Kirsty Holmes
Specialist role (Human Resources / Organisational Design)
Context The challenge How I used AI Prompting examples Pre-checks Risks What worked wellI Human input required Outcomes When not to use AI Good practice tips
Context (role & task)
As an Organisational Design Specialist, I recently had to present a policy proposal to a senior management committee for buy-in and approval. I had written the proposal from the perspective of good HR specialist practice, but I knew the audience would approach it very differently. Management would naturally focus on operational implications, cost, institutional risk, and implementation concerns — some of which I had not fully considered myself.
I also realised that the audience I was presenting to was very different from the HR audiences I normally engage with. The presentation therefore needed a different tone and structure if I wanted to secure meaningful buy-in.
The challenge
I needed to anticipate the kinds of questions they would ask and prepare myself to respond strategically rather than technically. In essence, I needed to stop thinking like me and start thinking like a senior manager.
How I used AI
To help me prepare, I used NotebookLM to review the policy proposal, supporting policy documents, and several research sources that informed the work. I asked it to help me create a structure for the presentation, along with concise speaking notes and cues that would help me communicate the proposal more effectively to the committee.
I gave the tool additional context, including the portfolios of the people attending the meeting, the amount of presentation time I had available, and some of the concerns that had already been raised in earlier discussions. One of the reasons NotebookLM worked particularly well for this task is because it is source grounded. It relies on the documents you upload rather than drawing information broadly from the internet. This reduces the likelihood of hallucinations or unsuitable outputs and allows it to provide citations linked directly to the uploaded material.
In a university environment where accuracy, traceability, and context matter, this became extremely useful. It also helped me process large amounts of information quickly by identifying patterns, surfacing contradictions, and connecting themes across multiple documents. Most importantly, it helped me organise, simplify, and consolidate my own thinking.
Example Prompting Approach
I found that the quality of the output depended heavily on the quality of the prompts I used. I approached this in several stages.
- First, I used role prompting to establish context:
- “I am an Organisational Design Specialist at a higher education institution in South Africa.”
- “I need to present to a team of senior managers, including the Chief Finance Officer and Property and Custodial Services executives.”
- I then used contextual prompting to explain the institutional challenge and the intended solution:
- “This is the current institutional challenge, and these are the proposed solutions.”
- “This is the way that the University currently operates, but these are the risks and problems we experience as a result. This is how we intend to solve that problem. I need help to get them to buy-in to the idea.”
- “I only have five minutes of presentation time at the end of a two-hour meeting.”
- From there, I added instructional prompts such as:
- “Focus on high level strategic benefits, cost saving, risk mitigation and operational framework.”
- “Compare these versions and provide a table showing the differences.”
- I also included examples of concerns that had already been raised and used negative prompting such as:
- “Don’t use corporate jargon.”
- Finally, I used prompt chaining to test my thinking further (where one response informed the next question) by asking questions like:
- “What concerns are likely to be raised by senior management?”
- “Can you provide an analogy suitable for higher education that I can use to strengthen understanding?”
- “What have I not considered?”
This helped deepen the analysis and improve the usefulness of outputs.
What I checked before using the output
In practice, I use AI mostly as a thinking partner — something to help me brainstorm, challenge my ideas, structure information, or prepare my thinking before engagement with stakeholders. In this case, the AI-generated content functioned as speaking notes and preparation cues. The audience never saw the actual AI output itself.
I use AI to help with the heavy lifting — much like using Excel to work through large data sets — but I never treat the output as automatically correct. I double check it, validate it, and then rework it into my own voice and ideas before using it professionally.
Risks I considered
There were also important ethical and privacy considerations. The information I used was not confidential or sensitive, and there was nothing that required the removal of UCT’s name. If the material had been sensitive, I would have used an institutionally approved environment such as Microsoft Copilot instead. I also appreciated that Google explicitly states that sources, notes, and uploaded documents remain secure and are not used to train its models.
Where AI contributes meaningfully to written work, I also believe in being transparent about its use, even if it is simply a short acknowledgement such as “Thanks to NotebookLM!”.
What worked well
One of the most valuable aspects of the process was that AI helped me see my own work differently. I know I can sometimes become too immersed in the detail of a proposal, and the tool helped me step back, summarise more effectively, and identify what actually mattered most for the audience I was addressing. It also helped me test my assumptions by presenting perspectives I had not initially considered.
What still needed human input
The quality of outputs depended heavily on the quality of prompting, the context provided, and a clear understanding of the intended audience.
For example, I had not included in my prompts that the proposal would also be presented to Deans. As a result, the output focused more heavily on operational risk, cost, and PASS management concerns than on teaching, learning, or research considerations. This highlighted how important audience-specific prompting is when working with AI tools. What you ask AI for is what you get, and the quality of what you get out is all about what you give it.
Outcomes
AI became integrated into several parts of my workflow, particularly in policy development, consultation and presentation.
AI tools help improve efficiency, support synthesis across multiple information sources, and create more space for strategic thinking and review. However, interpretation, contextualisation, and final decision-making remain my responsibility. I rework what AI gives me into my own voice and ideas.
When I would NOT use AI
I would avoid using AI when confidential HR matters are involved, when identifiable staff information is included, or when institutional data classifications restrict sharing. I would also avoid relying on AI outputs in situations requiring nuanced institutional judgement without appropriate human review.
Good practice tip for colleagues
For colleagues wanting to explore AI in similar workflows, my advice would be simple: the quality of what you get out depends entirely on what you put in. Start with clear context. You need a strong understanding of your own content area in order to critically evaluate the output. AI outputs improve significantly when the audience, purpose, and institutional setting are clearly explained in the prompt. It is also important to treat AI as a support tool for thinking and structuring rather than as an authoritative source.
The more intentional the prompting process becomes, the more useful and reliable the outputs are likely to be.
Footnote: What is NotebookLM and how can UCT staff access it?
NotebookLM is a Google AI-powered research and synthesis tool that allows users to upload source documents such as PDFs, reports, meeting notes, or presentations and then interact with the material through summarisation, questioning, comparison, and note-generation features. Unlike general AI chat tools, NotebookLM grounds its responses in the documents uploaded by the user, which can help improve traceability and contextual relevance. UCT staff can access the tool using their institutional Google Workspace credentials where enabled through UCT’s Google environment.