Academy / How How · From capability gap to a team that runs it

How we work

What we offer, to whom, and how we build it. The programmes and formats, the frameworks we teach, the eight-stage method, the teams and technology behind it, and the path to hand-over.

01 · Audiences

Three audiences, each with its own way in

Each audience gets a different format. All of them learn from the same reviewed knowledge, so leaders and teams use one vocabulary.

Leaders

CXO, ExCo and board members

Decide on AI with a shared vocabulary. Briefings use only human-reviewed knowledge, frame the decision, and show the regulatory context.

Executive briefings · steering-committee teach-ins

Practitioners

Architects, engineers and delivery teams

Build and run AI, cloud and data platforms under regulation. Learners prove skills in graded labs, capstones and hackathons on synthetic sandbox data.

Bootcamps · labs · capstones · hackathons · accelerators

Capability owners

L&D, CoE leads and knowledge stewards

Run the factory yourselves. We transfer the Brain, the skills and the gates, and we train your people in the roles that operate them.

Knowledge Steward · Learning Architect · Facilitator

Available soon GovKit Core, then a sector track for government, banking or telco See the courses

02 · Formats

Seven formats, from a short briefing to an accelerator

Each format ends with evidence of learning. We choose the format from the objective, not from habit.

Format	For	Shape	Evidence of learning
Executive briefing	CXO, ExCo, board	45–90 minutes · Wardley map and ArchiMate overview · regulatory context	One decision exercise
Technical bootcamp	Architects, engineers	Micro-lessons of 6–12 minutes, cases and labs, classroom or hybrid	Graded labs and TILT assessments
Hands-on lab	Engineers	Lab environment as infrastructure as code · synthetic data	Deterministic attester grades the result
Capstone	Bootcamp graduates	Challenge brief tied to objectives · milestone gates · mentors	Rubric review and demo day
Hackathon	Cross-functional teams	One to three days · sandbox data · judging rubric	Working solution and demo
Accelerator track	A cohort with a real use case	Cohort over several weeks · mentor plan · milestone gates	Gate reviews toward a pilot
Online micro-learning	Any audience	6–12 minute units on your LMS or LXP	Retrieval quizzes and item analysis

03 · Content

We teach with each industry's own frameworks

Each sector has its own reference frameworks: BIAN for banking, TM Forum for telco, TOGAF and ArchiMate across government and the enterprise. We use them as the language of the course.

Banking

BIAN service domains

Map AI and data use cases to the BIAN service landscape and to the controls a banking regulator expects.

BIANModel risk

Telco

TM Forum ODA and Frameworx

Work with ODA components and their widely used predecessors: eTOM processes, the SID information model and the TAM application map.

ODAeTOMSIDTAM

Enterprise architecture

TOGAF, Well-Architected and ArchiMate 3.2

Describe target states formally. Every architecture exercise uses ArchiMate notation and Well-Architected review questions.

TOGAFWAFArchiMate

Strategy

Wardley mapping and systems thinking

See the landscape before choosing a move. Map value chains, evolution and feedback loops in complex systems.

Wardley mapsSystems thinking

Engineering

Cloud, data and AI operations

DevOps, DataOps, MLOps, LLMOps, AIOps and ScienceOps as standards. Labs in Python, TypeScript, Go and Rust. Documentation as code.

MLOpsLLMOpsDocs as code

Governance

AI regulation and standards

Apply GCC regulator guidance with the EU AI Act, NIST AI RMF and ISO/IEC 42001 as reference frames for controls.

ISO/IEC 42001NIST AI RMF

04 · Method

One value stream, eight stages, five human decisions

Every programme moves through the same stages, from a change in the world to a renewed package. Agents do the production work. People make the decisions. Select a stage to see what happens.

Build the package

Run, measure and renew

stage 8 feeds stage 1 — the loop never closes

1Factory value stream Business process view · eight stages · source: AI Training Factory blueprint

Quality gates before release

Scripts run each gate. A package with a failed gate does not reach a learner.

Gate	Checks	Run by
G1	The knowledge bundle is valid against OKF v0.2 and the training profile	validate_bundle.py
G2	Each objective has its evidence and its activities	alignment_check.py
G3	Learner text follows ASD-STE100 and uses exact framework terms	ste_lint.py + review
G4	Each lab grader passes right answers and fails wrong ones	grader_test.py
G5	No concept in the package is past its review date	stale_report.py

Decisions that stay with people

The factory prepares the inputs and stops. It does not decide for you.

ID	Decision	Decided by
D1	Accept a concept as stable	Knowledge Steward
D2	Approve the learning plan	Learning Architect with client sponsor
D3	Release the package	Learning Architect
D4	Choose the lab data source	Client data owner · default synthetic
D5	Renew or retire a package	Knowledge Steward and Learning Architect

05 · Operating model

Four team types, each with one job

We organise the factory with Team Topologies. Stream-aligned teams own the programmes for one domain. Three other teams make that work fast and safe.

