Courses

The full course catalogue

The courses run across 10 months: 20 shared courses every Igniter takes together, plus dedicated track courses per direction. The catalog is living — we adjust it to where the market is.

courses in total: 44
shared courses (all): 21
unique courses per track: 8

Showing every course in the cohort

Felleskurs — tas av alle sporSporkurs — spesifikt for Builder/Tuner/Shaper

Foundations

September 20262 courses in total

Responsible AI
Alle
Establish the ground rules for AI work: accountability, traceability, client confidentiality.
Wed 9 Sept90 min masterclass (onsdag 15:00–16:30)
An introduction to responsible AI delivery. What does it mean to build systems that are transparent, auditable and respect regulatory requirements? The course sets the quality standard every later course builds on.
What you leave with
Can explain the difference between bias, hallucination and drift in AI systems
Can classify an AI solution against the EU AI Act risk classes — and knows what applies now versus from December 2027
Can explain the most common agent risks (prompt injection in agent flows, tool misuse) with concrete examples
Can identify ethical risks in a client case and propose mitigations
What you learn
The EU AI Act after the Digital Omnibus (July 2026): what is enforced now (transparency, GPAI) and what arrives in 2027/2028
Norway: the EEA process is under way — Nkom becomes the supervisory authority
Agent risks in plain terms: instructions hidden in content the agent reads, tools being misused, and why humans must approve consequential actions
Sources of bias: data, model, prompt, evaluation
Transparency, explainability and traceability as delivery principles
Data governance and privacy (the GDPR intersection)
Audit trails: what has to be logged to be able to defend the solution
Teknologier
EU AI Act (post-Omnibus)
OWASP Top 10 for Agentic Applications (introduksjon)
ISO/IEC 42001
NIST AI RMF
GDPR
Forutsetninger
No technical prerequisites. A basic grasp of AI as an industry term is enough.
LLM Foundations
Alle
The groundwork: how language models reason, answer and use context.
Wed 23 Sept90 min masterclass
How LLMs actually work: tokens, context, attention, temperature. An engineer's mental model of a language model — not marketing simplifications.
What you leave with
Can explain token economics and the context window as engineering decisions
Can steer model reasoning with effort levels — and knows why temperature and top-p are history
Can get guaranteed valid JSON out of a model using native structured outputs
Can estimate cost per request — and knows that prompt caching is the biggest cost lever
What you learn
Tokenisation — how text becomes numbers (and why tokenizers are model-specific)
The context window in the 1M era: what the model actually sees, and why more context is not always better
Prompting and context engineering: write good instructions AND curate what the model gets to see
Reasoning control: adaptive thinking and effort levels — the successor to temperature
System, user and assistant messages, streaming and tool use
Structured outputs: guaranteed JSON schema, and how the providers differ
Robustness from day one: handle refusals and stop_reason before you read the answer
Prompt caching: up to 90% cheaper when the prompt is structured correctly
Teknologier
Anthropic Claude API
OpenAI API
Google Gemini API
Python
TypeScript
Structured outputs
Prompt caching
Effort / adaptive thinking
Forutsetninger
Has called an LLM API at least once (or has taken the Responsible AI course).

