Generative AI in the workplace: practical rules for protecting documents and business data

Generative AI in the workplace: practical rules for protecting documents and business data

Generative artificial intelligence has entered workplaces through tools for writing, research, translation, analysis, and software development. An employee might use it to summarize a document, draft a reply to a customer, or organize a meeting. The benefit is immediate, but that same action can transfer information outside the company that should have remained protected.

The answer is neither to ban every tool nor to rely solely on individual caution. Companies need understandable rules, approved services, and safeguards proportionate to the value of their data. This guide offers a practical method for using generative AI in day-to-day work without losing control of documents, accounts, and decisions. Its recommendations should be adapted to the industry, contracts, and procedures already in place.

Why a company rule is needed before the first prompt

Without guidance, each person decides which service to use and what content to upload. A sales proposal may end up in a personal account; a contract may be pasted into a tool to obtain a summary; a support screenshot may contain names, addresses, or credentials. Even when the intention is legitimate, the company may lose sight of where that data has gone and who can access it.

An internal rule should answer a few concrete questions: which tools are allowed, which data may be entered, who approves exceptions, who checks the output, and what to do if something goes wrong. A clear, regularly updated page with examples is more useful than a general ban that nobody knows how to apply.

Start with use cases, not product names

Before choosing a service, list the tasks for which AI might be useful. Examples include outlining an article, summarizing notes that contain no confidential information, suggesting variations of a public message, or helping someone understand a non-proprietary code function. For each task, identify who performs it, what information it requires, and what kind of error would be acceptable.

Not every use case carries the same level of risk. Drafting a public webpage differs from analyzing medical records, payslips, confidential bids, or a client’s documents. Separating use cases makes it possible to approve straightforward activities quickly and assess the others individually.

Classify data in a way everyone understands

Effective classification should not become a bureaucratic exercise. It may be enough to distinguish public content, internal documents, confidential information, and data requiring special protection. Each category should include examples drawn from real work and clear guidance on what may be shared with approved AI services.

  • Public: previously published texts, catalogs available online, and information the company freely distributes.
  • Internal: operating procedures, drafts, and organizational information that is not intended for the public.
  • Confidential: unpublished prices, contracts, strategies, proprietary code, and customer information.
  • Especially sensitive: credentials, health data, personal financial information, and other content requiring greater care.

Classification also applies to photographs, audio recordings, screenshots, spreadsheets, and attachments. A seemingly harmless image may show a password on a monitor or a customer’s name in the background.

Choose approved tools and company accounts

The company should publish a short list of authorized tools, stating their permitted uses and the internal contact person for each one. Assessment should cover more than the quality of answers: it should include account management, administrative controls, contract terms, data location, retention periods, and deletion options. Settings may differ between a free plan, an individual plan, and an organizational plan.

Company-managed accounts are preferable for business activities. They allow the organization to assign roles, revoke access when someone changes jobs, and review settings centrally. Shared credentials, by contrast, make it difficult to determine who uploaded a document or changed a configuration.

Read the terms governing how content is handled

Before uploading non-public data, the person responsible should check what the provider says it does with submitted content. It is important to distinguish between processing needed to provide the service, temporary storage, conversation logging, and possible use to improve models. The answers depend on the product, the contract, and the settings actually enabled; they cannot be inferred from the service’s name alone.

It is useful to document who the provider is, what data it receives, where that data is processed, and what controls it offers. When personal data is involved, the assessment should include those responsible for data protection within the organization. A marketing claim such as “secure for businesses” does not replace a review of the terms that apply to the company’s own account.

Protect access and limit permissions

An AI account can become a gateway to conversations, uploaded documents, and connections to other systems. Protect it with a unique password, multifactor authentication where available, and controlled recovery procedures. Administrators should use a limited number of privileged accounts, separate from ordinary user accounts.

If the service connects to email, cloud storage, or work systems, grant only the permissions it needs. An assistant tasked with summarizing one folder has no reason to read the entire company’s files. Before enabling an integration, check what content it can see, whether it can change that content, and how to disconnect it. Review permissions whenever roles, projects, or providers change.

Write useful prompts without copying secrets

A good prompt describes the task, audience, tone, and expected format. It does not necessarily require pasting the original document. To draft a sales reply, for example, the industry, objective, constraints, and a summary of the points to cover are often enough; customer names and confidential profit margins can be left out.

Before submitting content, ask: is this information essential to getting the answer? If not, remove it. If it is essential, check that the tool and its contract are appropriate. The same rule applies to apparently technical requests: logs, configuration files, and code snippets can contain tokens, internal addresses, and sensitive paths.