Enabling Learning Method Coaches teams in backward design, cognitive load and assessment

Stream-aligned teams

Banking

Telco

Government

Cloud and AI engineering

Complicated subsystem Lab and Assessment Engineering Builds labs as code, graders and attesters

Platform Factory Platform Brain, skill runtime, CI gates and lab sandbox, offered as a self-service platform

facilitating

collaboration

X-as-a-service

2Team Topologies view Team types and interaction modes · after Skelton and Pais

Roles and what each one decides

People make every release decision. Agents prepare the inputs. The decisions D1 to D5 are set out in the method.

Role	Owns	Decides
Knowledge Steward	The Brain: sources, trust tiers, review dates	D1 · D5
Learning Architect	Design and release of each package	D2 · D3 · D5
Production Engineer	Runs the skills and edits the assets	—
Facilitator	Delivery in class, online and at hackathons	—
Client sponsor	The capability need and the budget	D2
Client data owner	Data classification and lab data rules	D4

How the teams interact

Three interaction modes keep the cognitive load of each team low.

Mode	Between	What happens
X-as-a-service	Platform → stream teams	Teams use the Brain, skills and gates without a ticket.
Facilitating	Learning Method → stream teams	Coaches join a team for a design, then step out.
Collaboration	Lab Engineering ↔ stream teams	They build one lab and its grader together, for a set time.

06 · Technology

An AI Training Factory built on knowledge as code

The Brain is one Open Knowledge Format (OKF) v0.2 bundle in Git. Seventeen agent skills turn its concepts into learning packages. CI runs the gates on every pull request.

Business
roles

Knowledge StewardVerifies concepts

Learning ArchitectOwns design and release

Production EngineerRuns skills, edits assets

FacilitatorClass, online, hackathon

Business
processes

Curate knowledge

Design learning

Produce assets

Assure release

Deliver · measure · renew

Application
components

BrainOKF v0.2 bundle in Git

Skill runtimeFactory agents

Freshness SentinelFinds stale concepts

Lab GraderDeterministic, no LLM

LMS / LXPClient channel

Learning analytics

one source of truth for every package

Data

OKF conceptMarkdown + YAML front matter

Learning packageOne audience, one format

Attestation receiptRuntime evidence only

Technology

Git platform and CIGates on pull requests

LLM platform

Lab sandboxSynthetic data

SkillsSKILL.md files

AttestersGrader code

Lab IaCEnvironments as code

In scope Client context Focal component Notation: ArchiMate 3.2, monochrome

3Layered view ArchiMate 3.2 · simplified for the web · source: ai-training-factory.archimate.xml

Knowledge as code

Every fact has a source and a review date

Concepts are Markdown files with front matter. Each one records its sources, trust tier and the date after which it is stale.

Freshness

Change finds its own impact

The Sentinel compares review dates with today and follows links to every package that uses a stale concept.

Attested labs

Graders that do not guess

Lab graders are deterministic code, not an LLM. They are tested against right and wrong answers before release.

Agent skills

Seventeen skills, one per job

Skills harvest knowledge, design syllabi, write scripts, build labs and check language. A router skill stops at each human decision.

An OKF concept · Illustrative

---
id: bian-service-domain
type: domain-concept
status: stable
sources:
  - BIAN Service Landscape
trust_tier: primary
stale_after: 2027-04-30
verified_by: human:steward-01
---
# BIAN service domain
A service domain is one discrete
business capability. It has one
control record and one owner.

Field names are simplified for this page. The OKF v0.2 training profile defines the exact schema.

07 · Adoption

A pilot proves the factory before it scales

Each client moves through four phases. A gate review closes each phase, so you decide to continue with evidence in hand.

0 Foundation
- Brain repository and OKF training profile
- Core skill set
- CI gates G1 to G5
1 Pilot
- One learning stream
- One bootcamp
- One hackathon
2 Scale
- More learning streams
- Freshness Sentinel runs on a schedule
- Integration with your LMS or LXP
3 Transfer
- Your teams own the Brain
- Factory runs as an internal service
- We coach only

4Adoption roadmap Implementation and migration view · phase lengths set per client

08 · Measures

We measure the gap closing, not hours trained

Each measure comes from data the factory already records. Findings go back into the Brain against the concept they concern.

Measure	What it tells us	Source
Lead time from change to trained staff	How fast a new rule or release reaches people who must apply it	Brain log and cohort dates
Objective attainment	The share of learners who show each objective in the evidence	Assessments, lab receipts, rubrics
Gate pass rate on first run	The quality of production before review	CI results for G1 to G5
Stale concepts in live packages	Content risk. The target is zero.	Freshness Sentinel report
Work owned by client teams	Progress toward hand-over	Git history and release approvals

Work with us

Choose a first programme

Tell us the audience and the capability gap. We propose the format, the first package and the path to run it in-house.

Visit majlisofai.com See the courses