Agenda

• Day 1 Foundations - Making AI Reliable Enough to Use
  • 1pm-4pm GMT
    Session 1: AI in Treasury – Choosing the Right Tool for the Right Problem

    The opening session sets the frame for the course. Rather than treating AI as a single category, the session distinguishes between large language models, Python-based automation, machine learning and agent-style workflows, and where each is genuinely useful in treasury. Live demonstrations across the major assistants show the practical differences that matter when a treasurer is choosing a tool for a real task.

    • Common treasury use cases for LLMs, Python, machine learning and agents
    • Where deterministic logic is still preferable
    • Practical criteria for choosing the right tool for the task
    • How to think about productivity, control and judgement when deciding what to automate

    Session 2: Prompting, Context and Reliable Outputs

    The second session covers the prompting and context discipline needed to get treasury-quality output. The emphasis is on obtaining useful, reliable results on real treasury tasks rather than generic prompting theory. Participants complete a pairs exercise interrogating the model on a topic they know well, then work through the C.A.S.H. framework and build the first version of their own memory document.

    • Writing prompts that reliably produce treasury-usable outputs
    • Using context, examples and policies to lift output quality
    • Building and maintaining a personal memory document
    • Limitations of generative AI and the continuing role of human review

    Assignment – Build your personal treasury AI memory document and try it out on one real prompt from your working week. Bring your chosen real problem for Half-Day 2.
• Day 2 From Prompt to Build - Your First Treasury Workflow
  • 1-4pm GMT
    Session 3: The FX Exposure Workflow – From Raw Exposures to Decisions

    The session opens with a short debrief on the memory-document assignment, then moves into a live build. Participants follow the facilitator through a full FX exposure workflow directly inside their own AI assistant: uploading debtors, creditors and rates, filtering and pivoting the data, calculating net positions, converting to base currency and generating buy or sell recommendations against a policy threshold. Every line of code is generated by the assistant in real time. The build makes the deterministic-and-generative pattern concrete.

    • Uploading and validating treasury data inside the assistant
    • Identifying and structuring FX exposures
    • Applying policy-aware rules and thresholds
    • Producing recommendations that remain subject to human review

    Session 4: Guided Build on Your Own Problem

    Participants take the pattern from Session 3 and apply it to their own problem in their own AI assistant, with facilitator support in breakout rooms. The objective is to leave the half-day with a working, if rough, first version. Two or three participants demonstrate their result in a short closing gallery.

    • Reusing the Session 3 pattern on your own data structure
    • Working out where your problem is deterministic and where it is generative
    • Recognising the failure modes when you move from demo data to real data
    • Leaving with a first-cut prototype

    Assignment – Prepare your Showstopper spec: what task do you really need help with, and what do you and other users need? This will be used in the Week 2 sessions.
• Day 3 Documents, Agents and Building Tools That Decide
  • 1-4pm GMT
    Session 5: Treasury Document Intelligence – From Unstructured Documents to Structured Data

    Document work is one of the strongest applied use cases for treasury AI today. In this session participants work through a build that extracts, structures and classifies information from intercompany and loan-related documents, identifies missing or inconsistent terms, and produces a clean structured output that can be used downstream. The build runs inside the assistant, with participants uploading a set of redacted example documents.

    • Extracting important information from treasury documents
    • Classifying facilities and funding arrangements
    • Identifying missing, inconsistent or unusual terms
    • Producing structured outputs for onward use

    Session 6: Building Your First Treasury Agent – Task, Tools, Guardrails, Escalation

    The session cuts through the marketing around agents by defining an agent simply as a task with a defined set of tools, guardrails and escalation points. A short live demonstration is followed by a guided build in which each participant designs an agent for one step of their own workflow using the agent-authoring surface in their chosen assistant (Custom GPT, Claude Project or Copilot agent). The session closes with the discipline of putting the human-review point at exactly the right place, not one step earlier or later.

    • What an agent actually is, without the marketing
    • Designing an agent for a specific treasury step using your assistant’s native agent surface
    • Guardrails, escalation points and controlled judgement
    • Deciding when to trust and when to hold
• Day 4 Governance, Deployment and the Showstopper
  • 1-4pm GMT
    Session 7: Governance, Controls and Moving from Prototype to Internal Use

    The session translates governance frameworks (NIST AI RMF, ISO/IEC 42001) into language a treasurer can use in front of internal audit, risk and their own board. It covers proportionality (governance that fits the risk of the workflow, not one-size-fits-all), the engine-and-policy separation as a governance tool, and the practical controls that need to be in place before a solution moves from personal use to team-shared use to sanctioned deployment.

    • Governance proportionate to workflow risk
    • The engine and policy separation as a governance tool
    • Logging, change control, thresholds and approval points
    • Moving from a working prototype to repeatable internal use

    Session 8: The Showstopper – Present Your Solution

    The Showstopper brings the Week 2 builds and the Showstopper spec together. Each participant presents their built solution (roughly five minutes each in breakout rooms, then two or three in plenary) and receives peer and facilitator feedback focused on whether it solves a real problem, whether the governance frame is right-sized and what they would build next. The session closes with a practical view of how teams move from ad hoc AI use to a repeatable, governed practice.

    • Presenting a built solution and receiving structured feedback
    • Deciding where your solution lives in the enterprise stack
    • Building a treasury AI operating model that scales beyond one enthusiastic user
    • Continued learning paths and an open door for follow-up questions