Chapters on Eriks training corner

AI for managers

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

AI for managers

Traditional IT vs AI

	Traditional software	AI
Principle	rules written by a programmer	patterns learned from data
Output	same input → same output	same input → similar output
Error	a bug, gets fixed	statistically expected
Testing	once at acceptance	continuously, on your data

The management question is not “does it work?” but “how often is it wrong, and what happens when it is?”

Three families of AI

graph LR P["Predictive AI forecasts and classifies VMS failure, WIM anomalies"] --> G["Generative AI creates text, images, code bids, summaries, translations"] G --> A["Agentic AI plans and acts with tools Cowork, Scout, custom agents"]

More autonomy means more value – and more risk and need for governance.

Use cases: what works and what failed

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

Use cases: what works and what failed

Four value patterns

Pattern	TRAXIA examples
Create	bids, case studies, job ads, translations
Find and summarise	tender documents, manuals, contracts, site meetings
Analyse and decide	project variances, tender win/loss, ticket triage
Act	Cowork assembles the monthly report, an agent creates a ticket

Departments – success and failure

Department	Use cases	Success	Failure
Marketing and sales	tailored bids, meeting prep, win/loss	Lumen: ~4 h/week per seller (Microsoft figure)	Chevrolet 2023: chatbot “sold” an SUV for $1
HR	job ads, onboarding agent, training	Unilever: AI in early hiring rounds	Amazon: CV model penalised women
Finance	variance commentary, payment forecasts, invoice checks	JPMorgan COiN: 360,000 h/year	Zillow 2021: algorithm overpaid for homes
Service	triage, technician assistant, complaint replies	+14 % / novices +34 % (study of 5,179 agents)	Air Canada 2024: the company is liable for its chatbot
IT and development	code, documentation, helpdesk, security	GitHub Copilot: task 55 % faster	PocketOS 2026: agent deleted the production DB

AI in ITS – processes vs products

AI in how we work – the company is a deployer: bids, service planning, reporting, technician know-how.
AI in what we sell – the company can be a provider: WIM anomalies and calibration, video analytics, predictive maintenance of VMS and level crossings, adaptive junction control.

AI Act

AI systems intended as safety components in road-traffic management are high-risk (Annex III). Obligations from 2 December 2027.

Strategy, roadmap and ROI

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

Strategy, roadmap and ROI

Why pilots fail

95 % of GenAI pilots with no measurable P&L impact (MIT NANDA, 2025)
buying from a specialist succeeds ~2× more often than building in-house
40 %+ of agentic projects will be cancelled by end of 2027 (Gartner)

AI strategy on one page

graph TB A["Ambition – why, what by 2028"] --> U["Priority use cases – 3 to 5"] U --> P["Platform and data"] P --> G["Governance and risk"] G --> L["People and skills"]

Fit test (0–2 points)

Volume · text/data work · data available · error tolerance · baseline · process owner. 9–12 candidate, 6–8 rework, 0–5 not now.

Security, trust, legislation

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

Security, trust, legislation

Hallucinations

The model optimises plausibility, not truth. Types: invented sources, wrong numbers, outdated information, a true fact from the wrong source, an “over-eager” answer. What helps: citations and checking them, grounding, calculations by tool.

Review ladder

Level	Output	Check
1	Internal drafts	the author checks it
2	Decision support	verify numbers and sources, disclose AI use
3	External outputs	second person + approval
4	Legal, financial, safety	AI only prepares, a human decides
STOP	Automated decisions about people, emotions	prohibited

Grounding

graph LR Q["Question"] --> S["Search company documents"] S --> M["Model + retrieved passages"] M --> A["Answer with citation"]

Answer quality ≤ document quality. Watch for oversharing – Copilot sees everything the user can access.

Choosing and deploying tools

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

Choosing and deploying tools

Four levels

Level	For what	Microsoft	Alternatives
Use	ready-made assistant for everyone	M365 Copilot, Copilot Chat	ChatGPT Enterprise, Claude, Gemini
Configure	agent on company knowledge	Agent Builder, Copilot Studio	custom GPTs, Claude projects, Gemini Gems
Delegate	agent carries out multi-step work	Copilot Cowork, Microsoft Scout	agent modes in ChatGPT, Claude
Build	custom solution, integrations, products	Microsoft Foundry	AWS Bedrock, Google Vertex AI

Decision tree

graph TB Q1{"Everyday office work of an individual?"} -- yes --> U["USE: Copilot"] Q1 -- no --> Q2{"Does a team need answers from company knowledge?"} Q2 -- yes --> C["CONFIGURE: Agent Builder / Copilot Studio"] Q2 -- no --> Q3{"A long multi-step task across files?"} Q3 -- yes --> D["DELEGATE: Copilot Cowork"] Q3 -- no --> Q4{"Is there a ready specialist product?"} Q4 -- yes --> B["BUY"] Q4 -- no --> F["BUILD: Foundry"]

Copilot Cowork

Generally available since 16 June 2026. You delegate a task, not a prompt: it plans the steps, works across files, mail and Teams, runs in the cloud, has checkpoints. Requires an M365 Copilot licence and is charged by consumption (Copilot Credits) – the task cost is visible, with limits per tenant, group and user. Off by default; audit, eDiscovery, sensitivity labels.

Change management and adoption

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

Change management and adoption

10 – 20 – 70

AI success: 10 % algorithms, 20 % technology and data, 70 % people and processes (BCG principle). The strongest adoption factor in the UK DBT study was the line manager’s attitude.

People’s fears and answers

Fear	What to say
“It’ll take my job.”	what changes and where freed time goes – don’t lie
“I won’t be able to do it.”	training, champions, room for mistakes
“They’ll monitor me.”	clear rules – AI metrics don’t rate people
“The results are bad.”	feedback, a better use case, better data

Good: IKEA (2023) – bot Billie handled 47 % of queries; 8,500 call-centre staff were retrained as design advisers. Bad: Klarna, Commonwealth Bank, Duolingo – “AI will replace people” messaging led to lower quality, backlash and U-turns.