Why HR Won't Use AI At Work
By Paul Marsh (not AI)
HR’s hesitation to embrace AI is understandable. HR teams handle salaries, health information, performance reviews, disciplinary matters and other details that can have serious consequences if misused. Staff may worry that entering personal information into an AI tool sends it somewhere they cannot control. Those concerns sit alongside practical barriers: limited confidence with AI, uncertainty about what it can do, and fear of getting an answer wrong.
But treating “AI” as one category (and rejecting it outright) can leave organisations stuck. The question is not whether HR data exists in digital systems - it already does. Organisations store employee records in HR platforms, payroll software, email, shared drives and recruitment systems. People also publish professional information on LinkedIn and other public sites. This does not mean that data is free to reuse or that any AI provider can safely receive it. It does show that the real challenge is governing access, purpose, retention and security across every system, including AI.
GDPR and data protection law need not be a blanket ban on AI. They require organisations to handle personal data lawfully, fairly and transparently, and to apply safeguards appropriate to the risks. HR teams should therefore start with the task and the data involved. Drafting a general interview guide or summarising a fictional policy scenario is very different from asking a public chatbot to analyse named employees’ performance records.
A practical approach is to introduce AI in stages. Begin with low-risk work that uses no personal data: drafting job-ad templates, creating onboarding checklists, or turning a generic policy into plain language. This gives staff room to build skills and judgement before sensitive cases are involved. Training should cover how to check outputs, recognise bias and avoid treating generated text as a decision-maker.
For more sensitive uses, minimise the data. Remove names and direct identifiers, use fictional examples and share only the detail needed for the task. Pseudonymising information can reduce exposure, but it does not make data anonymous if someone could still be identified by combining details. Avoid putting special-category data (such as health stuff) into a tool unless the organisation has assessed the use and put suitable protections in place.
Choose tools through procurement and security review, not convenience. Confirm whether prompts and uploaded content are used to train models, how long they are retained, where they are processed, who can access them, and what deletion and access controls are available. Set clear rules on approved tools, prohibited information, human review and record-keeping. For high-risk uses, and especially those that affect recruitment, promotion or employment decision, complete a data protection impact assessment and involve the data protection officer, security team and affected staff.
The aim is not to make HR less cautious. It is to turn caution into a workable framework. With clear boundaries, safer tools and practical training, HR can test useful applications while protecting people’s information.
Lightbulb © 2026
