Using AI with caution and vigilance

I am a Research Administrator in the EthicsLab, Department of Medicine – Health Sciences. My role routinely involves handling sensitive information and managing programme-related documentation. This requires constant vigilance around confidentiality and compliance with institutional policies.

A regular operational challenge in my role is drafting professional correspondence, emails, and event-related communication efficiently while maintaining a high standard of clarity and professionalism. I also frequently need to research and compare venues, quotations, flights, and accommodation options for conference planning.

I decided to use AI as a support tool — not to replace my professional judgment, but to enhance the efficiency and quality of routine writing tasks. However, I remain very aware that my role involves sensitive information, so I approach AI adoption with deliberate caution and clear safeguards.

The challenge is balancing efficiency with confidentiality, accuracy, and compliance responsibilities. While AI could potentially save time on repetitive administrative tasks, my work often involves sensitive information, contractual considerations, and institutional accountability.

I also need to determine where AI was genuinely helpful and where the risks or inaccuracies outweighed the benefits.

I use Microsoft Copilot as my primary tool for work-related tasks because it was communicated as being safer for confidentiality and institutional context. I prefer Copilot for anything involving university communications, programme information, or potentially sensitive materials.

I also use ChatGPT occasionally for non-work-related purposes or for obtaining a second opinion on non-sensitive tasks, without inserting any work content.

In addition, I use online PDF tools for routine document management tasks such as merging, compressing, converting, and splitting PDF documents where no sensitive content was involved.

My main use of AI is for professional writing support. My approach is iterative and controlled: I draft content myself, insert it into Copilot, and prompt the AI to help refine the readability, wording, clarity, and tone of the correspondence. This allows me to retain control over both the message and my own writing style. I am not asking AI to generate original correspondence, but rather to refine and polish my own work.

I have also used AI for:

• comparing venue quotations, flights, and accommodation options for conference planning
• generating initial summaries of venue information to support decision-making
• obtaining initial research overviews before conducting detailed verification
• verifying historical exchange rate information for audit-supporting documentation

One example involved calculating a colleague’s subsistence funding using a historical Euro exchange rate. Since manually backdating the exchange rate through external platforms would have been time-consuming, I used AI to help confirm the period during which the exchange rate was R18.00 per Euro. This significantly reduced research time while still supporting compliance requirements regarding exchange rate verification.

My prompting approach is iterative and guided by the intended tone, institutional context, and task requirements.

For writing-related tasks, I would first draft the communication myself and then use prompting to refine the wording, tone, and readability. Where I do not agree with AI-generated wording, I prompt further for outputs that are polite, professional, collaborative, urgent, or aligned with university and financial-policy contexts.

For comparative tasks such as conference planning or research verification, I use prompts requesting summaries, comparisons, or confirmation of information before conducting further manual checks.

I found that the more clearly I explained the context and intended outcome, the more useful the outputs became.

For all AI-generated outputs, I apply professional judgment informed by prior training and experience in POPIA, contracts, and policy compliance.

For correspondence refinement, I always review the tone and wording carefully to ensure it aligned with my intent before sending it. For research or comparative analysis tasks, I verify key facts against independent and authoritative sources.

For sensitive or complex matters, I consult colleagues or supervisors where appropriate.

Data protection and POPIA compliance are central to my approach to AI use.

Although I prefer Copilot because it was communicated as being safer for confidentiality, I do not assume any AI tool was entirely risk-free. I exercise my own professional judgment throughout.

I avoid uploading sensitive documents such as contracts or research material. I also refrain from including names or identifying information in prompts and only use summaries or selected sections where necessary.

For any content that could identify individuals, reveal confidential research details, or disclose sensitive institutional decisions, I do not submit it to AI systems at all, regardless of the platform being used.

As an administrator handling sensitive information and programme-related documentation, I apply professional judgment to ensure information is not shared on unsafe or inappropriate platforms that could make it accessible to unauthorised parties or other institutions.

Professional writing refinement saved significant time and effort, particularly for rewriting and polishing correspondence under tight deadlines. AI helped reduce repetition, improve tone, and strengthen the professionalism of written communication.

Comparative analysis for venues and costs was also useful for generating quick overviews to support planning decisions.

Research verification tasks, such as the exchange rate example, helped confirm specific information efficiently when cross-checked against reliable sources.

I also found that AI worked best for tasks involving refinement, iteration, and administrative support rather than final decision-making.

I learned very quickly where AI was useful and where it was not.

I tried using AI for meeting minutes and summaries but found the outputs unreliable. The inaccuracies and the amount of checking required meant there were very limited efficiency gains. I no longer use AI for this purpose because the risk of error in meeting documentation outweighs the time saved.

For accuracy-critical tasks such as financial calculations, compliance documentation, and research summaries, I do not rely on AI without extensive verification.

My experience with meeting minutes became an important learning point. It reinforced that responsible AI use also means recognising when not to use AI.

AI integration into my workflow has been deliberate, cautious, and bounded. I have not integrated AI into every aspect of my work. Instead, I identify specific low-risk tasks where AI adds clear value, particularly professional writing refinement, routine conference-planning comparisons, and research verification.

The outcomes included:

• reduced turnaround time for correspondence
• more polished and professional communication
• improved efficiency when conducting comparative planning research
• greater awareness of the limitations and risks associated with AI use

My workflow remains human-centred throughout. AI assisted with refinement and efficiency, but I remain responsible for verifying, validating, and approving all outputs representing myself or the institution.

I specifically do not use AI for meeting minutes as there are limited efficiency gains because of inaccuracies and the need for extensive checking. I also don’t use AI for confidential correspondence, contract-related work, or tasks involving sensitive institutional information.

I also avoid using AI for tasks where accuracy is non-negotiable unless extensive verification is possible.

Where I am uncertain, I default to not using AI.

Use AI as a support tool rather than a replacement for professional judgment. Always verify AI-generated information against reliable sources and avoid uploading sensitive or confidential information.

For colleagues in administrative and support roles, especially those working with compliance-related or confidential material, it is important to understand institutional policies, POPIA responsibilities, and the limitations of AI tools before integrating them into workflows.

Additionally, colleagues should be particularly cautious with tasks involving legal compliance, confidentiality, or high-consequence errors. Test AI first on routine, low-risk tasks to build judgment about where it is genuinely useful before using it for critical tasks. Remember; you are accountable for the quality and appropriateness of all work that carries your name, whether AI-assisted or not.

The decision to say “no” to AI for inappropriate tasks is just as important as knowing when it can add value.