Demystifying AI: DIY vs Tech Guru

Notes

1. Opening: The Hype vs. The Reality (8 min)

• The productivity paradox: research shows real gains, but time saved is often consumed by verification, debugging, and adapting outputs to business processes
• The "brilliant but unreliable intern" framing — introduce it here as the lens for the whole session (Jamie Bartlett, How to Talk to AI)
• A brief, honest map of where AI sits today: genuinely transformative in some areas, genuinely overhyped in others

2. What LLMs Actually Are—and Aren't (10 min)

• Creativity vs. Knowledge: LLMs excel at making unexpected semantic connections, not at storing and accurately regurgitating facts
• They are pattern-completion engines, not search engines or databases — a crucial distinction most users never grasp
• The knowledge cutoff and the confidence problem: models don't know what they don't know, and they'll fill gaps with plausible-sounding fiction
• Hallucinations and error rates: not a bug that will be fixed — a structural feature of how these systems work

3. Practical Applications: Where to Actually Use This (18 min)

High-impact use cases:

• Parsing legalese, contracts, and long documents — summarise, extract obligations, flag ambiguities
• Drafting and rewriting communications — emails, reports, policy documents: AI as a first-draft engine, human as editor
• Code assistance and scripting — even for non-developers: automating repetitive Excel/Power Automate tasks, writing SQL queries, generating regex, basic Python scripts for data wrangling
• Meeting notes and synthesis — feeding transcripts or rough notes into an LLM to produce structured summaries and action items
• Research triage — rapidly surveying an unfamiliar topic to know what questions to ask, not to get final answers
• Internal knowledge base queries — using AI interfaces on top of your own documentation (with appropriate tooling)

4. Getting Better Outputs: Prompt Engineering That Actually Works (12 min)

• Roleplay and persona prompting — assigning a role unlocks out-of-the-box thinking and shapes tone and framing
• Think out loud (chain-of-thought) — instructing the model to reason step-by-step before answering dramatically reduces errors
• Asking for confidence scores and caveats — explicitly prompting the model to flag uncertainty
• Giving context, not just commands — the more the model knows about your situation, constraints, and audience, the less you have to fix
• Iterative refinement — treating a prompt as a first draft, not a one-shot request; the conversation is the tool
• Negative prompting — telling the model what not to do is often as important as telling it what to do

5. The Risks You Actually Need to Manage (8 min)

• Deskilling — over-reliance degrades the human expertise needed to catch AI errors; the intern analogy bites back
• Data privacy and confidentiality — what you should never put into a public LLM (client data, IP, regulated information); brief mention of enterprise/private deployment options
• The verification burden — every output needs a human owner; AI shifts work, it doesn't eliminate accountability
• Bias and blind spots — models reflect their training data; outputs on sensitive topics need extra scrutiny
• The audit trail problem — AI-assisted decisions are harder to document and defend; worth flagging for compliance-conscious IT environments

6. Our Expertise (4 min)

• Where Cegos Group's white paper insights are most relevant: large-scale deployment, change management, risk frameworks
• Our workshop micro apps use LLMs to help participants surface insights, spot connections, and make sense of their own experiences in real time