Watch the video with chapters & transcript →
AI IN FINANCE
In short: Three AI agents built for actual finance clients — a due diligence analyst, a scheduled news brief, and an Excel bug-finder — run in real enterprise environments.
Most AI-agent tutorials on YouTube are built on private computers and would never survive a real enterprise environment. Here, Ani Björkström, a tech consultant based in Stockholm, shares three finance agents actually delivered to clients, including the prompts behind them.
A due diligence agent reads fund documents the way a senior analyst would and answers investor questions using only those documents. The first demo builds one for a Sweden-registered equity fund classified as Article 9 under SFDR, with a prompt structured in four parts: who you are, what I need, how to do it, and rules and constraints.
The agent lives in a Claude project rather than a chat. A project has a larger context window and persists, so users can return the next day without losing anything; once the answers satisfy the team, the knowledge is saved as a skill and reused on other funds.
Create a project with a four-part prompt, connect the Gmail connector, and schedule the run for every weekday at 7:00. The client had to present 24 hours of sustainability news at a 9:00 daily brief, which forced her into the office at 5 or 6 in the morning; the agent now finds, summarizes, and drafts the news for her commute.
Precision comes from iteration with the actual user, not from the first prompt. The brief was scoped to the Nordic market at the client's request, and instructions were updated after each review round until the agent answered exactly the way she wanted.
Yes — Claude in Excel located the source of a NAV discrepancy between the front-office model and the fund administrator's figure. NAV is calculated daily by both front office and back office, and because it is published to investors, any per-share discrepancy must be resolved fast.
Claude reads every formula in every tab, so one prompt asking it to trace the model found that total liabilities summed only rows B11 to B13 while rows 14 and 15 had been added without updating the formula. After a manual approval, it corrected the sum to B11:B15, leaving a difference of 0.00001 against a tolerance of 0.01.
| Agent | Setup | Output |
|---|---|---|
| Due diligence analyst | Claude project referencing a local folder of fund documents | Answers on objectives, risk category, and annual costs from the documents only |
| Morning brief | Project with web search, Gmail connector, weekday 7:00 schedule | Top five Nordic sustainability news items as a Gmail draft with sources |
| Excel bug-finder | Claude opened inside the Excel workbook | Root cause of a NAV discrepancy plus the corrected formula |
A cheaper model is often enough — extracting data from given documents is not the most advanced thinking task, so a lower-tier model performs well at lower cost.
Start with manual approval, where Claude pauses before each action, and shift to automatic approval once you are comfortable with how the agent behaves.
No. The demo runs on Microsoft 365 in the browser, and the same setup works on Excel running locally on your computer.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?