Anonymizing means more than changing a name

Replacing “Mario Rossi” with “Customer A” may not be enough if the rest of the text still contains an address, order number, occupation, event date, and other distinctive details. The combination of information may still identify the person. To reduce the risk, remove unnecessary elements and generalize details that do not affect the task.

A suitable example might describe “a customer with an annual contract” instead of reproducing the entire clause with identifying references. If the task genuinely requires the original document, treat it as the processing of real data and assess it before uploading. Anonymization is a process that must be checked, not a label applied automatically.

Check attachments, metadata, and file versions

Files contain more information than is visible at a glance. A document may include comments, tracked changes, authors’ names, and hidden text; a spreadsheet may contain hidden sheets; a PDF may include forgotten pages or attachments. Before uploading a file, open the exact version that will be sent and inspect all its contents.

Whenever possible, create an extract tailored to the task instead of sharing an entire archive. Use clearly named working copies, remove unnecessary information, and check the result after export. Screenshots and images require the same care: an area outside the main subject may reveal data that the prompt never mentions.

Be careful with assistants connected to documents

An assistant that searches company files is convenient because it can find information without anyone copying it manually. For that very reason, however, it inherits existing permission problems: if an archive is shared too broadly, its answers may disclose content to people who should not have been able to read it. Configuring the AI service does not automatically fix a folder with inappropriate sharing settings.

Before connecting an archive, check who can see each folder, which documents will be indexed, and how they will be updated or removed. Test questions using accounts with different roles and verify that the answers respect their access rights. Assigning an owner to every connection helps prevent forgotten integrations from continuing to operate after a project ends.

Recognize hidden instructions in external content

AI systems may read webpages, emails, and documents the company does not control. Such content can contain text that tries to act as an instruction to the assistant—for example, telling it to ignore the user’s request, reveal data, or follow an unrelated link. This risk is called prompt injection and particularly affects tools that can access files, browsers, or other actions.

Employees should not trust an instruction simply because it appears in a document retrieved automatically. Actions that change data or communicate outside the organization require human confirmation, limited permissions, and separate checks. OWASP’s guidance on prompt injection explains why external text can change the behavior of an AI application.

Check every answer before using it

Fluent writing can still contain errors, nonexistent references, or overly confident conclusions. Anyone using AI should check numbers, names, sources, dates, and operational steps before sharing the result. This is especially important for proposals, customer communications, security instructions, and material that could influence a financial decision.

The review should have an identifiable owner. An assistant can prepare a first draft or suggest alternatives, but it does not replace the expertise of the person who signs the document. When an answer cites a source, open the original document and check that the cited passage actually supports the claim. A plausible citation is not evidence until it has been verified.

Use AI in customer communications

A generated reply can make a message clearer, but it should be reviewed before it is sent. Check that it does not promise timelines, prices, or results the company cannot guarantee. It helps to prepare approved templates for recurring requests while leaving the employee responsible for adapting them to each case.

If a customer’s request contains personal data or contractual information, do not paste it into an unauthorized service. Instead, describe the issue in general terms or use an environment approved for that type of processing. When a chatbot replies directly on the website, it must also be clear what information it may use, which cases it passes to a person, and how errors can be corrected.

Protect source code and configurations

Developers may ask for an explanation of a function or suggestions for a test, but they must carefully review what they submit. Code can embed confidential business logic; configuration files can contain credentials; logs can include real users’ data. Remove secrets and unnecessary references before sharing a snippet.

AI-suggested code also requires review, testing, and security checks. It should not be run in production merely because it “looks right.” For a fix to a live system, follow the normal development process: a working copy, a review of changes, tests, approval, and a way to restore the previous version. NIST’s practices for secure AI system development provide a broader reference.

Handle HR tasks and assessments of people with care

Résumés, interviews, performance reviews, and internal reports contain sensitive information and can have real consequences for people. They are not suitable for improvised trials using personal accounts or tools chosen by a single department. Before using AI in these processes, involve human resources, those responsible for data protection, and whoever assesses the applicable obligations.

AI can help improve the wording of a job advertisement or organize criteria that have already been defined, but decisions about people must not be delegated to generated text without oversight. The organization must be able to explain what information was used, who checked the result, and how errors or biased assessments can be corrected.

Create images and content without exposing confidential material

Image generation is useful for articles and presentations, but a prompt can reveal an unannounced project, a customer’s name, or features of a product under development. Use general descriptions when specific details are unnecessary. If uploading a reference image, check that its use is authorized and that it does not show identifiable people or elements without a sound reason.

Before publishing, check quality, consistency with the brand, possible visual errors, and the chosen service’s terms of use. A synthetic image should not be presented as documentary photography of a real event. Keeping the prompt and the editorial review can help explain how the material was created.