Discover

October 20267 courses in total

Provider-independent LLM architecture
Builder
Why you should speak three languages with LLMs.
Sun 18 Oct90 min Builder-breakout på helgesamling 1 (søndag 18. oktober)
A practical introduction to multi-provider architecture: why not to lock yourself to one LLM vendor, and how to switch without tearing the code apart. For the Builder track.
What you leave with
Can implement a provider abstraction in Python and TypeScript
Understands the trade-offs between Anthropic, OpenAI and Azure for different tasks
Can migrate a prompt set between providers without losing quality
Knows how to measure quality differences objectively
What you learn
Provider SDKs: what they share and where they differ
The model router pattern
Feature differences: tool calling, streaming, vision
Fallback strategies for provider failures
Teknologier
Anthropic SDK
OpenAI SDK
Azure OpenAI
AWS Bedrock
Model router-patterns
Google Vertex AI (Gemini)
Forutsetninger
LLM Foundations. Comfortable with Python or TypeScript.
AI economics and FinOps
Alle
Token cost, caching, routing — and how to defend an AI budget.
Wed 21 Oct90 min masterclass
Unit economics for AI solutions. What a request costs, how we price towards the client, where break-even sits, and when fine-tuning pays off.
What you leave with
Can calculate cost per request for a concrete architecture — including thinking tokens
Knows which levers move cost most: caching, effort level, model choice and batching
Can build a TCO model for an AI service across 12 months
Can weigh RAG against long context and fine-tuning on economics
What you learn
The pricing landscape (as of October 2026) — and why you learn to look it up live rather than memorise numbers
Input versus output — and thinking tokens as an invisible cost driver
Prompt caching economics: 0.1× on reads, break-even after two or three calls
Batch APIs: 50% discount at all three major providers
1M-context economics: what it costs to fill the window — and to do it fifty times in an agent loop
TCO: model plus infrastructure plus maintenance plus eval
Open models and self-hosting: when it (rarely) pays off
Markup strategies for consulting delivery
Teknologier
Token pricing (Claude Opus 5, GPT-5.6, Gemini, DeepSeek)
Prompt caching
Batch API
Langfuse
LiteLLM
TCO-modellering
Forutsetninger
LLM Foundations.
AI readiness assessment methodology
Shaper
Find out whether the client is actually ready — or just talking about AI.
Sun 18 OctShaper-breakout på helgesamling 1 (søndag 18. oktober)
AI readiness assessment: how to judge whether an organisation is ready to adopt AI, where the gaps are, and which moves actually help.
What you leave with
Can run a readiness assessment at a client
Knows how to score six dimensions: strategy, data, tooling, process, culture, governance
Can propose concrete moves for each maturity level
Has a framework for reporting to leadership
What you learn
The Capgemini AI Readiness framework
From adoption readiness to execution readiness: can the client actually scale?
Data modernisation as travelling companion: half of agent projects require it
Orchestration infrastructure and agent identity as new assessment dimensions
Interview technique that produces honest answers
Scoring and gap analysis
How to avoid the theatre assessment
Teknologier
Capgemini AI Readiness-rammeverk
Gartner AI Maturity
MIT Sloan AI Index
Notion-templates
Forutsetninger
"Business case for an AI initiative" is recommended (it runs 4 November — after this session); the core concepts are introduced in the session.
Model selection and reading benchmarks
Tuner
Which model should the client use — and how do you know you are right?
Sun 18 Oct2,5 timer Tuner-breakout på helgesamling 1 (søndag 18. oktober)
The first question you get at a client is "which model should we use?". This course gives you a testable answer instead of a gut feeling: reading public benchmarks critically, setting up a quick blind head-to-head on the client's own task, and weighing quality against cost and latency. It is the entry point to the Tuner track — the ground all later evaluation work rests on.
What you leave with
Can read a public benchmark critically — and explain what it does NOT say about the client's task
Can set up a blind head-to-head comparison of three models on a real task
Can justify a model choice on quality, cost and latency — and document the trade-off
Knows what should trigger a re-evaluation of the model choice later in the delivery
What you learn
The benchmark landscape: what they measure, and where they leak
Contamination and saturation — why a top score is not enough
Blind head-to-head: same prompt, same data, model name hidden
The decision matrix: quality versus cost versus latency
Model cards and re-evaluation criteria — the choice as a document
Teknologier
Claude
Microsoft Foundry
Python (notebook)
Regneark
Forutsetninger
The shared course LLM Foundations (23 September). No coding experience beyond running a prepared notebook.
Delivery quality in AI projects
Alle
How the Ignite team works. What "ready" and "done" mean on an AI project.
Wed 7 Oct90 min masterclass + workshop
Team practice for AI delivery: what counts as done, how we work, who owns what. The cohort's own Definition of Ready and Definition of Done.
What you leave with
Has a Definition of Ready and Definition of Done written for the team
Knows how a task moves from ready to done inside the team
Can facilitate a team check-in and resolve conflict
Has established shared language for delivery quality
What you learn
Definition of Ready versus Definition of Done — what they mean for AI cases
Dividing responsibility between Builder, Tuner and Shaper in a pair or team
Retrospective practice — fast and honest
Handling conflict, and knowing when to escalate
Teknologier
Definition of Ready / Done
Git workflow
GitHub
Notion / Confluence
Forutsetninger
None.
AI-assisted development in the client's codebase
Alle
Build software with AI as an active collaborator
Sat 17 OctWorkshop på helgesamling 1, 17.–18. oktober
Everyday development work has changed fundamentally — Claude Code, Cursor, Windsurf and similar tools let you program with AI as an active collaborator, not just autocomplete. This course is foundational for every Igniter regardless of track: everyone should be able to work effectively with these tools in 2026.
What you leave with
Can use Claude Code or Cursor effectively for real development work
Can work plan-first: describe, plan, break down, implement — and knows when a full spec is overkill
Knows the difference between instructions (CLAUDE.md/AGENTS.md), knowledge (Agent Skills) and enforcement (hooks and permissions)
Can work safely with AI tools inside a client codebase — and knows what is never shared
What you learn
The 2026 tool landscape: Claude Code, Cursor, Codex CLI, Devin — and why it moves so fast
Plan-first workflow and spec-driven development (GitHub Spec Kit) — and when that is overkill
CLAUDE.md/AGENTS.md: the open standard you will meet in client repos
The toolkit in plain terms: hooks enforce, Skills teach, subagents delegate
MCP servers against internal systems — with allowlisting and healthy scepticism
The verification loop: the agent proves it with tests rather than asserting it
The consultant's rule: the client's approved AI stack applies. Always.
Cost and speed — when does AI-assisted coding pay off?
Teknologier
Claude Code
Cursor
OpenAI Codex CLI
GitHub Copilot
AGENTS.md
Agent Skills
MCP
GitHub Spec Kit
Forutsetninger
Basic programming. Recommended after LLM Foundations.
Presentation skills
Alle
Confident delivery, and demos that land.
Sat 17 Oct2,5 timer workshop under helgesamlingen 17.–18. oktober
A two-hour workshop combining general presentation technique with AI-specific demo craft. You learn to present an AI delivery with structure, run a live demo without freezing when the model misbehaves, and handle sceptical questions with calibrated confidence and humility.
What you leave with
Can present an AI delivery in under five minutes with a clear problem-solution-value structure
Can run a live demo of an AI system without losing composure if the model fails
Can answer "what does the AI do when it goes wrong" without overselling
Knows how to calibrate confidence and humility in an AI pitch
What you learn
The structure of a good technical presentation (hook, problem, solution, evidence, proposal)
Pitch narrative for AI: concrete rather than abstract
Live demo craft: preparation, fallback, and what you do when it fails
Voice, pace, pauses — the vocal toolkit
Q&A under pressure — handling critical questions
Ethics and honest calibration of AI promises
Teknologier
Minto Pyramid
Slides (Google / Figma)
Loom
Live demo-håndverk
Forutsetninger
No prerequisites. Every Igniter takes this early in Foundations so they can use it through the rest of the year.

