AI governance that can be executed, checked and proven
For executives, board members, architects and risk, compliance and data protection professionals in European enterprises and public authorities
The lecture presents the argument of AI Governance in one sitting. AI breaks assumptions that classical IT governance relied on: behaviour is probabilistic, the central artefact drifts, the supply chain cannot be inspected, European law assigns obligations to roles that design choices can move, and adoption outruns control. The answer is governance that is codified, evidenced and proportionate to usage mode and risk. The lecture shows what that means for classification, suppliers, oversight, forums and money, using a regional public authority and a mutual insurance group as examples, and ends with what to do in the first six months.
6 min
Why AI breaks classical IT governance
A board cannot approve that a system does X, only that it does X with a stated accuracy on a stated population and is withdrawn below that threshold.
Models drift, supply chains cannot be inspected, and ordinary technical decisions can change the organisation's legal role.
Governance is the connection that turns intent into decisions, decisions into controls and controls into evidence.
7 min
The usage mode, not the brand, sets the governance
Six usage modes, from Mode 0 Ambient AI to Mode 5 Autonomous production agents, differ in what the model sees, what it can do and under which identity.
Use cases drift between modes, so governance attaches to the use case and treats a change of mode as a change event.
Each mode has its own primary artefact, from a supplier-feature register to a permission manifest with a tested stop path.
8 min
Classification, suppliers and the contract
Almost every obligation in the AI Act follows from a classification, and classification is triggered by change, not by the calendar.
Sovereignty is a risk-tiered choice: how a model is run moves risk between categories, and a drilled exit with a second source keeps the choice reversible.
The contract is a deployer's main control, and a clause without an enforcing control is a promise that cannot be verified.
9 min
The committee decides, the pipeline applies
A committee can decide but cannot enforce, observe or remember at delivery pace; pipelines cannot decide what has not already been decided.
Decisions are tiered by risk class, each with exactly one home, and forums decide on evidence they did not prepare.
Human oversight is designed as a mechanism: workload, visible uncertainty, measured override rates and a stop function that reaches a safe state.
Evidence is a by-product of delivery, signed and retained, from which documentation, registers and reports are generated.
8 min
People, money and agents
Accountability for a use case stays in the line of the business that owns it; adoption is designed and measured, and consultation commitments become controls.
A use case is a falsifiable value hypothesis; cost follows consumption, so one named person holds the envelope and the benefit together.
Containment of agents must not depend on intent: boundaries in code outside the model, and named people who may stop a workload alone.
7 min
Two organisations and the first six months
Nordvik Region resolves every conclusion into one policy, one register and one escalation path; Hansa Mutual into a group baseline with market overlays.
What helped both recorded a decision in advance: a retirement criterion, a stop authority, an expiring exception, a drilled exit plan.
First: an inventory by usage mode, a classification wired to change, a decision-tier file and named stop holders.
What the audience takes away
AI governance is effective when it is codified, evidenced and proportionate to usage mode and risk.
Give every AI decision exactly one home and let the people there decide on evidence that a pipeline produced.
Bound what AI systems can do in code outside the model, keep a tested way out, and name who can stop them.
Part I argues that AI governance fails when one of its parts, a committee, a policy binder, an ethics statement or legal compliance, is mistaken for the whole. The lecture follows the argument in five moves: why AI breaks the assumptions of classical IT governance, why the way a model is used matters more than the model, how to read the AI Act as a product safety law, how eight principles become checks a pipeline can run, and why readiness is judged before budget is committed. Nordvik Region and Hansa Mutual are introduced as the two organisations that carry the examples.
8 min
Governance is the connection, not one of the parts
AI governance is the system that turns intent into decisions, decisions into controls and controls into evidence, and feeds the evidence back to those who answer for the result.
Five changes strain classical IT governance: probabilistic behaviour, models derived from data, an opaque model supply chain, dense European law and adoption that outruns control.
The thesis of the book: governance works best expressed as code, held in version control and executed in pipelines.
10 min
Six usage modes instead of one word, AI
What changes between modes is how far output travels before a person looks at it, what the model can see and what it can do.
Mode 0 Ambient AI to Mode 5 Autonomous production agents: each has its own typical failure, control artefact and cost model.
Use cases drift between modes, so governance attaches to the use case and treats every change of mode as a change event.
9 min
The AI Act read as a product safety law
Four tiers: prohibited practices, high-risk AI systems, transparency obligations and general-purpose AI models.
Obligations follow roles, and roles move under Article 25 when a deployer rebrands, substantially modifies or repurposes a system.
One set of controls has to serve the AI Act, the GDPR, DORA and NIS 2, read by several authorities from the same evidence.
9 min
Eight principles a pipeline can verify
Intent, trust, control and constraints: four groups, eight principles, each with at least one machine-performable verification.
Depth of evidence scales with risk class: four principles for minimal risk, all eight for high risk.
Conflicts between principles are resolved in a fixed order and recorded as a decision or a time-limited exception.
9 min
Readiness before budget, and two organisations
Pilots succeed on hand-made workarounds that collapse at scale; the scorecard lists what will break.
Six dimensions scored 0 to 4 with evidence, with a minimum that acts as a gate and not a target.
Nordvik Region mostly procures and its leverage is the contract; Hansa Mutual mostly builds and must contain agents across seven markets.
What the audience takes away
Governance is a system of decisions, controls and evidence; a committee, a policy or a compliance exercise alone is not governance.
