RLHF — Reinforcement Learning from Human Feedback: humans rank answers, the model learns what's preferred
RL — Reinforcement Learning: learning by reward/feedback instead of fixed examples
Constitutional AI — the model critiques & revises its own answers against written principles (Anthropic's method for Claude)
What they do really well ✅
Summarizing long text
Drafting emails and documents
Translation and rephrasing
Extracting structure from chaos
Explaining things in plain words
Generating and explaining code
Ideas and brainstorming
Classification and sorting
Where they fail ⚠️
Hallucinations
They sometimes invent facts — confidently.
No access to your data unless you connect it (that's tomorrow's work!).
Weak at exact math and counting.
Their knowledge has a cut-off date.
So the human stays in control — AI gives a draft, not a verdict.
The 2026 landscape — who's who
Provider
Region
Models
Notable for
Microsoft + OpenAI
🇺🇸 US
GPT‑5
In M365 Copilot & Copilot Studio
Anthropic
🇺🇸 US
Claude (Opus, Sonnet 4.5/4.6)
Also selectable in Copilot Studio
Google
🇺🇸 US
Gemini
Google Workspace
Mistral
🇪🇺 France
Mistral Large 3, Ministral 3
European, open‑weight (Apache 2.0)
Alibaba — Qwen
🇨🇳 China
Qwen 3.7 Max / Plus
Leading Chinese flagship
DeepSeek
🇨🇳 China
DeepSeek V4 / V4 Pro
Open‑weight, low‑cost, self‑hostable
In Copilot Studio you pick the model — GPT‑5 and Claude are both available. Open‑weight models (Mistral, DeepSeek, open Qwen) can even run on your own infrastructure — relevant for data privacy.
Sources: Microsoft Copilot Studio 2026 release wave 1 (learn.microsoft.com); Mistral AI — Mistral 3 (Dec 2025); Alibaba Cloud — Qwen 3.7 (May 2026); DeepSeek V4 (2026).
Open vs closed weight
Weights = the billions of learned numbers that are the model.
🔓 Open weight
You can download the weights and run them yourself
Host on-prem, fine-tune, inspect — no vendor lock-in
e.g. Mistral, DeepSeek, Qwen (open), Llama
✅ Data privacy ⚠️ needs GPU infra, you maintain it
🔒 Closed / proprietary
Weights stay with the vendor — API / cloud only
Vendor handles infra, updates and safety
e.g. GPT‑5, Claude, Gemini
✅ Easiest, often most capable ⚠️ data leaves your tenant, ongoing cost
Note: open weight ≠ open source — you get the trained weights, but usually not the training data or recipe.
Module 2 · 45 min
The art of the prompt
Same model, garbage or gold — it depends on the prompt.
Weak prompt → weak result
❌ "Write an email"
The model doesn't know about what, to whom, in what tone. It returns something generic.
✅ With context
"You are a consultant. Write a short, professional email to a client who is postponing a meeting to next week. Polite, 4 sentences, propose a new date."
The RCTFE framework
Letter
Element
Example
R
Role
"You are an experienced legal assistant…"
C
Context
"…we're working on a vendor contract…"
T
Task
"…extract all the deadlines."
F
Format
"Return as a table: clause | deadline."
E
Examples
"e.g. Payment | 30 days."
≥3 of these 5 = already a strong prompt.
Patterns that work
Give a role: "Act as…" focuses tone and knowledge.
Ask for a format: table, bullets, JSON, email.
Step by step: "Think step by step" for complex tasks.
Give an example (few-shot): show one desired output — the model copies the style.
Iterate: "shorter", "more formal", "add a risks section".
A note on grounding
A prompt only uses what's in front of the model.
To answer from your documents, you must give them — paste the text, or connect a knowledge source (tomorrow, Mission 4).
If you don't, the model guesses from general knowledge → higher hallucination risk.
🎬 Mini-exercise (5 min)
Take a weak prompt and improve it live with RCTFE.
Start: "summarize this" Target: a 3-bullet summary + 1 recommended action, professional tone.
This is a direct warm-up for Mission 2 tomorrow.
Module 3 · 40 min
Microsoft AI ecosystem
From chat to your own agent — and the bridge to your world.
The three surfaces
M365 Copilot Chat
Free AI chat with web access. Your starting point.
Agent builder
Build your own agent with instructions — no code, inside Copilot Chat.
Copilot Studio (Full)
Full low-code portal — knowledge, actions, connectors, publish to Teams.
Source: learn.microsoft.com — "Agents for Microsoft 365 Copilot Chat" & "What is Copilot Studio" (2026).