Build

November–December 202610 courses in total

The EU AI Act in practice: DPIA and documentation
Shaper
Produce the documentation pack before the regulator knocks.
Wed 25 Nov90 min masterclass + case-arbeid
Practical AI Act implementation for the Shaper track: not just understanding the law, but actually rolling out compliance processes at a client.
What you leave with
Can lead an AI Act compliance workshop at a client
Has a template for the risk register and the documentation pack
Can set up compliance gates in the delivery pipeline
Knows how to train the client's team to maintain compliance
What you learn
The adviser's message after the Omnibus: postponed is not cancelled — use the breathing room for a gap analysis
Compliance as part of the delivery process, not an add-on
Risk register template and maintenance
Gate-based compliance checklist
The likely audit scenario, and preparing for it
Teknologier
EU AI Act compliance-prosesser
ISO/IEC 42001
Risk register
Audit-prep
Forutsetninger
EU AI Act risk classification, AI readiness assessment methodology.
Evaluation craft: rubrics and LLM-as-judge
Tuner
When the machine does the judging — calibrate it against people you trust.
Wed 9 Dec90 min masterclass + 3 timer lab
A follow-on from the shared course "Evaluation in practice", where everyone learned to set up an eval suite. Here you go deeper into the craft itself: writing rubrics several assessors agree on (inter-rater agreement), calibrating an LLM judge against human judges, and knowing when the judge cannot be trusted. This is the skill that separates a Tuner from someone who has merely set up evals.
What you leave with
Can write an eval rubric with 85% or better inter-rater agreement
Can calibrate an LLM judge against human judges
Knows how to detect eval drift
Can present eval results to a non-technical audience
What you learn
Rubric design and iteration
Inter-rater reliability (Cohen's kappa, Fleiss')
LLM-as-judge: bias and calibration
Eval drift and how to mitigate it
Reporting: charts decision-makers understand
Teknologier
LLM-as-judge
Cohen's / Fleiss' kappa
Rubrikk-design
Langfuse
Statistical evals
Forutsetninger
Evaluation in practice.
Robust AI services in production
Builder
Streaming, retries, circuit breakers — the patterns that keep agents alive in production.
Wed 11 Nov90 min masterclass
Production patterns for AI services: rate limits, retries, circuit breakers, idempotency, observability. Not so different from any distributed system — but LLMs have quirks of their own.
What you leave with
Can build a resilient LLM service with retries and a circuit breaker
Understands token rate limits and how to stay clear of them
Can design idempotent calls that involve an LLM
Knows how to handle provider outages
What you learn
Retry strategies: exponential backoff, jitter, budget
Circuit breakers and rate limits — designing with gateway primitives
Idempotency in an LLM context
Structured outputs in production: native schema enforcement, value validation, retry design
Streaming combined with structured output
Error handling: refusals, truncation and tool failures
Teknologier
Retry / circuit breakers
OpenTelemetry
Sentry
Structured output (Zod/Pydantic)
Rate-limit handling
Forutsetninger
Provider-uavhengig LLM-arkitektur.
Business case for an AI initiative
Alle
Build an ROI model that survives the meeting with the client's CFO.
Wed 4 Nov90 min masterclass + workshop
How to write an AI business case that gets funded. From idea to numbers the client will actually accept.
What you leave with
Can write a business case aimed at the right decision-makers
Knows how to estimate ROI without overselling
Has tools to sanity-check their own assumptions
Can present the case in ten minutes to client leadership
What you learn
The problem statement: the right depth for the audience
From pilot ROI to scaling economics: why most pilots never reach the accounts — and what distinguishes the ones that do
Workflow redesign as the value driver: AI bolted onto an existing process rarely moves the needle
ROI calculation: time saved, quality, risk reduction
Outcome-based pricing: cost per completed task as a model
Assumptions versus facts — and sensitivity analysis (what if cost doubles?)
Decision-maker personas: what a CFO, a CIO and a head of operations actually ask
Teknologier
ROI-rammeverk
Sensitivitetsanalyse
Stakeholder-mapping
Notion/Figma
Forutsetninger
AI economics and FinOps.
Process redesign with AI
Shaper
Map the existing process, find the AI leverage points, redesign the value stream.
Wed 9 Dec90 min masterclass + workshop
AI without process gives fragmented value. You learn to map a client process (BPMN-style), identify where AI genuinely provides leverage versus where it only adds complexity, and redesign the value stream so people and AI complement each other. Includes real cases from Norwegian banking, energy and the public sector.
What you leave with
Map an existing client process with structured methods
Identify AI leverage points with a quantified business case
Design a new process that balances automation and human control
Lead a redesign workshop with the client and subject-matter experts
What you learn
BPMN basics and value-stream mapping
Leverage analysis: where AI actually pays off
Human-in-the-loop design and approval points
Time and cost modelling, before and after
Change strategies and role transitions
Case studies: banking, energy, public sector
Teknologier
Celonis process mining
Miro / FigJam
BPMN
Service blueprints
Forutsetninger
The shared business case course and the Discover phase as a whole.
Context engineering: what the model actually sees
Tuner
Manage the context window like an engineer, not like a writer
Wed 6 Jan90 min workshop + 2 timer lab
De fleste kvalitetsproblemer i en LLM-løsning er ikke modellproblemer — de er kontekstproblemer. Feil ting kom inn i vinduet, riktig ting falt ut, eller ingen vet hva som lå der da svaret ble feil. Dette kurset handler om å behandle konteksten som et budsjett du styrer og en artefakt du versjonerer: hva som må inn, hva som kan komprimeres, hva som skal ut, og hvordan du måler effekten av en endring i stedet for å gjette.
