Token ≈ 0.75 words. The context window is the AI's "working memory": everything that fits within a single conversation.
128K
ChatGPT ≈200 pages
200K
Claude Opus ≈300 pages
1M
Gemini ≈1,500 pages
1M
Claude API ≈1,500 pages
Important: When the context overflows, the AI starts "forgetting" the beginning of the conversation. Solution: start a new chat and provide the key context again.
Claude Pro for Documents
Your personal analyst
Capabilities
Extract data from PDF and Word files
Create analytical reports
Structure information
Retain context for 200K+ tokens
Work with multiple documents simultaneously
For Wealth Management
Processing bank statements
Analysing portfolio positions
Preparing Due Diligence reports
Creating Portfolio Reviews
Source of Wealth reports
Rule of thumb: Claude works best when you upload a document and provide a clear instruction on exactly what to extract from it.
Types of Documents We Process
What we handle on a daily basis
Bank Statements
UBS, Credit Suisse, CBH, One Swiss Bank. PDFs with positions, transactions, and balances.
Prompt for extracting data from a bank statement
Analyse the attached bank statement and extract:
1. Total Portfolio Value
2. Breakdown by asset class:
- Equities - name, quantity, value
- Bonds - ISIN, coupon, maturity date
- Funds - name, NAV, weight
- Cash - currency, amount
3. Top 5 positions by size
4. Currency allocation (% by currency)
5. Changes over the quarter (if available)
Present the results in a structured table format.
Tip: Upload the PDF directly into Claude. Do not convert it to text — Claude can read PDFs natively.
Julius Pro
A specialised data analytics platform
What It Is
AI-powered analyst for working with data
Upload Excel/CSV — receive analysis
Agentic approach: writes Python code
SQL queries on your tables
Visualisation: charts and diagrams
Advantages
Handles spreadsheets up to 32 GB
No hallucinations — precise calculations
Works with formulas, not text
Export results to Excel
Data visualisation
Key differentiator: Julius does not "make up" numbers. It writes Python code that runs on your actual data. The result is always accurate.
Julius Pro vs ChatGPT
When to use which tool
Criterion
Julius Pro
ChatGPT / Claude
Data size
Up to 32 GB
Limited by context window
Calculation accuracy
100% (executes code)
Errors possible
Visualisation
Built-in charts
Limited
Text analysis
Weak
Excellent
Report generation
Data only
Full-fledged documents
Cost
$20/month
$20/month
Takeaway: Use Julius for precise calculations and large spreadsheets. Use Claude for document analysis and report generation. Together they cover 90% of tasks.
What Is an AI Agent
Theory: from chatbot to autonomous system
AI agent — a system that does not merely generate text but can plan actions, use tools, and execute tasks step by step. Unlike a standard chatbot, an agent can decide on its own which tool to invoke.
Standard Chatbot
Question → answer
Text generation only
Cannot take actions
Does not self-verify
Example: "What is the USD exchange rate?" → text response (may be outdated)
Example: One agent parses the PDF, another runs calculations, a third writes the report
How Our Tools Work
What happens under the hood
Claude Pro
Large Language Model (LLM) + document parsing
You: upload a PDF
↓ Claude converts PDF to text (OCR/parsing)
↓ LLM analyses text according to your prompt
↓ Generates a structured response
= Text-based analytics, but does not execute code
Strength: understanding context, language, nuances. Weakness: may make calculation errors.
Julius Pro
LLM + Python interpreter (agentic approach)
You: upload Excel + question
↓ LLM writes Python code (pandas, numpy)
↓ Code executes on your actual data
↓ Result is returned to you
= Precise calculations, as the machine computes, not the LLM
Strength: 100% calculation accuracy. Weakness: does not understand document context.
Perplexity Deep Research
LLM + web search + research agent
You: ask a question
↓ Agent plans search queries
↓ Searches the internet (10-30 sources)
↓ LLM synthesises the information
= Up-to-date data with sources
Strength: timeliness, source citations. Weakness: does not work with your files.