Define retention and deletion rules for content

Conversations and uploaded files may remain available in an account or in system copies under the service’s terms. Establish a rule for who may retain conversation history, how long drafts remain useful, and how to delete material uploaded by mistake. Check the actual settings, because products do not all offer the same controls.

Separate materials the company must retain in its own archive from temporary prompts used during work. If a draft becomes an official document, the approved version belongs in the proper document system, with defined ownership and access. A chatbot’s conversation history should not be the only place where an important decision is recorded.

Prepare an incident response

If an employee accidentally pastes a password or a confidential file, the first step is to notify the internal contact promptly, not to hide what happened. A credential may need to be revoked and replaced. For a document, the company needs to establish which service and account received it, which settings were active, and whether the provider offers deletion tools or assistance.

The procedure should identify who gathers information, who assesses the impact, who contacts the provider, and who decides on any further steps. Recording the incident helps address its cause: an unclear rule, excessive permissions, or a process that forces people to copy unnecessary data.

Train staff using examples from real work

A generic presentation about AI is not enough if people must then decide on their own how to handle a contract or a complaint. Training should use examples from the organization’s work: a prompt that is allowed, one that needs changing, and one that must not be sent. People should learn to recognize confidential data, check sources, and report mistakes without fear.

In the European Union, Article 4 of the AI Act requires measures to support AI literacy among people who operate these systems on behalf of providers and deployers; the updated text takes account of the context of use and people’s experience. The European Commission explains the AI literacy obligation. Practical training remains useful regardless of legal obligations.

Launch a measurable pilot project

To begin, choose one or two low-risk use cases and define what “working well” means. Possible measures include time saved, draft quality after review, number of corrections, and satisfaction among internal users. Record errors, situations the tool could not handle, and information employees had to remove before using it.

A pilot should have a duration, an owner, and criteria for deciding what comes next. At the end, the company decides whether to expand, change, or stop its use. The NIST profile for generative AI recommends managing risks throughout the lifecycle, rather than treating the initial choice of product as sufficient.

Write an internal policy in ten points

The policy can be short, as long as it is specific and can be put into practice. An initial template might contain the following points, adapted to the organization’s context:

  1. A list of authorized tools and accounts.
  2. Tasks for which AI may be used and cases requiring approval.
  3. Categories of data that must not be entered into unauthorized services.
  4. Rules for attachments, screenshots, audio, and links to archives.
  5. Use of company accounts, authentication, and minimum necessary permissions.
  6. Human review of facts, sources, and content before use.
  7. Responsibility for public texts, customer communications, and code.
  8. Retention, deletion, and management of approved versions.
  9. A procedure for reporting errors or accidental sharing.
  10. Regular policy reviews as tools and processes change.

Each point should name a person or role people can turn to for help. A rule without a contact person risks becoming an obstacle rather than an aid to work.

Frequently asked questions about generative AI in the workplace

Can I paste a contract into a tool to have it summarized?

Only after checking that the tool is authorized for that type of document and that uploading it complies with the agreements with the client. If you only need to understand a general clause, you can often describe the issue without sharing the entire contract. If in doubt, ask the company contact before sending the file.

Does a paid account automatically make data secure?

No. Price and security are not the same thing. Check the terms of the specific plan, the active settings, access management, and the kind of information being submitted. Even a service with advanced controls can be used poorly if an account is shared or connected to archives that are too broad.

If I remove the customer’s name, is the text anonymous?

Not necessarily. Dates, addresses, amounts, or unusual circumstances may still identify them. Assess the information as a whole and reduce the details to the minimum needed for the task.

Can I publish AI-generated text immediately?

It is better to treat it as a draft. Before publication, check facts, sources, tone, rights to the materials used, and consistency with the company’s message. A person must take responsibility for the final version.

Conclusion: make AI useful without losing control

Generative AI can ease repetitive work and help produce better content. Its benefits, however, depend on the rules surrounding it: classify data, choose appropriate tools, limit permissions, check answers, and handle errors. There is no need to build a complex system immediately; start with a few clear use cases and improve procedures based on experience.

For a small or medium-sized business, the most useful next step is often simple: appoint a contact person, publish a short policy, train people with concrete examples, and review tools and access regularly. AI then remains a work aid, while important decisions and information stay under the company’s responsibility.

Would you like to explore this topic for your business? Azienda Digitale can help you assess the right solutions and put them into practice. Azienda Digitale

Lascia un commento

Your email address will not be published. Required fields are marked *

Your comment will be visible after approval by the administrator.

Il commento sara visibile dopo approvazione dell amministratore.