What you leave with
Kan sette opp et kontekstbudsjett for en konkret oppgave: hva som er systeminstruksjon, hentet materiale, historikk og verktøysvar — og hva hver del koster
Kan komprimere samtalehistorikk og hentede dokumenter uten å miste det svaret faktisk avhenger av
Kan behandle prompter som versjonerte artefakter med endringslogg, og A/B-teste en endring mot et evalueringssett i stedet for å vurdere den for hånd
Kan designe structured output med skjemavalidering som ikke knekker når modellversjonen byttes
Kan finne ut hvorfor caching ikke slår inn, og hva som må endres for at den skal gjøre det
What you learn
Kontekstbudsjett: de fire delene av vinduet og hva hver av dem koster
Komprimering av historikk — oppsummering, vindusstrategier og hva som går tapt
Retrieval-budsjett: hvor mange dokumenter er nok, og når flere gjør svaret dårligere
Prompt som versjonert artefakt: endringslogg, gjennomgang og utrulling
Structured output og skjemavalidering som overlever modellbytte
Cache-strategi: hva som må stå fast for at et treff skal skje, og hvorfor korte prompter ofte ikke kvalifiserer
Feilsøking: hvordan finne ut hva som faktisk lå i vinduet da svaret ble feil
Teknologier
Azure OpenAI / Microsoft Foundry
Anthropic og OpenAI API-er
Skjemavalidering (JSON Schema, Pydantic eller tilsvarende)
Versjonskontroll for prompter
Tokenisering og telleverktøy
Forutsetninger
Felleskurset LLM Foundations og felleskurset om evaluering. Du bør ha kjørt minst én prompt mot et API selv.
Quality gates in the delivery pipeline
Builder
Regression-test prompts and agents — break the build when quality drops.
Wed 9 Dec90 min masterclass + lab
Evaluation inside the CI/CD pipeline. How to run evals on every pull request, stop regressions, and hold a release gate for quality.
What you leave with
Can set up eval runs in GitHub Actions or Azure DevOps
Knows how to keep costs under control in CI
Can define release gates based on eval score
Has a reporting dashboard for eval trends over time
What you learn
Cost-aware eval scheduling
Handling flaky evals
Parallel execution and caching
Release gates and branch protection
Setting up the eval dashboard
Teknologier
GitHub Actions
Langfuse / LangSmith
Promptfoo
Release gates
Cost-aware scheduling
Forutsetninger
Evaluation in practice, Robust AI services in production.
RAG engineering and hybrid search
Alle
How to find the right context and deliver grounded answers.
Wed 16 Dec90 min masterclass + 4 timer lab
RAG — retrieval-augmented generation. From simple vector search to hybrid pipelines, re-ranking and graph RAG. What actually works on enterprise data. If you do not code: you work in a pair with a Builder in the lab and own the quality and content side — knowing why a RAG solution answers wrongly is gold in client meetings, whatever your track.
What you leave with
Can design a RAG pipeline based on the client's data formats
Knows how to choose an embedding model and a chunking strategy
Can measure retrieval quality with recall@k and MRR
Is comfortable with hybrid search (vector plus BM25) and re-ranking
What you learn
From pipeline RAG to agentic retrieval: the agent decides when and what to fetch
Context architecture: curation, compression and just-in-time retrieval
Managed knowledge bases versus DIY: the decision frame
Hybrid retrieval and re-ranking as the production baseline
Retrieval storage: pgvector as the default — and when a specialised database is genuinely needed
Chunking and embedding choices (background)
Retrieval evals as the governing discipline: recall@k, MRR, nDCG
Graph RAG: when a knowledge graph beats vectors
Teknologier
pgvector
Azure AI Search (agentic retrieval)
Bedrock Managed Knowledge Base
Hybrid retrieval
Re-ranking (Cohere, Voyage)
Langfuse
Forutsetninger
LLM Foundations.
EU AI Act risk classification
Alle
Place a system in the right category and document why.
Wed 18 Nov90 min masterclass + lab
The EU AI Act in depth: risk classification in practice, documentation requirements for high-risk systems, and how to build compliance into the delivery rather than bolting it on.
What you leave with
Can classify a concrete AI solution under AI Act Article 6
Can explain the new timeline after the Digital Omnibus: what applies now, what arrives in 2027/2028
Can write risk management documentation that satisfies Article 9
Knows what is required for logging and human oversight (Articles 12 and 14)
What you learn
The Digital Omnibus (2026): the new timeline and what it means for clients
The risk classes: prohibited, high-risk, limited, minimal
Transparency obligations for generative AI (Article 50) — in force NOW
Risk management system — Article 9
Data governance — Article 10
Human oversight — Article 14
Documentation — Annex IV
Norway: EEA status and Nkom as the incoming supervisory authority
Teknologier
EU AI Act (Art. 6, 9, 10, 14, 50)
Digital Omnibus (2026/1744)
Annex IV-dokumentasjon
Risk register-maler
ISO/IEC 42001
Forutsetninger
Responsible AI.
Evaluation in practice
Alle
Golden sets, LLM-as-judge and an eval rig you can trust.
Wed 2 Dec90 min masterclass + 2 timer lab
Evaluation culture as craft. From ad-hoc smoke tests to regression-safe evals, golden sets and calibrated LLM judges. If you do not code: evaluation is at least as much method as code — you contribute test design and assessment criteria, which is the heart of the discipline.
What you leave with
Can build an eval suite that runs on every pull request
Knows how to calibrate an LLM-as-judge against human judges
Can maintain a golden set as requirements move
Understands the false positive rate in their own evaluation
What you learn
Rubrics and the scoring manual
Golden set curation and maintenance
LLM-as-judge: calibration and inter-rater reliability
Regression evals in CI
Tracking eval drift over time
Teknologier
Promptfoo / Braintrust
DeepEval / Ragas
Anthropic eval-rammeverk
Langfuse
LLM-as-judge
Golden sets i git
Forutsetninger
LLM Foundations.