DeepL
Neural Machine Translation (NMT)
You: upload text/document
↓ Encoder analyses the source language
↓ Decoder generates the target language
↓ Terminology post-processing
= High-quality literal translation
Strength: terminology accuracy, speed. Weakness: does not understand task context.
The Agentic Loop
How an AI agent solves tasks step by step
The "Think → Act → Observe" Cycle
1
Think
The agent analyses the task and decides which tool to use
2
Act
Invokes a tool: code execution, search, API call, file reading
3
Observe
Reviews the result: did it work? Are there errors?
4
Repeat
If not ready — repeats the cycle. If done — delivers the result.
Example: Julius Calculates Returns
Thinks: "I need to load the data and calculate returns"
Acts: Writes code: df.pct_change()
Observes: "There are NaN values, need to clean them"
Acts again:df.dropna()
Result: Accurate returns table
Example: Perplexity Checks Sanctions
Thinks: "I need to check the name against sanctions lists"
Acts: Searches: "Name OFAC SDN list"
Observes: "No exact match, but a similar name found"
Acts again: Refines by date of birth and country
Result: "Clear — namesake from a different country"
Why does this matter? Understanding how an agent works allows you to assign better tasks. An agent is not a magician but an executor with a set of tools. The more precise the task, the better the result.
Portfolio Reviews: The Problem
40+ hours per week of manual work
40+
hours/week on Portfolio Reviews
6+
banks different formats
4-6
hours per single review
Current Pain Points
Data is fragmented (multiple brokers)
Unstructured PDF statements
Errors in broker statements
Manual reconciliation takes hours
Every bank uses a different format
Scale of the Problem
Each manager: 10-20 hours/week
Team of 3-4 people = 40+ hours
+ time for reconciliation and corrections
+ report formatting
= half of working time spent on routine tasks
Automating Portfolio Reviews
4 steps from statement to report
1
Extraction
Upload bank statements. Extract positions, balances, and transactions.
2
Reconciliation
Automated valuation checks. Identify discrepancies and errors.
3
Structuring
Populate Excel. Allocate by asset class, currency, and region.
4
Report
Generate the presentation. Charts, tables, and recommendations.
Result: A manual Portfolio Review takes 4-6 hours. With AI, it takes 1-2 hours. A time saving of 60-70%.
Portfolio Review Prompt
Copy and use
Prompt for Portfolio Review
You are an analyst in the Wealth Management department
of FP Wealth Solutions.
Analyse the attached bank statements and prepare
a draft Portfolio Review:
1. PORTFOLIO SUMMARY:
- Total Assets Under Management (AUM) in USD
- Breakdown by custodian/bank
- Currency allocation (%)
2. ASSET CLASS ALLOCATION:
- Equities / Bonds / Funds / Cash / Alternatives
- Current vs target allocation
- Deviations from mandate
3. PERFORMANCE:
- Quarterly performance (%)
- Benchmark comparison
- Top 3 best and worst performers
4. RECOMMENDATIONS:
- Rebalancing (if required)
- Current portfolio risks
- Optimisation opportunities
Format: tables + brief commentary.
Weekly Cash Balance Reconciliation
Automating a routine process
Before (Manual)
Request statements from each bank
Manual data entry into Excel
Reconciliation with the previous week
Searching for discrepancies
Updating client portfolios
Time: 3-4 hours every week
After (With AI)
Upload statements to Claude/Julius
Automatic balance extraction
Auto-reconciliation with the previous week
Anomaly highlighting
Ready report in minutes
Time: 30-60 minutes
Document Translation
DeepL vs Claude: when to use which
DeepL (Literal Translation)
Legal documents
Technical specifications
Compliance documents
Investment agreements
Principle: Word for word. Preserves terminology accuracy.
Claude (Contextual Translation)
Marketing materials
Client correspondence
Presentations
Books and articles
Principle: Considers context, style, and cultural nuances.
