AI in HR for Small Teams: Where to Start and What to Check

You have a new starter arriving on Monday, a policy update to explain and a job advert that still needs writing. AI can help produce a first draft for each. The useful question is whether the result takes less work to check than it would take to write yourself.

For a small team, start with one recurring task whose output you can verify. Agree which tool and information people may use, who checks the result and when it is ready to share. These five examples offer a starting point.

Five tasks to try first

1. Turn a role brief into a job description

Give the tool the role's responsibilities, essential skills, working arrangements and approved salary information. Ask it to organise those facts into a clear advert and flag anything missing.

Compare the draft with the actual job. Remove invented benefits, unnecessary qualifications and wording that could exclude suitable applicants. The hiring manager should confirm every commitment before publication. Keep candidate CVs out of this drafting exercise.

2. Build an onboarding checklist

Use a generic role description and approved onboarding steps to draft a first-week plan. Ask for a suggested owner and timing for each task, then have the people involved confirm them.

Check access, equipment, training and introductions against how your company works. Local registration deadlines and employment requirements need a separate check. A plausible checklist can still omit something essential. Our European onboarding checklist is a useful starting reference.

3. Explain an approved policy in clearer language

Once the tool is approved for the document's confidentiality level, provide the current policy and ask for a short explanation with references to the relevant sections. Tell it to identify questions the policy does not answer.

Compare the explanation with the source, especially exceptions, eligibility and deadlines. Keep a link to the approved policy so employees can check the full wording. Ask the person responsible for the policy to answer questions about an employee's own entitlement.

4. Draft routine employee communications

A reminder about a training session or a change to office opening hours can be a manageable first experiment. Supply the confirmed dates, audience and action required. Ask for a short, direct message.

Check that the draft creates no unintended promises and reaches the right people. For the initial pilot, leave disciplinary matters, grievances, health information and individual pay discussions outside the approved scope.

5. Prepare a translation draft

AI can produce a working translation of a welcome message or general announcement. Give it the audience, preferred level of formality and terms that should remain unchanged.

Have someone competent in both languages compare the meaning, including dates, instructions and tone. Contracts, privacy notices and legal rights need appropriate linguistic and subject-matter review before use. A fluent translation can alter an obligation without making the error obvious.

Decide what information can go into the tool

Start with public text, invented examples or internal material the tool is approved to handle. Internal policies may contain confidential information even when they contain no employee records, so check the tool's terms and your organisation's rules before uploading them.

Before using personal data, establish the purpose and lawful basis, limit the information to what is necessary and assess any additional conditions for special-category data, such as health information. Consent is usually unsuitable in employment because employees may not feel free to refuse their employer. EDPB: lawful processing

Check whether a data protection impact assessment is required before starting. GDPR Article 35 requires one where the proposed processing is likely to create a high risk to people’s rights and freedoms. GDPR, Article 35

Removing a name does not necessarily anonymise a record. A role, location and distinctive event may still identify someone in a small team. Pseudonymised information remains personal data when it can be linked back to an individual. EDPB: anonymisation and pseudonymisation

Ask the provider who processes prompts and outputs, where processing and support access occur, how long information is kept, whether it is used for training and who can retrieve it. Where a supplier acts as your processor, an Article 28 arrangement is required. International transfers need the relevant GDPR safeguards; an EU hosting location alone does not settle the assessment. GDPR, Articles 28 and 44–49

If a provider uses inputs for its own purposes, assess that processing separately. Keep personal data out of the pilot until the arrangement is assessed and approved. Our provider questions cover the checks in more detail.

Give the reviewer something concrete to check

Use a prompt that ties the draft to known facts. For example:

Draft a first-day checklist for a fictional office-based administrator. Use only these confirmed tasks: collect equipment from IT, meet the manager, attend the scheduled safety induction, review the approved handbook and test system access. Suggest an order. Mark missing times and owners as “to confirm”. Do not invent company policies or legal requirements.

Check that every confirmed task appears, no unsupported requirement has been added and unresolved details remain visible. The instruction helps define the task; it does not guarantee the model will follow it.

For the first few attempts, record drafting time, checking time and corrections. Avoid putting personal data into that pilot log. If checking repeatedly takes longer than doing the task yourself, change the task or stop the experiment.

Treat decisions about people as a separate project

Ranking applicants, scoring performance or recommending dismissals requires a different assessment from drafting a generic welcome message. GDPR Article 22 concerns solely automated decisions with legal or similarly significant effects, subject to specified exceptions and safeguards. GDPR, Article 22

A human approval step does not by itself settle the legal assessment. The AI Act separately addresses certain employment systems as high-risk, and its requirements have different application dates. Check the intended use and current rules before introducing such a system. European Commission: AI Act

Train people on the approved tasks, permitted information and likely mistakes. Include practice at checking an incorrect answer and a clear route for reporting problems. Article 4 of the AI Act requires providers and organisations deploying AI systems to take measures supporting the AI literacy of people using those systems on their behalf. The Commission says no particular certificate is required. Keep a record of the measures taken. European Commission: AI literacy

Your first-pilot checklist

Use this as a practical planning aid alongside the legal and organisational checks relevant to your use case.

• Choose one recurring drafting task and a small trial period.
• Name the approved tool, account and person responsible for the pilot.
• Specify permitted information and what must stay outside the tool.
• Check supplier terms, access, retention and any personal-data processing.
• Check whether introducing the tool requires consultation or agreement with employee representatives under applicable national rules.
• Provide a current source document or confirmed facts.
• Give the reviewer time, knowledge and authority to correct or reject the output.
• Show the team common errors and how to report a concern.
• Compare the complete effort, including review, with the current process.
• Decide whether to continue, change or stop before expanding the scope.

A useful first result is a task your team can complete more easily while keeping the information accurate and the responsibility clear.