AI for Leaders

This opening module makes the case that AI is a leadership question, not an IT project. It traces a short history from Alan Turing's 1950 paper to ChatGPT and the Welsh Government's AI Cymru plan, uses Local Government Association and Work Trend Index evidence to show AI is already in use, and sets out four kinds of AI, five questions to ask of any use, and the principle of owning the decision.

This module explains, without mathematics, how large language models produce answers by predicting probable text, and why plausible is not the same as correct. Drawing on Agrawal, Gans and Goldfarb's idea of AI as a prediction machine, the Ayinde v Haringey fake citations case and Parasuraman and Manzey's work on automation bias, it sets out where AI helps, where it fails, and practical literacy habits for checking.

Evidence first, enthusiasm second. This module reviews the strongest studies of AI and productivity: Brynjolfsson, Li and Raymond on customer support, Dell'Acqua and colleagues' jagged technological frontier, the METR trial of experienced developers, the civil service Copilot evaluations and the contested MIT NANDA claim that 95 per cent of pilots fail. It shows why time saved is not value created and why value is measured in outcomes.

Strategy begins with what residents and staff need, not with the tool. This module builds a layered picture from purpose and outcomes down to governance and trust, compares three strategic postures, and offers a value and feasibility matrix for choosing use cases. It shows how to set out an AI strategy on one page, what boards should ask, and a test, learn, then scale or stop cycle in which stopping a pilot counts as a result.

Policy sets the weather, but you choose the route. This module maps UK, Welsh and European AI policy: the 2023 White Paper's five principles, the AI Opportunities Action Plan, the UK government's AI Playbook, the Welsh Government's AI Cymru plan and Workforce Partnership Council guidance, and the EU AI Act. It shows that while there is no single UK AI law, plenty of existing law applies, and accountability stays local.

Every AI output inherits the strengths and flaws of the data beneath it. This module sets out five foundations beneath any AI system, five tests of data quality, and the idea that a model can only learn from the history its data contains. Through the 2020 A level grading algorithm and Robodebt it shows how data problems harm people, then covers data protection essentials, who is missing from your data, and practical steps to get data ready.

You can buy the tool, but you cannot outsource the accountability. This module covers the choice to build, buy, partner or wait, supplier due diligence questions, and why awkward questions belong before signature. Using the Deloitte Australia AI-assisted report case, it sets out five safeguards for commissioned work, five contract essentials, ways to avoid supplier lock-in, and a route from idea to procurement in which the contract makes values enforceable.

AI does not sit outside the law. This leader's guide, which is not legal advice, maps data protection, equality, public law, human rights and sector rules, explains the reform of automated decisions under the Data (Use and Access) Act 2025 and why a rubber stamp is not a safeguard, and sets out the EU AI Act's timetable and four risk tiers. It closes with the regulators who will ask questions and why lawful is the floor, not the ceiling.

Fairness is a choice, not a setting. This module traces five places bias enters AI and shows how bias scales inside a system, using Amazon's experimental recruiting tool, the COMPAS and ProPublica debate and the ICO's audit of AI recruitment tools. Drawing on Kleinberg, Mullainathan and Raghavan and on Chouldechova, it explains three meanings of fairness that cannot all be met, the OECD AI principles, and an ethics check before go-live.

Oversight you cannot see is oversight you do not have. This module separates hollow from meaningful human oversight, insists on a named owner for every AI system, and sets out five building blocks: a register, rules for use, impact assessment, monitoring, and challenge with openness. It applies the three lines model and NIST's Govern, Map, Measure, Manage cycle, reviews standards boards hear about, and describes four board postures.

When systems fail, people pay. This module studies Robodebt in Australia, the Dutch childcare benefits scandal, the Post Office Horizon scandal, the 2020 grading algorithm and the DWP model for Universal Credit advances. Comparing them side by side, it draws out five recurring patterns, from warnings ignored to accountability diffused, and argues that efficiency which harms people is not efficient.

Tasks change before jobs do. This module reads global projections from the IMF, the World Economic Forum and Eloundou and colleagues, explains why exposure is not replacement, and offers a spectrum for deciding which tasks to automate, augment or keep human. It reviews early evidence from Stanford's Canaries in the Coal Mine study and British Progress, the Klarna U-turn, and why the first rung of the career ladder deserves watching.

Do it with people, not to them. This module sets out the Workforce Partnership Council's five principles for managing technology that manages people, with five steps before implementation and six expectations afterwards. It covers transparency, algorithmic management and the ICO's guidance on monitoring workers, routes to worker voice in Wales and England, the 2023 Hollywood writers' strike, and the difference between consultation as formality and as design.

Everyone needs enough AI literacy to ask good questions, but while the tool is cheap, skill is not. This module sets out literacy at four levels, why literacy is now an expectation under the EU AI Act, the UK skills target, AI Cymru and LGA evidence, and the expertise pipeline problem, illustrated by the legal profession. Drawing on Bainbridge's ironies of automation, it offers ways to protect how people learn and a skills plan.

Your staff are already using AI. This module uses Microsoft and LinkedIn's Work Trend Index and LGA survey evidence to explore shadow AI, compares banning with safe enablement, and sets out four responsibilities for leading a team through adoption and what makes an acceptable use policy people follow. Drawing on Amy Edmondson's psychological safety and William Bridges on transition, it attends to feelings as well as facts and offers a team adoption rhythm.

Public trust is the licence to use AI at all. This module explores how AI can help analyse consultation responses but cannot weigh them, five guard rails for AI in engagement, and the government's Consult tool. Through Moffatt v Air Canada it shows that if your AI says it, you said it. It also covers digital exclusion, chatbot essentials, planning for a deepfake of a leader, and Onora O'Neill's idea of well-placed trust.

This capstone puts the programme to work through two illustrative, invented cases: AI note-taking in social work at Cyngor Bro Dyffryn, read through seven lenses, and a repairs chatbot at Cartrefi Glas that failed tenants reporting damp and mould. It draws out the principles that the tool was the easy part and that you automate the routine but never the urgent, then offers a maturity matrix, a ninety-day plan and one loop for leading AI well.