The Right Translation Prompt
Poor vs Effective prompt
Poor Prompt
Translate this document into English.
Result: a generic translation without accounting for specifics
Effective Prompt
Translate this document from Arabic into English.
Context: legal document for banking KYC.
Requirements:
- Literal translation of terms
- Preserve formatting
- Transliterate proper names
- Dates in DD/MM/YYYY format
Perplexity Deep Research
Research with real-time data from the internet
What Perplexity Can Do
Real-time search for up-to-date information
Sanctions screening & Adverse Media
Verification of legal entities and beneficial owners
Market and sector overviews
Links to primary sources
When to Use
Due Diligence: client verification before account opening
Portfolio Review: current market data
Compliance: sanctions lists, PEP screening
Research: company and competitor analysis
Prompt for Perplexity Deep Research
Conduct deep research on the following client for Due Diligence:
Name: [CLIENT NAME]
Country: [COUNTRY]
Company: [COMPANY NAME]
Find and analyse:
1. SANCTIONS: check against OFAC, EU, UN sanctions lists
2. PEP STATUS: politically exposed person?
3. ADVERSE MEDIA: negative mentions in the press
4. CORPORATE RECORDS: company registrations, beneficial owners
5. LITIGATION: court cases and proceedings
Provide sources (links) for each item.
Language: English.
Perplexity Search Use Cases
6 query types for Wealth Management
1. Sanctions Screening
Check if [NAME] from [COUNTRY] appears on OFAC SDN, EU consolidated sanctions list, or UN Security Council sanctions. Include PEP status check. Provide sources.
2. Adverse Media Check
Search for negative news, investigations, lawsuits, or regulatory actions involving [NAME / COMPANY] in the last 5 years. Focus on financial crime, fraud, money laundering. Provide links to each source.
3. Market / Sector Overview
Provide a current market overview of [SECTOR, e.g. US tech stocks / European bonds / EM equities]. Include: YTD performance, key drivers, outlook for next quarter, main risks. Sources required.
4. Company Verification
Research [COMPANY NAME]: registration details, beneficial owners, financial performance, recent news, any regulatory issues. Country: [COUNTRY]. Provide corporate registry links where available.
5. Fund / ETF Analysis
Analyze [FUND/ETF NAME]: AUM, expense ratio, top holdings, 1Y/3Y/5Y performance, risk metrics (Sharpe, max drawdown), comparison with peers. Morningstar rating if available.
6. Regulatory Changes
What are the latest regulatory changes in [JURISDICTION, e.g. Switzerland / UAE / EU] affecting wealth management, KYC/AML requirements, or cross-border investments? Focus on 2024-2025 updates.
Manus AI
Autonomous AI agent for complex tasks
What It Is
Fully autonomous AI agent
Opens a browser and searches for information on its own
Creates files: Excel, Word, PDF, presentations
Works in parallel (multiple tasks)
You assign the task — it handles everything
For Wealth Management
Deep Research: comprehensive report on a company/client
Market Reviews: data collection from 20+ sources
Comparative Analysis: funds, ETFs, strategies
Compliance: regulatory requirements verification
Reports: ready-made documents with charts
Prompt for Manus
Conduct a comprehensive analysis of [COMPANY NAME] for Due Diligence:
1. Find registration data, beneficial owners, financial statements
2. Check against sanctions lists (OFAC, EU, UN)
3. Find media mentions over the last 3 years
4. Compare with sector competitors
5. Present results as an Excel spreadsheet + brief report in Word
Language: English. All sources must be cited.
Difference from Perplexity: Perplexity provides a text response with links. Manus creates ready-to-use files (Excel, Word, PDF) and can execute multi-step tasks autonomously.