Operate

January–February 202713 courses in total

AI rollout and launch
Shaper
How to take an AI delivery from finished to actually in use
Wed 20 Jan90 min masterclass + caseøvelse
The Shaper focus in the Operate phase: rollout strategy for AI solutions. Phased deployment, A/B testing of AI features, GO/NO-GO criteria, and communication during quality incidents. Launch is often where AI projects still fail — even after a technically successful build.
What you leave with
Can design a phased rollout plan for an AI delivery
Knows how to measure success in an AI launch — beyond "it shipped"
Can communicate during quality incidents without losing trust
Can define GO/NO-GO criteria for an AI launch
What you learn
Phased rollout patterns: canary, dark launch, percentage rollout
A/B testing of AI features (with evals as the baseline)
GO/NO-GO criteria for an AI launch — decision frameworks
Rollback strategies and the kill switch
Launch communication: internal versus external
Handling "the AI did something odd" incidents
Teknologier
Feature flags (LaunchDarkly, Unleash)
A/B-testing
Statsig
Canary / blue-green deploy
Forutsetninger
Assumes an understanding of the AI delivery process — taken after the Build phase.
Distillation and small models for cost and latency
Tuner
When training pays off, and when you should simply not.
Wed 17 Feb90 min masterclass + lab
Fine-tuning: when it pays off, which models, which data. From LoRA to a full fine-tune — and why RAG is often the better answer.
What you leave with
Can judge whether fine-tuning is the right move
Can carry out LoRA fine-tuning on an open model
Understands data requirements and evaluation after fine-tuning
Knows the cost and maintenance trade-offs
What you learn
The decision first: RAG versus long context versus fine-tuning — fine-tuning is rarely the answer
The commercial case: distillation into small or open models for cost and latency
SLMs in practice: when a small model does the job at a fraction of the price
QLoRA on open models (Qwen, DeepSeek, Mistral) with Unsloth and TRL
Data curation: 500 to 2,000 curated examples beat volume
Evals before and after: overfitting and distributional shift
Model management after tuning: versioning and rollout
Teknologier
Unsloth · TRL/PEFT
Åpne modeller (Qwen, DeepSeek, Mistral)
Modal / RunPod
Bedrock custom models / RFT
Hugging Face
LoRA / QLoRA
Forutsetninger
RAG engineering, Evaluation in practice.
Azure deployment for Norwegian clients
Builder
Cloud architecture that meets Norwegian data requirements and survives the security committee.
Wed 6 Jan90 min masterclass + Bicep-lab
Azure dominates Capgemini Norway's delivery work. You learn to architect AI solutions that respect Norwegian data sovereignty — region choice, BYOK patterns, private endpoints — and how to argue the cloud choice in front of security-conscious clients in finance, energy and the public sector.
What you leave with
Choose the right Azure region for Norwegian client data
Set up Azure OpenAI / AI Foundry with private endpoints
Implement a BYOK pattern with Azure Key Vault
Argue for a cloud architecture in front of a security committee
What you learn
Azure North Europe versus Sweden Central — data storage and residency
Azure OpenAI / AI Foundry architecture
Private endpoints and VNet isolation
BYOK with Azure Key Vault and double encryption
DPAs, GDPR and residency requirements
Bicep and Terraform for AI infrastructure
Teknologier
Azure OpenAI
AWS Bedrock
Google Vertex AI
Bicep
Terraform
Container Apps / Fargate / Cloud Run
Forutsetninger
Basic Azure familiarity (subscription, resource group, RBAC).
Agentic architecture and orchestration
Alle
Build an agent that plans, uses tools and handles state.
Wed 13 Jan90 min masterclass + 4 timer lab
From a single prompt to an agentic system. How to design, build and operate agents that actually carry out tasks against external systems — with traceability. If you do not code: you take part in the design and evaluation half of the lab. Understanding how agents interact is core competence when explaining solutions to client leadership, and nobody expects you to write the code.
What you leave with
Can implement a multi-step agent with tool use
Understands the difference between ReAct, plan-and-execute and hierarchical orchestration
Can design approval points and fallback handling
Can trace agent decisions for audit
What you learn
Agent patterns: ReAct, plan-and-execute, and harness/subagent patterns
Tool and function calling across providers
MCP (Model Context Protocol) — the 2026 spec, stateless core
Agent Skills and A2A: the standards around the agent ecosystem (awareness level)
LangGraph, Pydantic AI (v2) and Microsoft Agent Framework — a landscape overview
Approval points and human-in-the-loop
Observability for agent decisions
Teknologier
LangGraph
Pydantic AI (v2)
Claude Agent SDK
MCP (2026-spec)
Agent Skills
A2A
Microsoft Agent Framework
Forutsetninger
The shared course LLM Foundations. Builders also benefit from the multi-provider course, but it is not a requirement.
Dataset craft and labeling
Tuner
Build golden datasets that genuinely teach the model — and the client — something.
Wed 20 Jan90 min masterclass + lab
Model quality is set by the data. You learn to design golden datasets, run labeling pipelines, calibrate annotators, and iterate on data when evals show where the model fails. This is the daily craft of the Tuner role — and what separates a consultant who understands the model from one who only swaps the prompt.
What you leave with
Design a golden dataset that mirrors real client usage
Set up and calibrate a team of annotators
Use active learning and hard-case mining to improve the model
Uncover hidden bias and coverage gaps in a dataset
What you learn
Dataset design and statistical coverage
Annotator guidelines and inter-rater agreement
Active learning and uncertainty sampling
Hard-case mining from production traffic
Bias and coverage analysis
Synthetic data — when and how
Teknologier
Annoteringsverktøy (Label Studio eller tilsvarende)
Label Studio / HumanSignal
Active learning
Syntetisk datagenerering
Statistisk dekningsanalyse
Forutsetninger
The shared course Evaluation in practice — you need to be able to run an eval loop before you design datasets.
MCP integrations and custom server design
Builder
Connect agents to enterprise data and actions with the Model Context Protocol.
Wed 3 Feb90 min masterclass + 3 timer lab
The Model Context Protocol (MCP): how to expose internal systems to AI agents in a standardised way. Build your own MCP servers and connect them to clients such as Claude Desktop and your own stack.
What you leave with
Can build an MCP server in Python or TypeScript
Understands the protocol semantics: resources, tools, prompts
Can connect an MCP server to internal APIs with authentication
Knows how to test an MCP server with Claude Desktop and their own tooling
What you learn
The MCP 2026 spec: stateless core, header-based routing, the extensions framework
Resources, tools, prompts — the three primitives
Tasks and MCP Apps: long-running jobs and server-rendered UI
Enterprise-managed authorisation: OAuth/OIDC integration at clients
Security: allowlisting, scoping and the NSA/CISA guidance
Testing and debugging MCP servers
Rollout strategy inside the client organisation
Teknologier
Model Context Protocol (2026-07-28-spec)
MCP-server SDK
Enterprise-Managed Authorization
JSON-RPC
MCP Inspector
Claude Desktop / Cursor som klient
Forutsetninger
Agent orchestration.
Operations and service agreements for AI solutions
Shaper
What the client is actually buying once the solution is delivered
Wed 17 Feb90 min masterclass + case-arbeid
Et AI-prosjekt slutter ikke ved lansering — det går over i forvaltning, og det er der avtalen avgjør om leveransen ble en suksess eller en tvist. Dette kurset handler om alt som må være på plass når løsningen skal leve videre hos kunden: hva en SLA kan love når svaret ikke er deterministisk, hvem som betaler modellkostnaden når forbruket dobles, hvordan en AI-tjeneste kommer gjennom kundens endringsråd, og hvor hendelser havner når noe går galt. Dette er rådgiverjobben som skiller et konsulenthus fra en frilanser.
What you leave with
Kan sette opp en forvaltningsavtale for en AI-løsning, med opsjoner for videreutvikling og et definert eierskap etter overlevering
Kan formulere en SLA som er til å holde når svaret er sannsynlighetsbasert — og forklare kunden hvorfor «99 % riktige svar» ikke er en forsvarlig forpliktelse
Kan gjøre rede for hvem som bærer modellkostnaden i en fastpris- kontra løpende avtale, og hva som skjer kommersielt når forbruket dobles
Kan føre en AI-tjeneste gjennom kundens endringsråd og driftsgodkjenning, og vet hvilke dokumenter som kreves før produksjonssetting
Kan plassere AI-hendelser i kundens eksisterende hendelsesprosess i stedet for i en egen kanal, og vet hva som skal eskaleres og til hvem
What you learn
Forvaltningsavtale: omfang, opsjoner, responstider og hva som er inkludert
SLA for sannsynlighetsbaserte tjenester — tilgjengelighet er lett, kvalitet er vanskelig
Kostnadsmodell etter overlevering: hvem eier token-regningen, og hva skjer ved forbruksvekst
Endringsråd og driftsgodkjenning: veien fra ferdig løsning til produksjon hos kunde
Hendelseshåndtering i kundens prosess — ikke i en egen Teams-kanal
Løpende etterlevelse etter lansering: hva kunden må gjøre videre, og hvem som følger det opp
Overgangen fra prosjektteam til forvaltningsteam: dokumentasjon, kompetanse og bemanning
Teknologier
