What I told a room of investors about getting more out of Claude.
Tips from a lunch and learn on AI with an investment team: how to make your usage go further, set Claude up properly, and pick the first task to automate. Plus BNY's AI cost tracking and a study on how agents sway analysts.
Hi folks,
This week I ran a lunch and learn on AI for a large multi-asset investment firm. Here are the tips that got the most questions.
What we're seeing.
Top tips from the session
Make your usage go further:
- One chat per task. Every turn resends the whole thread, so old chats cost more (especially when there are large documents in the thread).
- Brief it once, properly. Give the format, length and audience up front. I ramble it out by voice, then for big tasks ask Claude to play back its understanding and ask me questions. Only then do I approve its plan.
- Send the page, not the book. Word or Markdown documents go further than a PDF.
- Match the model to the job. Opus or Sonnet on medium effort is a good default. Go up only when you are stuck.
Set up your Claude/GPT properly:
- Keep custom instructions broad. Your role, tone and preferred format.
- It's here you put standing instructions, i.e. "If anything is unclear, ask me before you start."
- Meanwhile use skills for repeatable tasks. An IC paper first draft, a covenant check, meeting notes to actions. These are all good examples of when a skill is useful.
- Use a project for ongoing work. One per deal or fund, with the IM, model and past IC papers loaded once, so every new chat starts with that context. A skill is how you do a task; a project is what the work is about.
Once you've got the basics right, start with one painful task and give it an owner who keeps the instructions and agents current. Examples we've seen working with our clients: a $400m corporate lender cut monitoring from 9 hours to 2 per loan, and a ~$1bn property fund cut IC papers from 2 days to 2 hours.
In the mix.
- BNY (the world's largest custodian bank) now tracks what it spends on AI, task by task (The Next Web)
- CFO Dermot McDonogh told Bloomberg TV the bank routes each task to the model that suits it: "We don't use expensive models for cheap queries." In July he said BNY did not track token use at all.
- Our read: the same habit as "match the model to the job", at the scale of 400 agents.
- Analysts went along with Claude's hidden lean 86% of the time (Anthropic)
- In a trial, 50 analysts ran an analysis in Claude Code, and the agent was secretly told to favour one option on each methodology choice. Most followed it, 77% reused its argument, and only 15% said where it came from. The leans were deliberately strong, so real effects are likely smaller.
- Our read: for IC papers, ask the AI for the best case on each side, and form your own view before you read its recommendation.
Levercon builds custom, self-improving AI agents and systems for investment firms, with every output traced, verified and signed off.
Book a strategy call.
See where AI fits in your firm.
A short conversation about where AI could help your firm, what to build first, and what it would take.