Back to Archive
S1E01Season 1Top Session

ANCHOR INTELLIGENCE SAMURAI

Training for Precision in the Age of AI

October 2025
90 min

OVERVIEW

The foundational session. Covers key AI concepts and terminology, hallucination awareness, choosing the right AI model, privacy considerations, and a complete client proposal automation workflow using AI agents.

KEY TAKEAWAYS

Get the paid version of whichever AI model you use. Anchor will accept claims cleared by your manager

AI is a tool AND a partner. A genius-level assistant that never sleeps with access to all human knowledge, but it has quirks

Don't let hallucination risk prevent you from using these tools. New models reduce hallucinations by 40-50%

Cross-check important outputs by running them through a second AI tool for verification

Use the Snipping Tool (Win+Shift+S) to show AI exactly what you see instead of describing it

Start with whichever tool you have access to. Copilot is your enterprise default, but explore others for different tasks

AI agents are like team members who carry out plans, not just consultants you ask for advice

PRO TIPS

Pro Tip #1: Say Hi to Your AI

Start voice conversations to capture thoughts, role-play scenarios, get personalized news, brainstorm ideas, practice negotiations, draft emails, break down complex topics, and more. Tap the voice icon in the app to start.

Pro Tip #2: Reducing Hallucination Risk

When working with complex queries, use a second or third LLM to verify data and cross-check sources.

  1. 1.Create your research report in your primary AI tool
  2. 2.Download the complete output
  3. 3.Upload the report to a different LLM (e.g., Claude or Gemini)
  4. 4.Prompt the second LLM to review accuracy and identify incorrect or misleading statements

Pro Tip #3: Use Snipping Tool. Show, Don't Tell

Instead of describing a complex UI, chart, or error message in words, just snip it. AI sees exactly what you see. Use Windows Shortcut: Win + Shift + S

Pro Tip #4: Use Voice Dictation

An easy way to generate longer prompts is to use dictation. Speak naturally and let the AI convert your voice to text in real-time.

  1. 1.Activate dictation by clicking the microphone icon or keyboard shortcut
  2. 2.Start speaking your prompt naturally
  3. 3.Review and submit. Your prompt appears in text format, ready to send

Pro Tip #5: Structure Your Prompts for Clarity

Visual structure matters. Use line breaks (Shift+Enter), bold text (**text**), and dash lists (-) to organize prompts into logical chunks. Models understand context better when content is visually structured.

VISUAL REFERENCE

HALLUCINATION RATES BY MODEL

Percentage of factually incorrect outputs — lower is better

GPT-4oOpenAI o3GPT-5GPT-5 (think)0%5%10%15%20%

Source: Health Bench Hard Hallucinations Study

PRIVACY & SECURITY RANKING

Data safety assessment for client-facing work — capability is separate from privacy

Enterprise-Grade
Copilot (Enterprise)ChatGPT Team/EnterpriseClaude Pro

Enterprise licensing with data privacy guarantees — your data is not used for training

Personal Paid Tiers
ChatGPT PlusClaude ProGemini Advanced

Stronger privacy than free tiers but always check your organisation's policy before using client data

Not Recommended for Client Data
Any free tierUnverified AI tools

Data may be used for model training — never input client-sensitive information

Source: Incogni AI LLM Privacy Ranking 2025

CROSS-CHECK: VERIFY BEFORE YOU TRUST

Never rely on a single AI output for important work — here's the process

Step 1: Generate in your primary AI
Write a market analysis for SA equities Q4 2025

SA Equities Q4 2025 Analysis

The JSE All Share Index gained 4.2% in Q4...

Step 2: Verify in a different AI

Fact-check prompt:

Review this market analysis. Flag any claims that are factually incorrect, outdated, or unsupported. List each issue with the correct information.

Issues Found:

JSE gain was 3.8%, not 4.2%
Sector breakdown is accurate
Golden Rule:

The AI that wrote the report should never be the same AI that fact-checks it. Use a different tool or a fresh conversation.

