Watch the video with chapters & transcript →
AI IN FINANCE
In short: A due diligence review that once meant months of reading hundreds of data-room pages becomes a Claude project with instructions co-written with analysts.
Ani Björkström, a freelance consultant in Stockholm educated in both finance and technology, builds data and AI solutions for asset managers. She recently demonstrated three real client use cases: a due diligence analyst assistant, a sustainability news summarizer, and error detection in the Excel sheets behind daily NAV calculations. None require a developer: the end users are analytical but non-technical.
Due diligence is how an investment company verifies that an investment object — a fund or a company — is what it claims to be. Traditionally an analyst reads hundreds of data-room pages hunting risks and red flags, which can take months.
In the demo, Björkström downloaded the public documents of Handelsbanken Energy, a Swedish fund, and created a Claude project referencing that folder. The critical part is the instructions, never written alone: she sits with analysts and follows a fixed structure — who Claude is, what is needed, how to do it, and rules such as "use only the documents I gave you." The team iterates for a week or two until analysts are satisfied, then asks Claude to create a skill from the conversation so the method can be reused on another fund or company. She is explicit that this is not an AI agent — it is interactive question-and-answer, not a workflow acting on its own.
European asset managers must provide sustainability reporting toward investors, so ESG specialists read the news daily and often present the last 24 hours at a morning meeting. The client wanted that news summarized and sent as an email draft readable on the commute.
This project starts from scratch: Claude gathers everything by browsing. Two settings matter. Permissions: choose "ask before acting," so nothing unplanned happens. Effort levels, introduced with Opus 4.8: high suffices for summarizing news; max suits heavier analytical work if tokens are no concern. With the Gmail connector activated, Claude browsed the sources, produced a draft citing valid references, and scheduled the brief for 9:00 every morning. One caveat: it is a desktop app, so the computer must be on for the schedule to run — hence the "keep awake" option in projects. A scheduled workflow runs on its own, so this case comes closest to a true AI agent.
Yes — and this targets a highly stressful recurring back-office problem. NAV, the net asset value per share of a fund, is calculated daily from holdings, cash and liabilities, then reconciled against the fund manager's figure under hard deadlines. Discrepancies are common, and with a hundred holdings, finding the source manually takes time nobody has.
Using the Claude add-in for Excel — requiring at least the 20-dollar Pro plan — Björkström asked why a synthetic but realistic fund sheet would not reconcile. It identified a sum covering B11 to B12 instead of B11 to B15: rows added without extending the formula, an extremely common failure. Her rule is to verify the diagnosis before letting Claude apply the fix, because in a regulated industry a person, not the tool, is accountable. The add-in also helps newcomers by explaining unfamiliar, badly named spreadsheets in progressively simpler terms.
| Use case | Before AI | With Claude |
|---|---|---|
| Due diligence | Months reading hundreds of data-room pages | Project over the document folder, instructions refined with analysts |
| ESG news brief | Early-morning prep before the meeting | Scheduled 9:00 browse-and-summarize with an emailed, sourced draft |
| NAV reconciliation | Manual hunt through formulas under deadline pressure | Excel add-in locates the broken formula and proposes the fix |
Never use a personal Claude account for this work. Pro and Enterprise plans include a data processing agreement: breaches must be reported, deletion can be requested, and data is not used for training — though with Anthropic directly, data still resides in the US. If it must stay in Europe, Claude's models are reachable via Google Vertex or Amazon Bedrock, but only as an API, so a developer is needed. In practice, banks on the Microsoft stack often default to Copilot, while competitive venture capital and private equity firms sign team or enterprise contracts with Anthropic.
No. An agent needs a workflow of tasks acting on its own; this setup is interactive questioning. The scheduled daily news brief comes closest to a genuine agent.
No. If the machine was off, Claude offers to run the brief on next open, but the intended 24-hour window is lost. The "keep awake" option prevents this.
Not the consultant alone. They come from sitting with domain experts, and Claude itself can improve a draft prompt or even interview the team.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?