Cheat sheets on Eriks training corner

AI policy template

Mon, 01 Jan 0001 00:00:00 +0000

Template: a one-page AI policy

A template for a department or a smaller company. Adapt it with IT, your data protection officer (DPO) and a lawyer – it is not legal advice.

Valid from: ………… Owner: ………… Review: every 6 months

1. Why we use AI

To have more time for work that needs people – customers, problem solving, quality. AI is a helper; a human is always responsible for the result.

Prompts for managers

Mon, 01 Jan 0001 00:00:00 +0000

Cheat sheet: AI as a partner in management decisions

Prompts for strategic work with Copilot. The Copilot-day principle Goal · Context · Source · Expectations still applies.

Finding opportunities

I lead [department]. Here is a list of our main activities with volumes: [list]. Which three would gain most from AI, and why? For each, say what we'd need to have ready (data, people, rules).

@Researcher How do companies in intelligent transport systems and road management in the EU use AI in service and equipment maintenance? Give concrete examples with links to sources; separate independent sources from vendor marketing.

Business case and challenge

You are a sceptical CFO. Here is my AI business case [paste]. Ask me 5 uncomfortable questions and for each say what evidence would convince you.

Propose measurable success criteria and exit criteria for this pilot. The criteria must be measurable within 8 weeks and start from today's state: [baseline].

Risks

Review this AI use-case proposal and place it on the review ladder (1 internal, 2 decision support, 3 external, 4 legal/financial/safety). List the GDPR and AI Act risks and propose controls. Flag it if it could be a high-risk system.

Communicating with the team

Draft a short e-mail to my team announcing that we're starting a pilot of [use case]. Tone: open, no hype. Explain why we're doing it, what changes, what does NOT change, where the saved time will go and whom they can contact. Max 150 words.

My people are worried AI will replace them. Prepare 5 honest answers to the most common fears for a team meeting. Don't promise anything I can't guarantee.

Choosing a vendor

I'm preparing a meeting with a vendor of an AI solution for [use case]. Suggest 10 questions covering: value, data protection and EU residency, security, integration, costs at higher usage, evidence of quality, lock-in and exit, support. Add red flags to watch for in the answers.

Glossary

Mon, 01 Jan 0001 00:00:00 +0000

Glossary

Term	Explanation
Agent	AI given a goal, tools and permissions that carries out several steps by itself (e.g. Copilot Cowork, a custom Copilot Studio agent).
Agent 365	Microsoft’s tool to inventory, identify, govern and block AI agents in a company.
AI Act	The EU regulation on artificial intelligence. It classifies AI by risk: prohibited, high risk, transparency, minimal.
AI literacy	Companies’ duty to support people who use or decide about AI in understanding it (AI Act Art. 4).
Baseline	The measured state before a change – without it, benefit can’t be proven.
Copilot Cowork	A Microsoft 365 Copilot mode to which you delegate a longer multi-step task; charged by consumption.
Copilot Studio	A tool for building agents with actions and connectors to company systems.
Deployer / provider	AI Act roles: who uses AI (deployer) vs who develops it and places it on the market (provider).
DPIA	Data protection impact assessment – for high-risk processing (GDPR Art. 35).
Evaluation	Systematic testing of AI quality on a test set of questions – accuracy, grounding, safety.
Generative AI	AI that creates new content: text, images, code (Copilot, ChatGPT, Claude, Gemini).
Grounding (RAG)	AI answers from retrieved company documents, not “from memory”, and cites the source.
Guardrails	Limits – content filters, topic and action restrictions, abuse protection.
Hallucination	A plausible-sounding but false AI answer.
Jagged frontier	AI excels at some tasks and fails at similar-looking ones – you can’t see it from outside.
Microsoft Foundry	An Azure platform for building custom AI apps and agents (formerly Azure AI Foundry).
Microsoft Scout	Microsoft’s always-on personal agent (preview since 6/2026); works without prompting.
Oversharing	Documents shared too widely – AI finds and uses them.
Predictive AI	Machine learning that forecasts or classifies (failure, anomaly, demand).
Prompt injection	An attack where hidden instructions in an e-mail or document take control of the AI.
Realisation factor	The share of saved time that actually turns into value.
Shadow AI	AI tools employees use without company approval.

Stories and sources

Mon, 01 Jan 0001 00:00:00 +0000

Stories and sources

Examples from the course with sources. Status as of October 2026 – treat vendor figures as illustrations, not benchmarks.

Research and surveys

Topic	Key finding	Source
Jagged frontier (HBS & BCG, 2023)	758 consultants: inside the frontier +12 % tasks, 25 % faster, +40 % quality; outside it 19 pp fewer correct solutions	SSRN 4573321
Generative AI at Work (2023)	5,179 support agents: +14 % issues resolved, novices +34 %	NBER w31161
METR (July 2025)	experienced developers 19 % slower with AI, believed they were 20 % faster	METR
UK DBT – M365 Copilot (2024–25)	72 % satisfied, no robust evidence of productivity gains	The Register
GitHub Copilot (2023)	controlled experiment: task 55.8 % faster	arXiv 2302.06590
MIT NANDA (2025)	95 % of GenAI pilots with no P&L impact; buying ~67 % success	Fortune
Gartner (2025)	40 %+ of agentic projects cancelled by end of 2027	Gartner

Successes

Example	What happened	Source
Lumen (2024)	sellers save ~4 h a week, ~$50 M a year (Microsoft customer story)	Microsoft
JPMorgan COiN (2017)	reading loan agreements – 360,000 hours of work a year	ABA Journal
Unilever (2017–2019)	AI in early rounds of graduate hiring	Business Insider
IKEA (2023)	bot Billie 47 % of queries, 8,500 staff retrained as design advisers	Retail Gazette

Failures

Example	What happened	Source
Chevrolet of Watsonville (2023)	chatbot “agreed” to sell a Tahoe for $1	AI Incident Database
Amazon (2014–2018)	CV model penalised women, scrapped	CNBC / Reuters
Zillow Offers (2021)	algorithm overpaid for homes, unit closed, −25 % staff	CNBC
Air Canada (2024)	the company is liable for what its chatbot says	CanLII 2024 BCCRT 149
Klarna (2024–2025)	AI “doing the work of 700”, then lower quality and rehiring	Entrepreneur
Mata v. Avianca (2023)	6 invented rulings, $5,000 fine	Wikipedia
Deloitte Australia (2025)	report with invented citations, partial refund	CTV News
Samsung (2023)	source code in public ChatGPT, AI ban	The Register
McDonald’s (2024)	end of the IBM AI drive-thru test	Nation’s Restaurant News
EchoLeak (2025)	zero-click prompt injection in M365 Copilot, fixed	The Hacker News
Commonwealth Bank (2025)	voice bot replacing 45 people, U-turn and apology	Bloomberg
Duolingo (2025)	“AI-first” memo, user backlash	Fortune
OpenClaw (February 2026)	agent deleted hundreds of e-mails, ignored instruction	OECD AI Incidents
PocketOS (April 2026)	coding agent deleted the production database and backups	Zenity

Tools and legislation

Topic	Source
Copilot Cowork – GA 16 June 2026	Microsoft 365 Blog
Microsoft Scout – launch (Build 2026)	TechCrunch
Microsoft Scout – admin and prerequisites	Microsoft Learn
Agent 365 – GA May 2026	CloudEagle
AI Act after the Digital Omnibus – timeline	Future of Privacy Forum
AI Act – Omnibus changes	Morgan Lewis
AI regulation in Slovakia	CMS AI Regulation Scanner