ServiceNow og tilsvarende ITSM-verktøy
Azure Cost Management
Azure Monitor / Application Insights
Forvaltningsavtale som dokumenttype
Endringsråd (CAB) som prosess
Forutsetninger
Felleskurset om overlevering og driftsovertagelse, og Shaper-kurset om rollout og lansering.
GenAIOps and observability
Alle
See what happens in production and catch the failures before the client does.
Wed 27 Jan90 min masterclass + lab
GenAI Ops as a discipline: logging, tracing, monitoring and alerting for AI services. What every ML team has wrestled with for a decade, but with LLM-specific dials. If you do not code: your focus is understanding the operational lifecycle well enough to ask the right questions at the client — the conceptual track through the lab carries equal weight.
What you leave with
Can set up distributed tracing for an LLM chain
Has alert rules for the typical LLM failure modes
Can analyse token drift and detect model regression
Knows how to respond to production incidents involving an LLM
What you learn
Agent traces as the primary object: runs, steps, tool calls
LLM observability: Langfuse, LangSmith, Arize
OpenTelemetry for AI services (the GenAI conventions)
Online evals: LLM-as-judge on production traffic
Drift detection for LLMs: semantic, not statistical
Cost and latency dashboards
Incident response for LLM failures
Teknologier
Langfuse
LangSmith
Arize
OpenTelemetry (GenAI semconv)
Grafana
LiteLLM (gateway-metrikker)
Forutsetninger
The shared courses LLM Foundations and AI-assisted development. Builders also benefit from Robust AI services in production, but it is NOT a requirement — the course has a conceptual track for Tuner and Shaper.
FinOps for AI services
Alle
Send the right request to the right model. Cut cost without losing quality.
Wed 10 Feb90 min masterclass
Cost management for AI services. How to hold the budget as traffic grows and models get more expensive. If you do not code: cost management is a business discipline — the arithmetic happens in spreadsheets, not code, and Shapers are often the strongest people in the room on this.
What you leave with
Can build a cost dashboard for an AI service
Knows how to set per-user and per-tenant budgets
Can implement prompt caching and batch pricing
Has a strategy for routing between cheap and expensive models
What you learn
The AI gateway as control plane: routing, budgets, fallback
Per-tenant budgets and throttling at the gateway layer
Cost attribution: per feature, per prompt, per tenant
Caching strategies: provider cache, semantic cache and layering
Batch pricing and async workflows — 50% at all three major providers
Model routing and tiering — and thinking tokens as a cost driver
FinOps practice: dashboards, spike alerts, monthly cost review
Teknologier
LiteLLM
Portkey
Langfuse
OpenMeter
Prompt caching
Batch API
Redis / Valkey (semantisk cache)
Forutsetninger
AI economics and FinOps, GenAIOps and observability.
Agent governance in production
Tuner
Identity, oversight and control once the agents go live
Wed 3 Feb90 min masterclass
This course gives you the tools to govern autonomous agents in production: agent identity and access, registries with ownership and lifecycle, oversight mechanisms such as guardian agents, and intervention design that keeps the human meaningfully in the loop.
What you leave with
Can set up an agent registry with owner, autonomy level and tool access per agent
Can design agent identity on least privilege (short-lived, scoped credentials)
Can explain the guardian agent category and when an oversight agent is the right move
Can define intervention thresholds and a revoke path for an agent in production
What you learn
Why classic governance does not cover agents: cascading failure, scope creep, the attribution gap
The agent registry in practice: identity, owner, lifecycle, policy bindings
Agent IAM: short-lived credentials, scoping, least privilege
Guardian agents: reviewers, monitors, protectors
OWASP Agentic Top 10 from an operations perspective
Singapore's Model AI Governance Framework for Agentic AI
Intervention design: thresholds, approval points, kill switch
Teknologier
OWASP Top 10 for Agentic Applications
Singapore MGF for Agentic AI
NIST AI Agent Standards Initiative
Agent registry (Microsoft Agent 365, Microsoft Entra Agent ID)
ISO/IEC 42001
SPIFFE/SPIRE (kjenne til)
AI deployment and infrastructure as code
Builder
From notebook to production — build the infrastructure your AI solution actually lives in.
Wed 17 Feb2 × 3 timer (teori + hands-on lab)
Many AI projects stall in Jupyter notebooks. This course is about what happens afterwards — when the solution has to live in production, handle real users and meet the security, cost-control and uptime requirements Capgemini holds its deliveries to. You learn to provision Azure infrastructure with code rather than clicking through the portal, wire up CI/CD pipelines that build and deploy AI models automatically, and use managed identity and private endpoints so secrets are never exposed. The result: you can go from "it works on my machine" to a repeatable, auditable and scalable deployment process — and argue architecture choices on their merits in front of clients and technical architects.
What you leave with
Provision Azure AI resources (Azure OpenAI, AI Search, App Service) with Bicep or Terraform, and commit the infrastructure to Git
Set up a CI/CD pipeline in GitHub Actions that builds a Docker image, runs tests and deploys to Azure with no manual steps
Use managed identity and private endpoints so API keys and connection strings never sit in code or environment variables
Run a blue-green or canary deployment to roll out new model versions without downtime
Size and defend infrastructure choices (SKU, region, network isolation, cost) in a technical review with a client
What you learn
IaC foundations: Bicep versus Terraform — what they share, where they differ, and when to pick which
Azure AI resources as code: Azure OpenAI, Cognitive Services, AI Search, App Service and Container Apps
CI/CD with GitHub Actions: a build, test and deploy pipeline for AI solutions
Blue-green deployment and feature flags for safe model upgrades
Network isolation: private endpoints, VNet integration and managed identity in practice
Infrastructure review: presenting and defending architecture choices to a technical audience
Teknologier
Bicep
Terraform
GitHub Actions / Azure DevOps
Azure Container Apps / App Service
Private endpoints og managed identity
Forutsetninger
Completed Azure deployment for Norwegian clients. Basic Git and terminal comfort assumed.
Security for AI agents: injection, STRIDE and MAESTRO
Alle
Threat-model an agent solution as if an adversary is trying to break it.
Wed 24 Feb90 min masterclass + threat modeling workshop
Security for AI services: STRIDE-based threat modelling, prompt injection defence, data exfiltration and the model supply chain. If you do not code: STRIDE threat modelling is a structured workshop method, not programming — groups are formed across tracks.
What you leave with
Can threat-model an AI service with STRIDE for the service layer and MAESTRO for the agent layer
Can explain and demonstrate indirect prompt injection against an agent with tools
Can design defence in depth: guardrails, least privilege, human-in-the-loop and sandboxing
Can run a basic red-teaming round against their own agent with established tools
What you learn
Threat modelling: STRIDE for the service layer, MAESTRO for the agent layer
Prompt injection in the agent era: direct, indirect, multimodal — and the lethal trifecta
OWASP Agentic Top 10 (2026) as the backbone, LLM Top 10 as backdrop
Defence in depth: vendor guardrails — and why they are never enough on their own
Agent IAM and least privilege: over-permissioned credentials as the leading root cause
Model and agent supply chain: registries, MCP servers, dependencies
Red-teaming agents: PyRIT, Garak and the NIST AI 100-2 taxonomy
Teknologier
OWASP Top 10 for Agentic Applications
OWASP LLM Top 10
MAESTRO (CSA)
STRIDE
PyRIT · Garak
NIST AI 100-2
Prompt Shields / Bedrock Guardrails / LlamaFirewall
Forutsetninger
The shared course LLM Foundations. Builders also benefit from Robust AI services in production, but it is NOT a requirement — threat modelling is done in groups across tracks.
Agent deployment and harness in production
Builder
From a working agent to an agent that survives production
Wed 20 Jan90 min masterclass
You learn to operate long-running agents: harness architecture, sandboxing of tool calls, checkpointing and resume, and deployment patterns for agents that work for hours without supervision.
What you leave with
Can design an agent harness with sandboxed tool calls
Can implement checkpointing and resume for long-running agents
Can choose the right deployment pattern: short-lived job, service or cloud session
Can set cost and time budgets that stop runaway agents automatically
What you learn
Harness architecture: main loop, tool layer, state and memory
Sandboxing: filesystem and network isolation for tool calls
Long-running agents: checkpointing, resume and idempotent steps
Deployment patterns: batch job, service, background agent, cloud session
Budgets and guardrails: tokens, time, cost — and automatic shutdown
Multi-agent in production: supervisor/worker and handoff (introduction)
Observability for agent runs: traces as the primary object
Teknologier
Claude Agent SDK
LangGraph (checkpointing)
Docker / Container Apps
MCP
OpenTelemetry
Git worktrees / parallelle agenter