Gamma
AI-powered presentation and document generation
Capabilities
Create presentations from text in minutes
Professional design generated automatically
Chart and diagram generation
Export to PDF and PowerPoint
Web pages and landing pages
For Wealth Management
Portfolio Review: polished presentation for the client
Investment Proposal: professional formatting of investment offers
Quarterly Reviews: market reports with charts
Onboarding: materials for new clients
Internal Reports: for senior management
Prompt for Gamma
Create a Portfolio Review presentation for a client:
Portfolio data:
- Total AUM: $5.2M
- Allocation: 45% equities, 30% bonds, 15% funds, 10% cash
- Quarterly return: +3.8%
- Benchmark: +2.1%
Include slides:
1. Portfolio overview (key figures)
2. Asset class allocation (pie chart)
3. Performance vs benchmark (bar chart)
4. Top 5 positions
5. Recommendations
Style: professional, minimalist. Colours: dark blue, gold.
Claude in Excel
AI assistant directly in your spreadsheets (Add-in, 2026)
Capabilities
Reads the entire workbook (all sheets, formulas, references)
Creates and edits formulas while preserving existing ones
Builds pivot tables and charts
Conditional formatting on demand
MCP connectors: S&P Global, LSEG, FactSet
Keyboard shortcut: Ctrl + Alt + C — invoke Claude in any cell
For Wealth Management
Portfolio: automatic allocation and return calculations
Reconciliation: comparing statements from different banks
Example query in Claude Excel
Analyse the data on the "Portfolio" sheet:
1. Calculate each asset's share of Total AUM
2. Add a "Quarterly Return" column with formulas
3. Create a pivot table by asset class
4. Build a pie chart for the allocation
5. Add conditional formatting: green >5%, red <0%
How to connect: Microsoft 365 → Get Add-ins → Claude for Excel. Requires a Claude Pro subscription ($20/month) or Team/Enterprise.
Claude in PowerPoint
Native editable slides (launching February 2026)
Capabilities
Creates real PowerPoint slides (not images!)
Editable text, shapes, SmartArt
Respects existing design: layouts, fonts, colours
Can extend an existing presentation
Works via the Office JavaScript API
For Wealth Management
Portfolio Review: auto-generate slides from Excel data
Quarterly Report: update an existing template
Investment Proposal: structured presentation
Onboarding Pack: personalised materials
Board Report: executive summary for management
Example query in Claude PowerPoint
Create 5 slides for a Quarterly Portfolio Review:
1. Title slide: "Q4 2025 Portfolio Review" + client name
2. Portfolio overview: AUM, return, benchmark
3. Asset class allocation (use data from Excel)
4. Top 5 positions with quarterly returns
5. Recommendations for the next quarter
Style: match the current presentation template.
Gamma vs Claude PPTX: Gamma generates from scratch with beautiful design. Claude PPTX works within your company's existing templates.
How to connect: Microsoft 365 → Get Add-ins → Claude for PowerPoint. Available for Claude Pro/Max/Team/Enterprise.
Case Study 1: Bank Statement Analysis
Swiss banks and UAE brokers
Objective
Extract position, transaction, and balance data from statements of different banks (UBS, Credit Suisse, CBH, One Swiss) and UAE brokers. Consolidate into a single table.
Challenges
Every bank uses a different PDF format
Some PDFs are poorly structured
Errors in UAE broker statements
Different currencies and naming conventions
AI-Powered Solution
Claude: data extraction from PDFs
Julius: precise calculations and aggregation
Prompt specifying the output format
Mandatory manual verification of key figures
Case Study 2: Automating Due Diligence
Source of Wealth reports
Process
1
Questionnaire
Client completes a form: income, assets, business interests.
2
Extraction
AI extracts and structures the data.
3
Calculations
Currency conversion, Net Worth calculation.
4
Report
Word + Excel + Sanctions Screening.
Result: A Due Diligence report that previously took 6-8 hours is now prepared in 1.5-2 hours with AI. Quality improves through standardisation.
Source of Wealth Prompt
Key sections
Source of Wealth Report Structure
You are a compliance specialist at FP Wealth Solutions.
