examples of what may be used and what may not, tied to actual work, is effective. The aim of training is not merely to know the rules, but to be able to judge by them.
Rules are something to grow — a mechanism for review
The technology and services of generative AI change greatly in a short period of time. New services appear, contract terms change, and legal debates move forward. That is why it is important to build into the rules a mechanism for reviewing them regularly, rather than treating them as finished once made. Set an opportunity to review the rules once every six months or once a year. At review time, check whether the rules are being followed on the ground, and whether there are points that are hard to follow.
The material for review is the voices from the front line. Set up a way to collect voices saying the rules are hard to understand, or do not fit the work. Internal rules are something to be grown with the voices of the front line. If the rules take root on the ground, generative AI becomes a tool that can be used with confidence. Building internal rules as promises to be grown together with the front line, rather than as a document of prohibitions, is the basis for making generative AI useful at work.
It is also important to think in advance about how to respond when a violation of the rules occurs. Simply blaming the violation will make people on the ground start hiding their use of generative AI. Check why the violation happened — whether the rules were hard to understand, or did not fit the work — and use it to improve the rules. Cases of violation are also material for growing the rules. The attitude of adjusting the rules to the reality of the front line is what makes the rules last.
Related Terms
GLOSSARYRelated Articles
RELATEDWhy Semiconductor Plants Gather in the Regions — Reading the Logic of Factory Location
Why are new semiconductor plants built in places like Kumamoto and Hokkaido rather than near big cities? We explain the logic of location through water, power, land, people, subsidies and supply chains.
DX Basics — How 'Digitization' and 'DX' Differ
Digitizing paper and stamps is not the same as changing how the business itself works. This article explains why DX is demanded of Japanese companies, and how to think about its stages.
From Paper to Data — How Electronic Invoicing Is Changing the Face of Accounting
Sending an invoice by email does not make accounting work any easier. In the age of electronic invoicing, the invoice itself travels as data, and Peppol, an international standard, is beginning to spread in Japan. When receipt, bookkeeping, and payment can be connected automatically, how does accounting work change? This article traces the future of invoice automation from the mechanics of the system: the settling-in of the invoice system, its relationship with the Electronic Book Storage Act, and a realistic path forward for small and medium-sized businesses.