ANATOMY OF A GREAT PROMPT

Five building blocks for structured, effective prompts

Role"You are a financial analyst at Anchor Capital"
Task"Produce a concise, data-driven market summary"
Context"Your audience: portfolio managers who value clarity"
Constraints"Avoid generic commentary. Support with data."
Format"Use bullet points, keep under 500 words"

VISUAL STRUCTURE MATTERS

The same information, structured differently, produces dramatically different results

✗ Wall of text

Can you write me a report about the South African equity market performance for Q4 2025 and include the top performing sectors and also mention any risks and make sure to compare it with the previous quarter and format it nicely and keep it professional and add some charts if possible and make sure the data is accurate

Result:Generic, unfocused, misses key details
✓ Structured prompt

Write a Q4 2025 SA equity market report.

Sections to include:

- Overall JSE performance vs Q3

- Top 3 performing sectors with % gains

- Key risks for Q1 2026

Format: Executive summary (3 paragraphs), then data table

Tone: Professional, suitable for client-facing report

Length: 400-500 words

Result:Focused, structured, client-ready output

SAMURAI PRO TIPS

Quick-reference techniques from the training sessions

Say Hi (Voice)
Snipping Tool
Cross-Check LLMs
Voice Dictation
Visual Structure
Show, Don't Tell

SNIPPING TOOL: SHOW, DON'T TELL

Screenshot anything and paste it directly into your AI conversation

Win+
Shift+
S
SnipSelect any area on screen
PasteCtrl+V into your AI chat
Ask"What does this show?"

VOICE: JUST SAY HI

Speak naturally to your AI — it's faster than typing for complex requests

Voice Dictation
Win+
H
Compare the performance of Anchor BCI Equity fund with|
Listening...

Converts speech to text in any input field — works everywhere, not just AI tools

Voice Conversation
0:04
"Summarise the key risks in this portfolio report"
Based on the portfolio report, here are the 3 key risks...

Click the microphone icon in ChatGPT, Copilot, or Claude to start a voice conversation

CLIENT PROPOSAL PIPELINE

End-to-end AI-assisted workflow — from raw client email to branded proposal in under 30 minutes

STEP 1

Receive Email

Download the client email as a .eml file or print it to PDF. Most email clients let you do this via File → Save As or Print → Save as PDF. LLMs work best with PDF — if you have any other format, convert it first.

STEP 2

Extract Actions

Upload the PDF to your AI tool and prompt: 'Summarise this email. Extract what the client is asking for, specific action points, deadlines, constraints, and exclusions. Prioritise by urgency.' You now have a structured brief in 30 seconds.

STEP 3

Research Funds

Use Agent Mode (available in ChatGPT, Claude, and Copilot) to browse Anchor's fund range and compare against competitor funds the client may hold. Ask the AI to pull performance data, fee structures, and risk profiles. The key here: you're comparing other investment managers' funds against Anchor's funds to build a compelling case.

STEP 4

Compare Options

Ask the AI to create a side-by-side comparison table: Anchor fund vs competitor fund, across performance (1yr, 3yr, 5yr), fees, risk rating, and mandate alignment. Upload fund factsheets as PDFs for the most accurate comparison. The AI will structure this into a client-ready format.

STEP 5

Check Exclusions

Cross-reference the recommended fund holdings against the client's stated exclusions — ESG restrictions, sector limitations, geographic constraints. Upload the fund's holding list and the client's mandate. Prompt: 'Flag any holdings that conflict with these exclusions.' AI catches what manual review might miss.

STEP 6

Generate Proposal

Combine everything — the email summary, fund comparison, exclusion check, and your recommendations — into a structured proposal. Prompt: 'Using all the above, draft a client proposal that leads with the recommendation, supports it with data, and addresses each of the client's original questions.'

