Fund Finance AI Session at Project A
On 6 May 2026 I ran an AI session for the fund finance team at Project A. The goal: understand what an AI assistant actually is, and then build a first reusable automation on your own laptop. Thanks to Project A for hosting us, providing the space and the pizza 🍕.
What we covered
- What an AI assistant is: a foundation model (LLM) plus system prompt, interaction interface and orchestration (context window, memory, session state). Examples: Claude (app, Code, Cowork), ChatGPT / Codex, OpenWork.
- What an LLM is and how it is trained: text is split into tokens, the model predicts the next one, repeat. Training means guessing billions of next tokens and correcting the weights. The model predicts text, it does not know or verify facts.
- From LLM to assistant: adding instructions, documents and tools turns a text generator into something that can read, calculate, draft, format and send.
- MCP, skills and tools: skills are predefined instructions or workflows, tools are functions the LLM can call, MCP is the standard interface between the LLM and external tools and data.
- How to build an AI automation: pick a repetitive, time-consuming workflow, break it into small steps with clear inputs and outputs, plan, build, test, iterate. Two worked examples: summarising meeting notes into follow-up emails, and generating pro-rata capital call notices from a commitment schedule.
- Limits: hallucinations (ground answers in real documents, ask for source quotes and page numbers, human review on high-stakes output), data privacy (enterprise tier for client data, redact when prototyping, check internal policy first) and cost (pick the smallest model tier that handles the task, retrieve relevant chunks instead of pasting whole documents).
Hands-on
Two practical blocks:
- Sketch a workflow to automate: each participant identified one repetitive task, defined inputs and outputs, and decided where a human should stay in the loop.
- Build your first skill: in Claude Cowork, chat until the assistant produces the result you want, save it as a skill, run it again and check for consistency, then swap skills with a neighbour.
Material
Further reading
Alexandros Sivris
AI product discovery