Create a Source of Wealth report based on the questionnaire.
USE ONLY data from the attached document.
REPORT STRUCTURE:
1. Personal Information (full name, passport, addresses)
2. Family (family members)
3. Education (educational background)
4. Career (career history with periods and income)
5. Business Participation (shareholdings in companies)
6. Annual Income (income table by year in USD)
7. Net Worth (total wealth)
8. Real Estate (property)
9. Bankable Assets (accounts and brokers)
10. Sanctions & Adverse Media Screening
Currencies: convert to USD at the average annual rate.
Language: English only.
Day 2 Tools
8 AI tools for Wealth Management
Claude Pro
Document analysis, report creation, PDF processing. 200K token context.
Julius Pro
Precise data analytics in Excel. Agentic approach, no hallucinations.
DeepL
Literal translation of legal and technical documents.
Perplexity
Research with up-to-date data. Sanctions screening, fact-checking.
Translation of the final report for the client (RU → EN or vice versa)
Workflow: Due Diligence
Perplexity: Sanctions & adverse media screening
Claude: Data extraction from questionnaire, SoW generation
Julius: Net Worth calculation, currency conversion
DeepL: Translation of documents from Arabic / Russian
Workflow: Weekly Reconciliation
Claude: Balance extraction from new statements
Julius: Reconciliation with the previous week, anomaly detection
Perplexity: Exchange rate and market price verification
Claude: Final report generation for the team
Principle: Each tool does what it does best. Perplexity — real-time data. Claude — document analysis. Julius — precise calculations. DeepL — translation.
AI Hallucinations in Finance
What can go wrong and how to protect yourself
15-20%
of GPT-4 responses contain factual errors
5-10%
error rate in data extraction from PDFs
<1%
error rate with Julius (code + verification)
Common Hallucinations
Fabricated account details and account numbers
Incorrect exchange rates and dates
Non-existent sanctions records
Mixed-up positions from different tables
"Confident" answers without supporting data
How to Protect Yourself
Always verify figures against the source PDF
Use the prompt "USE ONLY data from the document"
For calculations — use Julius (code, not guessing)
Perplexity — for fact-checking (with source links)
Rule: AI generates the draft, humans verify
Key Takeaways
12 main lessons from Day 2
AI agents are not chatbots: They plan, use tools, and self-verify
Julius Pro for precise calculations: Writes Python code, executes on real data, no hallucinations
Context window is critical: Claude with 200K tokens is better suited for Portfolio Reviews than ChatGPT
Perplexity for fact-checking: Sanctions, adverse media, corporate data. Always with source links
Manus is an autonomous agent: Comprehensive research with ready-made files (Excel, Word, PDF). Delegate complex tasks
Gamma for presentations in minutes: Portfolio Review, Investment Proposal, quarterly reviews with professional design
Claude in Excel: AI directly in spreadsheets — formulas, pivot tables, charts, MCP connectors to financial data (S&P, LSEG, FactSet)
Claude in PowerPoint: Native slides within existing templates. Gamma for creating from scratch, Claude PPTX for working within templates
Effective prompting for translation: Legal documents — DeepL; marketing — Claude
Hallucinations are the primary risk: 15-20% of responses contain errors. "USE ONLY data from the document"
Combine your tools: Perplexity → Claude → Julius → DeepL → Gamma. Each one does its part
AI = draft, human = verification: Especially for financial data and client reports
Homework
2-3 hours of practice before the next session
1. Set Up an Assistant
Configure Claude Pro for working with your documents. Upload a sample statement and try extracting data from it.
2. Automate a Process
Choose one routine process: data extraction from a PDF, populating spreadsheets, or creating a report.
3. Try Julius Pro
Upload an Excel file with data into Julius and ask 3 analytical questions. Compare the results with manual calculations.
4. Prepare Documents
Gather sample documents for the next session: statements, spreadsheets, reports.
Next session: December 20, 2025 — Vibecoding and Claude Code