Adopt

March–April 20279 courses in total

User feedback and instrumentation
Builder
Build AI systems that learn how they are actually used
Wed 14 Apr90 min masterclass + lab
The Adopt phase is about the AI system actually being used. As a Builder you have to build in feedback mechanisms so we know whether the system helps. Explicit feedback (thumbs, comments) and implicit feedback (retries, abandonment) together give the full picture.
What you leave with
Can instrument an AI system with user feedback flows (thumbs up/down, comments)
Can choose metrics that genuinely correlate with user value
Knows how to connect in-app feedback to evals in production
Can avoid feedback shaming and build trust in the feedback mechanism
What you learn
Explicit versus implicit feedback — plus agent signals: interventions, escalations, aborted runs
OTEL-based tracing with feedback spans
Judge calibration against a human gold set
CI gating on eval score
One shared data layer: failing production cases grow the golden set automatically
Cohort analysis over time
Teknologier
Langfuse
PostHog / Amplitude
Tommel-feedback-mønstre
Sentiment-tagging
Forutsetninger
Assumes Robust AI services in production and GenAIOps and observability.
Change management for AI adoption
Alle
The five steps users must pass through before they actually use the solution.
Wed 10 Mar90 min masterclass + workshop
Change management specifically for AI rollouts. Why people say no to AI, and what actually works.
What you leave with
Can design a change plan for an AI rollout
Knows Prosci ADKAR and Kotter's eight steps
Can handle the three most common resistance patterns
Has a communication template for different audiences
What you learn
The Prosci ADKAR framework — still the backbone
AI fluency: the 4D framework (delegation, description, discernment, diligence)
Human-agent teaming and job architecture: who does what once the agent takes whole tasks
AI-specific resistance: fear for the job, for quality, for control
Communication strategy per persona
Measuring adoption over time
Kotter's eight steps as a reference frame (briefly)
Teknologier
Prosci ADKAR
Kotter 8 step
Stakeholder-mapping
Adopsjons-metrikker
Forutsetninger
The shared business case course and the Discover phase. Shapers gain extra from process redesign, but it is not a requirement.
AI governance: frameworks, registries and control points
Tuner
The frameworks clients ask about, and the control points actually used
Wed 17 Mar90 min masterclass
AI governance at enterprise scale: policies, review processes, model registry, AI audit.
What you leave with
Can set up a model registry with metadata requirements
Knows how to build review gates into an AI delivery
Knows the AI audit requirements from ISO 42001
Can write governance policies
What you learn
ISO 42001 and the NIST AI RMF — and why neither was designed for agents
Singapore's Model AI Governance Framework for Agentic AI: the four dimensions
From model registry to agent registry — the concept (operational practice is covered in its own course)
New risk classes: cascading failure, scope creep, the attribution gap
Review gates and approval workflows
Designing AI policy
Audit trails and reporting
Teknologier
ISO/IEC 42001
NIST AI RMF
Singapore MGF for Agentic AI
Model registry (MLflow, W&B)
Review-workflows
Forutsetninger
EU AI Act risk classification.
Executive communication for AI initiatives
Shaper
Talk AI to leadership without losing them.
Wed 17 Mar90 min masterclass + øvelse
Communicating to senior leadership: how to present AI initiatives in a way a CEO, CFO and CIO actually hear.
What you leave with
Can build a five-slide executive deck that lands in fifteen minutes
Is comfortable with the Minto Pyramid Principle
Can handle difficult questions without backing down
Knows what a CFO cares about versus a CIO
What you learn
The Minto Pyramid Principle
Writing the executive summary
The agent ROI vocabulary for a CFO: containment, cost per resolved task, intervention rate
Data-driven argument
Handling critical questions
Persona-differentiated messaging
Teknologier
Minto Pyramid
Executive briefing-rammeverk
Slide-design
Data storytelling
Forutsetninger
Business case for an AI initiative.
Portfolio and visibility
Alle
Show your work so clients remember you.
Wed 21 Apr90 min masterclass
How to present your own portfolio to the graduation jury and to future clients. Selection, phrasing, narrative.
What you leave with
Has a public portfolio ready for graduation
Can choose the three strongest deliveries and articulate them
Has a demo video, a sanitised case study and an article
Knows how the portfolio is used in a sales conversation
What you learn
The narrative behind a delivery
Sanitising case studies without losing substance
Demo video: script and editing
The groundwork of personal branding
Teknologier
GitHub Pages / personlig portfolio
Loom
Demo-videoredigering
Case study-mal
Forutsetninger
None.
Workshop facilitation for AI transformation
Shaper
Lead a workshop that moves from abstract problem to concrete plan.
Sun 14 MarWorkshop på helgesamling 2 (søndag 14. mars) — åpen bolk for hele kullet
Facilitating strategic AI workshops. From agenda design to psychological safety in the room. Run as an open session for the whole cohort at weekend gathering 2 — facilitating a good workshop matters whatever your track.
What you leave with
Can design a one-day AI strategy workshop
Knows how to create psychological safety in an unfamiliar room
Knows ten core facilitation patterns
Can handle dominant voices and silent rooms
What you learn
Workshop design and timing
Managing energy through the day
Liberating Structures techniques
Activating participants
Synthesising the outcome
Teknologier
Liberating Structures
Miro / FigJam
Workshop-design-mønstre
Action-Reflection-cycle
Forutsetninger
None.
User research and adoption metrics
Alle
Measure whether the solution creates real value, not just logins.
Wed 24 MarWorkshop på helgesamling 2 (søndag 14. mars)
User research for AI products: how to test whether the solution actually helps, not merely whether it works.
What you leave with
Can design a user research study for an AI solution
Knows what to ask about and what not to ask about
Can interpret qualitative data without confirmation bias
Can iterate the solution based on the insight
What you learn
Discussion guides for AI products
Observation protocols and think-aloud sessions with AI
Thematic analysis
Agent metrics: task completion, containment, intervention rate, cost per resolved task
Why seats and WAU measure access rather than adoption — and what to measure instead
The novelty effect quantified: 30-day repeat usage as the litmus test
Human versus agent activity in telemetry: measure them separately
Teknologier
Diskusjonsguider
Tematisk analyse
Dovetail / Notion
User interview-protokoller
Forutsetninger
None.
Handover and operations of AI solutions
Alle
Hand over a solution so the client can run it without you.
Wed 7 Apr90 min masterclass + workshop
Transfer from delivery to client operations. How to build a handover that is actually taken over.
What you leave with
Can put together a handover pack that survives the transfer
Knows what has to be written down and what has to be people
Can lead knowledge transfer to the client's team
Can close an engagement cleanly
What you learn
The seven elements of the handover pack — now with agent additions
Operational ownership: who owns the agent when it fails in production?
Escalation design: confidence- and risk-based intervention points, designed up front
Kill switch, revoke path and rollback of prompt and model versions
Cost budgets per agent in production
Build-operate-transfer as a handover model
Eval drift after handover: judge calibration as an operational task the client takes on
Runbook writing, on-call and escalation agreements
Teknologier
Run-books
On-call setup (PagerDuty / ServiceNow)
Knowledge-transfer-formater
Notion / Confluence
Forutsetninger
None.
The Phoenix Project — a DevOps simulation
Alle
A four-hour simulation: save Parts Unlimited together
Sat 13 Mar09:00–13:00 under helgesamlingen 13.–14. mars
A business simulation based on the book The Phoenix Project. In teams you run the company Parts Unlimited through a series of crises, learning flow, prioritisation and cross-role collaboration in practice — skills that hit AI delivery just as hard as classic operations. Run in person at the March weekend gathering.
What you leave with
Has felt first-hand how bottlenecks and queues wreck delivery flow — and how to remove them
Can apply the DevOps principles (flow, feedback, continuous learning) to their own deliveries
Has practised prioritisation and collaboration under pressure in a cross-functional team
Can connect the simulation's lessons to AI deliveries at a client
What you learn
The Three Ways: flow, feedback and continuous learning
Bottlenecks, WIP and queueing theory in practice
Prioritising when everything is urgent at once
Collaboration across roles: operations, development, business
Retrospective: what we take back to our own projects
Teknologier
The Phoenix Project (simulering)
Kanban
The Three Ways (DevOps)

