In the second of HEFMA’s webinar series with CompliMind, exploring the fundamentals of artificial intelligence and using it safely and productively within healthcare Estates and Facilities services, Carl-Magnus von Behr, CompliMind founder focused on how to make the AI-generated output more reliable. Fundamentally, this rests on how good the initial prompt to the AI is.
The prompt is essentially the brief, the problem you want to solve or the task you want to complete, and the more time spent at the outset defining the prompt the better the output is likely to be.
Prompts usually fail for one of four root causes:
• Missing assumptions, so not specifying the type of hospital, for example acute or mental health, and location, such as England as opposed to Wales or Scotland
• Failing to specify the output format, for instance, a chart, a table, a bullet-pointed list or a word-limited passage of prose; not being precise risks the AI generating multiple pages of material that might wander off-point
• Task stacking, that is, including too many tasks in one prompt, which could also result in a critical stage not being clearly specified, will stretch the AI potentially resulting in inaccuracies that are harder to identify later on; taking it step by step produces better output
• Being vague, not providing sufficient context, constraints or jurisdiction.
Using AI
There’s a distinction to be made between generative AI – which will generate text, calculations, designs or whatever you ask of it, confidently, plausibly and quickly, but not always accurately – and assurative AI – which can be used to verify content that has already been generated against checklists for correctness, compliance and missing requirements.
Carl-Magnus introduced two frameworks to apply to working with these technologies – CARE and CRIT:
• CARE = Context, Ask, References, Expectations and applies to the process of working with generative AI, helping to generate the output.
• CRIT = Clarify, Review, Improve, Test and applies to assurative AI, helping to clarify the output.
Key takeaways
• Spend time getting your brief right; AI will start to generate information as soon as you start to ask it, and will make assumptions to fill in the gaps which can be difficult to find once they are layered into a document
• Don’t outsource your critical thinking to AI
• Assumptions made by AI will be using US-centric data; so it’s critical to provide the context and specify references
• Treat the first output as a draft
• Task stacking degrades the output and makes it more difficult to identify mistakes
• Free AI tools will retain data for training purposes and therefore bring high exposure to risk; assume loss of control of your data if it’s input to a free AI tool
• Keep records; the engineer is accountable, not the AI tool so it’s essential that organisations can justify how AI has been used and that outputs have been verified with engineering expertise
• AI can support analysis, drafting and options, but humans must make the decisions
• Be aware of what risk level is attached to a process if you are using AI.
The HEFMA and CompliMind partnership aims to support HEFMA members in navigating the growing complexity of compliance, knowledge management and digital transformation. It establishes a member-led approach to the safe and meaningful adoption of AI within NHS estates and facilities services. It will enable the sharing of best practice across Trusts and empower estates and facilities teams to drive the development of AI as well as other digital tools, and their integration into everyday processes, in a way that is controlled, transparent and aligns with professional values.
Future webinars
These webinars are free to attend for HEFMA members. Please log into this site, navigate to MyRegion and you’ll find the registration links in your diary dates and Branch notices.
• Streamlining the paperwork: minutes, action logs, and reports with AI
Wednesday, November 18, 2026, 12:30 – 13:30
• AI in practice: real use cases, benefits, and lessons learned
Wednesday, December 9, 2026, 12:30 – 13:30
If you missed either of the first two webinars in the series, the plan is to repeat them early in 2027. Details of dates and how to register will be shared in due course.