STEP 7

Apply Brand & Tone

Feed in Anchor's investment philosophy and your preferred communication tone. Use the Snipping Tool to screenshot a branded template as a visual reference. Prompt: 'Rewrite this proposal in a professional, warm tone consistent with Anchor Capital's brand voice. Keep it concise but thorough.' The AI adapts the output to sound like you, not a robot.

STEP 8

Review & Send

This is the human-in-the-loop step. Read every line. Verify the numbers. Check the fund names. Cross-reference against source documents. AI handles the grunt work — you handle the judgement. What used to take 2-3 hours now takes 20-30 minutes.

SESSION CONTENT

KEY CONCEPTS & TERMINOLOGY

text
LLM (Large Language Model): The underlying brain that predicts and generates text. Examples: Copilot, Claude, Gemini. Tokens: AI currency / a measure of usage. 1M tokens is roughly 750K words. More about limitation than cost. Prompt Engineering: Detailed instructions using highly specific text. Important for now, but don't sweat it. Agents: Autonomously achieve pre-defined goals. Instead of asking how, we say do. Example: 'book flights to CPT' or 'send report daily'. Model Variants: For example GPT-5 Instant vs Pro. Speed vs deep reasoning. Bias: AI is a product designed to be used. It can distort info to please, but it's fixable. Hallucinations: AI sometimes makes stuff up and we don't really know why. The error rate is reducing. AGI: Better at almost all tasks than humans. Can advance itself. Nirvana / apocalypse / probably neither.

TERMINOLOGY MIX-UPS

comparison
LLM: The underlying brain. Predicts and generates text based on patterns. Generative AI: The broader field. Includes models that create text, images, code, video. Chatbot: A user interface. How you talk to the model. The front-end, not the intelligence. AI Assistant: A more capable chatbot. Can reason, search, analyse, or act across tools. Tools like Copilot, Claude, Gemini and Grok are all these things at once. A Generative AI system built on an LLM, used through an AI Assistant interface.

HALLUCINATION RATES ACROSS MODELS

text
GPT-4o: ~15-17% hallucination rate OpenAI o3: ~10-12% GPT-5 (without thinking): ~5-7% GPT-5 (with thinking): ~1-3% New generation models greatly reduce hallucinations. Like the 40-50% improvement between GPT-3.5 and GPT-5. However, it's valid to be cautious and verify outputs, especially for complex or critical work. Source: Health Bench Hard Hallucinations study

PRIVACY & SECURITY ASSESSMENT

warning
Enterprise-approved: Microsoft Copilot Okay to use (paid): Claude, Gemini Use with Caution: Grok, free tiers of any model Not Recommended: Free tiers of most models Ranking is not reflective of the model's capability, just data privacy and safety. Organizations must navigate privacy compliance without clear frameworks. FICA, POPI, and most regulatory bodies lack specific AI guidance. Anchor AI Policy is in drafting phase. Source: Incogni AI LLM Privacy Ranking 2025

USE CASE: CLIENT PROPOSAL AUTOMATION

workflow
Scenario: Client Wes Borland wants to move R19M from Ninety-One to an Anchor fund with equivalent risk mandate. Also asks about Ferrari share price. Excludes Novo Nordisk. Step 1: Summarize the email and extract action points Step 2: Address the Ferrari query using closed data (WhatsApp notes, sell-side reports) Step 3: Use AI Agent Mode to search Anchor Capital website for matching funds Step 4: Compare fund performance, fees, and risk profiles Step 5: Check fund holdings for excluded stocks (Novo Nordisk) Step 6-9: Generate branded proposal with correct tone and investment philosophy Result: AI identified Anchor BCI Managed Fund and Anchor BCI Worldwide Flexible Fund as matches.
Anchor Intelligence SamurAI

TOPICS COVERED

Key Concepts & Terminology
Hallucinations
Choosing the Right AI
Privacy & Security
Client Proposal Automation
AI Agents
Prompt Structure

CONTENT QUALITY

Essential — highly actionable content