Graduate

May–June 20273 courses in total

Norwegian sector advisory
Shaper
Sector-specific advisory for banking, energy and the public sector.
Wed 26 May90 min masterclass + sektor-case
Norwegian consulting is shaped by three large sectors: banking and finance (Finanstilsynet, the ICT regulation), energy (NVE, power contingency), and the public sector (DOFFIN, GDPR, the NSM basic principles). Each has its own compliance regime and buying patterns. You pick one sector as your main focus and learn to deliver credible advisory work there.
What you leave with
Understand procurement and supplier patterns in the chosen sector
Navigate the sector-specific regulatory landscape
Build client references and a portfolio aimed at the sector
Lead a discovery workshop using the sector's vocabulary and codes
What you learn
Banking and finance: Finanstilsynet, the ICT regulation, open banking, AML
Energy: NVE, power contingency, OT/IT security, smart grid
Public sector: DOFFIN, GDPR DPIA, the NSM basic principles, shared components
Decision processes and budget cycles per sector
Client references and portfolio strategy
Sector-specific language and buying behaviour
Teknologier
Sektor-spesifikk regulering (Finanstilsynet, NVE, NSM)
DOFFIN
GDPR
Sektor case studies
Forutsetninger
Experience from the discovery phase and change management. Choose your sector before the course starts.
Oral defence and pitch craft
Alle
Pitch your solution in three minutes. Defend it for fifteen.
Wed 5 May90 min masterclass + mock jury
Oral defence of your own delivery before a jury. Jury rehearsal with feedback.
What you leave with
Can present their own delivery in fifteen minutes to a jury
Handles critical questions without backing down
Knows what the jury is looking for
Can give and receive feedback constructively
What you learn
Presentation structure for a jury setting
Handling questions under pressure
Body language and voice
Mock jury rehearsals
Teknologier
Jury-presentasjon-formater
Q&A under press
Demo-fallback-strategier
Forutsetninger
Portfolio and visibility.
Alliance case: the final delivery
Alle
Assemble the whole Alliance year into one client delivery.
Wed 19 May90 min masterclass + kundeøvelse
Closing the Alliance case with the client: from delivery to handover to the client's own staff, and how to strengthen the relationship from there.
What you leave with
Can lead an Alliance handover conversation
Has agreed follow-up or scale-up with the client
Knows the scope of the non-poach agreement
Has planned their own follow-up rhythm after graduation
What you learn
Handover for an Alliance pair
The scale-up conversation
The non-poach agreement in practice
Follow-up rhythm
Teknologier
Handover-pakke
Oppskalering-samtale-mønstre
Non-poach-avtaler
Forutsetninger
Handover, operational transfer and AgentOps.

The program's six formats

An Ignite year moves between six formats. The core is the courses — 42 of them — delivered via masterclass and supplemented by dojo, frontier lunches, lab weekends, and the two large cohort gatherings that open and close the year.

28× / år
Masterclass
Cadence
Every other Wednesday, 15:00–16:30 (90 min). In person in the instructor's city, Teams for the rest
Purpose
The primary delivery of the curriculum. The shared masterclasses are taken by everyone; track masterclasses run on the in-between Wednesdays for their track — each Igniter attends 20–23 of the 28 sessions. The instructor goes deep on a topic, with code/method demos and Q&A. Every masterclass is recorded.
Attendance
Mandatory for the audience (all for shared courses, track-specific for track courses)
10× / år
Dojo
Cadence
One Thursday per month, 15:00–16:30 (90 min, late afternoon)
Purpose
Two parts: first everyone shows last month's assignment to a partner — that is how assignments are verified, with no hand-in and no grading. Then a reflection circle where Igniters share concrete cases from client work. No slides, just work and feedback.
Attendance
Mandatory for all Ignitere
10× / år
Frontier Friday
Cadence
First Friday of the month, lunch 11:00–12:00 (60 min)
Purpose
The cohort's own stage: an Igniter presents something they built or learned — often an assignment taken further — or picks a frontier paper, a new model release or a contested AI debate. 15 min intro + discussion. Presenting here counts as verifying your assignment.
Attendance
Optional for Ignitere, open to all of Capgemini Norway
2× / år
Lab weekend
Cadence
Two weekends per year (Saturday–Sunday): 17–18 October and 13–14 March
Purpose
Intensive build session. Small teams tackle a realistic client scenario without real client pressure — safe practice in production discipline. The example is taken all the way: evals, observability and adoption goals.
Attendance
Mandatory for all Ignitere
2× / år
Cohort gatherings
Cadence
Onboarding in September (4 days), Graduation in June (2 days)
Purpose
The two tentpoles of the program. Onboarding sets the tone: program intro, track confirmation and team formation. Graduation closes the year with deliverable pitches to a client panel, certification ceremony, and Alliance partners invited.
Attendance
Mandatory for all
40× / år
Courses
Cadence
Throughout the year — shared courses plus dedicated track courses per direction
Purpose
The academic core of the program. Every course has curriculum, recordings and exercises, delivered through an accompanying masterclass. The shared courses are taken by everyone; the track courses provide depth in Builder, Tuner or Shaper.
Attendance
Mandatory for the audience (all for shared courses, track-specific for track courses)

Next step

How we deliver

Each course builds systematic capability — from foundational AI mechanics to delivery-ready AI systems in production.

See the courses

The full course catalogue

Responsible AI

LLM Foundations

Provider-independent LLM architecture

AI economics and FinOps

AI readiness assessment methodology

Model selection and reading benchmarks

Delivery quality in AI projects

AI-assisted development in the client's codebase

Presentation skills

The EU AI Act in practice: DPIA and documentation

Evaluation craft: rubrics and LLM-as-judge

Robust AI services in production

Business case for an AI initiative

Process redesign with AI

Context engineering: what the model actually sees

Quality gates in the delivery pipeline

RAG engineering and hybrid search

EU AI Act risk classification

Evaluation in practice

AI rollout and launch

Distillation and small models for cost and latency

Azure deployment for Norwegian clients

Agentic architecture and orchestration

Dataset craft and labeling

MCP integrations and custom server design

Operations and service agreements for AI solutions

GenAIOps and observability

FinOps for AI services

Agent governance in production

AI deployment and infrastructure as code

Security for AI agents: injection, STRIDE and MAESTRO

Agent deployment and harness in production

User feedback and instrumentation

Change management for AI adoption

AI governance: frameworks, registries and control points

Executive communication for AI initiatives

Portfolio and visibility

Workshop facilitation for AI transformation

User research and adoption metrics

Handover and operations of AI solutions

The Phoenix Project — a DevOps simulation

Norwegian sector advisory

Oral defence and pitch craft

Alliance case: the final delivery

The program's six formats

Masterclass

Dojo

Frontier Friday

Lab weekend

Cohort gatherings

Courses

How we deliver