Key takeaways

AI can enhance creativity, efficiency, and problem-solving when used thoughtfully. Transparent, client-opt-in adoption builds trust and safeguards values. Human oversight is essential, and close attention is required to environmental, data privacy, fairness, and bias risks.

“AI” is wildly hard to define, we’re primarily talking about generative AI

The term “AI” is loosely applied across industries. AI ultimately learns by itself — either from data fed into a model or iteratively through feedback. Within AI there is machine learning, deep learning, and Generative AI, where deep or machine learning is applied to create outputs.

Basic terminology you may hear about

Key categories include the AI hierarchy (AI, Machine Learning, Deep Learning, and Generative AI), learning methods (supervised and unsupervised), and function types such as forecasting, classification, clustering, natural language processing, image recognition, and recommendation systems.

Our quest for meaningful positive impact

The organisation exists to “Unlock the potential of the new economy to leave behind a better world.” This involves scrutinising the tools used — considering their carbon footprint, data practices, profit flows, and intellectual property implications.

People and the planet are physically impacted by AI

Evidence shows AI causes tangible physical harm. Writing an email with AI consumes 3 litres of water (p.5), and AI water usage is predicted to reach 6.6 trillion litres per year in 2027 (p.5). CO₂ in the US is projected to increase by 0.4–1.9% purely from AI growth (p.5). One data centre uses as much water as 100,000 homes (p.5). Data centres also emit PM2.5 including Nitrogen Dioxide and Sulfur Dioxide (p.5). Additionally, 40,000 children aged seven and above are involved in cobalt mining in the DRC (p.5).

A trade-off decision

AI usage decisions are case-by-case and subjective. The most effective tools often deliver outcomes by an extreme margin, creating an internal conflict between productivity and ethics — with no clear line between the two.

AI bias & fairness

AI systems trained on skewed datasets can unintentionally reinforce stereotypes or under-serve marginalised groups. Human oversight, diversity of input, and context awareness are maintained as essential safeguards, particularly in recruitment, content guidance, and client-facing outputs.

Trials and tribulations / The illusion of productivity

AI has been trialled extensively across services. Despite impressive initial outputs, fabricated references and hallucinated content quickly emerge. Logic loops can cost significant time. Productivity gains are unclear and vary by task.

It can be thorny too

Security and privacy risks are significant. Documented AI-related breaches include over $1 billion in losses (p.9), covering Samsung’s semiconductor code leaks, Microsoft’s 38TB data exposure, Zillow’s $500 million algorithmic failure, and a McDonald’s recruitment chatbot breach affecting 64 million applicants. Security researchers consistently find that 30–50% of AI-generated code contains exploitable vulnerabilities (p.10).

You are unwillingly participating

AI is embedded across many platforms without explicit consent. The organisation disabled AI by default across Google Workspace, Slack, Hubspot, and Zoom, finding the benefit to be zero or extremely low in all four instances (p.11).

Become a willing participant

Five modes of AI use are outlined: No AI, Researcher, Assistant (Third-party), Assistant Generator, and Private. Each carries different levels of data sharing, risk, and productivity uplift. Clients are given the option to direct how AI is used in work conducted on their behalf.

When we use various modes of AI

Recommended uses include coding assistance, ideation, editing, wireframing, analytics, and survey design. High-risk or low-value uses to avoid include transcribing meetings, generating brand new visuals, direct data analysis with core systems access, and managing security credentials. The default position can be adjusted per client.