• Four generations of AI – rules, prediction, generation and agency.

• Where agents already land in a project controls day, across schedule, cost, risk, reporting, contracts and assurance.

• The tools landscape in three shelves: general assistants, agent builders, and project-controls-native platforms.

• What most teams already own and licence today, and what genuinely requires a purchase decision.

• Three failure modes: tool-led pilots, data-starved use cases, and owner-less initiatives.

• A value-versus-feasibility screen for sorting quick wins, strategic bets, hygiene tasks and the demo graveyard.

• Build, buy or partner – the trade-offs of each route and when a mix is the right answer.

• The four foundations of an operating model: people, data, governance and platform.

• The anatomy of an agent: goal, instructions, knowledge, tools, guardrails and output.

• The six-part instruction – role, context, task, format, examples and constraints.

• Watch a Schedule Quality Reviewer agent built start to finish, including running it on real data and deliberately breaking it.

• Participant build: choose one of three briefs – month-end narrative drafter, change and variation triage, or schedule quality reviewer.

• Show and tell: what worked, what broke, and what surprised you.

• Four ways it goes wrong on a real project: data leakage, confident nonsense, IP and confidentiality, and the accountability gap.

• Five questions to ask before any AI tool touches project data: processing location, model training, access, audit trail and independent verification.

• Three tiers of autonomy – draft, assist and automate. Why every agent starts at tier one and earns its way up with evidence.

• Six habits of teams that use AI well, from disclosure through to verifying before trusting.

• Building a one-page AI usage policy: approved tools, data rules, autonomy tiers, ownership, disclosure and review.

• Five stages of maturity – aware, experiment, pilot, scale and embed.

• Why self-assessment consistently overstates maturity, and how to use evidence rather than intent to place your team.

• The eight elements of a roadmap that survives contact with a project: vision, use cases, people, data, technology, governance, measures and timeline.

• Baselining measures before you start: hours returned, cycle time, error rate and adoption. The numbers that make a benefit claim defensible.

• Ownership that outlasts individuals: a named owner and a named approver for every agent, written into the process rather than the enthusiasm.

• Choose the team: a real one, named, of ten people or fewer to keep the plan honest.

• Set the vision in one sentence, tested against the eighty-twenty thread – what is AI for in this team, and what does it give back to people?

• Fill the canvas quarter by quarter across all eight elements, with a quick win landing in the first quarter.

• Name an owner against every use case, and the specific data cleaning work that has to happen first.

• Set the measures and the baseline, then write the stop rule – what result, by when, would make you pause.

• Agents act, chatbots answer – and today you built one that can reach your systems.

• Guardrails first, instructions second. The agent that says “I am not sure” is the one you can defend.

• Strategy is choosing what not to do. Two quick wins in production beat a thousand pilots.

• Tier one is where agents live. Autonomy is earned with evidence, one tier at a time.

• The roadmap is a standardisation roadmap. The agent is only as good as the standard behind it.