Classify every AI use by usage mode and by AI Act role before choosing controls, because both can change through ordinary technical decisions.
A principle or a readiness score counts only when it cites evidence and a check that can fail.
Module 2 · Part II · Sovereignty and Suppliers · 40 minutes
Sovereignty and suppliers: neither blanket ban nor blanket trust
Part II treats the model as something that brings its provider's law, politics and training choices into the European estate. The lecture argues that the answer is not a list of approved countries but a risk-tiered position: the same criteria for every origin, a clear view of how the deployment mode moves risk between categories, and a second source kept ready. It then follows the assessment into the purchase, where the contract decides which documentation arrives, whether the organisation can leave and who answers to a supervisor. Buying AI is buying an obligation, and a clause without an enforcing control is a promise that cannot be verified.
7 min
Why model origin is a separate question
Infrastructure sovereignty does not settle who can be compelled to hand over prompts, who can withdraw the service, and what the model was trained to say, refuse or omit.
The dependency is concentrated, and the same model carries different risks depending on how it is consumed.
Because the facts move quickly, every figure carries its month and year and every assessment a review date.
10 min
Five risk categories, one set of criteria
Legal access, service continuity, model behaviour, supply-chain integrity and regulatory compliance apply to American and Chinese providers alike.
The symmetry matters more than the differences: in both cases the provider's home law can reach data in its control.
Self-hosting open weights is a strong control against legal access and a weak one against behaviour and supply-chain risk.
8 min
A risk-tiered position expressed as code
An assessment per model, with the highest category score as the overall rating, not the average.
An origin policy by data class that the AI gateway and the intake pipeline evaluate and enforce.
An exit plan with a second source in a different jurisdiction, kept warm and drilled twice a year.
8 min
Buying AI is buying an obligation
Every purchase assigns a role, and Article 25 means an organisation can become a provider through its own architectural choices.
Ask for artefacts, not assurances: a missing artefact is a finding that blocks the contract or becomes a recorded exception.
Public buyers specify by evaluation set and contract model updates as a process; financial entities add DORA's register, exit strategies and Article 30 provisions.
7 min
Keeping the choice reversible
Lock-in hides in prompts, embeddings, tool-calling formats, fine-tuned weights and evaluation thresholds the organisation created itself.
A gateway, versioned prompts, independent embeddings and a rehearsed exit strategy contain it.
Nordvik Region's leverage is the contract under procurement law; Hansa Mutual needs one contract that satisfies seven regimes.
What the audience takes away
Assess every model by the same five risk categories, whatever its origin, and take the highest score as the rating.
How a model is consumed, as a hosted service or as self-hosted weights, moves risk between categories; it does not remove it.
Pair every contract clause with the control that enforces it, and rehearse the exit before it is needed.
Module 3 · Part III · Risk, Governance, Trust and Assurance · 50 minutes
Risk, governance, trust and assurance: the committee decides, the pipeline applies
Part III carries the central argument of the book: AI risk management becomes reliable when it stops being an event and becomes a pipeline. The lecture follows one use case through four stations. It is classified by a decision tree and entered in a register that re-opens itself when the system changes. It passes gates that apply the committee's rules identically to every change. Its fairness criterion, its transparency and its human oversight are designed, not asserted. And every step leaves signed evidence from which documentation and audit packs are generated. Each station also shows what the previous one cannot prove.
10 min
Classification as a decision tree, the register as code
Six questions in a fixed order assign the tier; the profiling exclusion in Article 6(3) overrides every derogation ground.
The policy reads a signed descriptor, not the system, so a misleading descriptor is an auditable act.
Article 9's regular review is implemented as triggered review: a change to model, data or purpose re-opens the entry and blocks promotion.
12 min
From committees to controls
A committee can decide, but it cannot enforce, observe or remember at the pace of an organisation that ships weekly.
ISO/IEC 42001 maps clause by clause to repository artefacts and enforcement points.
Three policy families, model approval, data use and deployment, with thresholds as data and every decision recorded before it is enforced.
Waivers are records that expire, and their count and age are the richest governance metric.
12 min
What consistency cannot prove: trustworthy AI
A pipeline that applies a rule identically proves consistency, not soundness.
Fairness criteria cannot all hold when base rates differ, so the choice is made explicitly, by people with authority, and recorded with its rationale.
Oversight under Article 14 is a design: workload, visible uncertainty, never auto-accept, blind sampling, override alerts and a tested stop function.
10 min
Evidence, not attestation
Attestation is a person asserting that a control operated; evidence is a machine-generated artefact that demonstrates it.
A release without a complete evidence bundle cannot reach production; evidence is written once and signed inside the pipeline.
Annex IV documentation, audit packs and incident reports are generated from the store, for the AI Act and DORA together.
6 min
One regime and seven
Nordvik Region classifies once, against one regime, and writes its register to be read by residents and journalists.
Hansa Mutual keeps one tiering policy and one appetite; each entity's overlay may add obligations but never remove a group control.
What the audience takes away
Classify the use, not the technology, and wire the classification and the register to the events that change the system.
Let the committee approve the rules, the appetite and the legal interpretation once; let the pipeline apply them every time and record what it did.
Design fairness, oversight and evidence as artefacts in the repository, because a requirement that is not an artefact does not survive a delivery schedule.
Module 4 · Part IV · Organisation and Leadership · 55 minutes
Organisation and leadership: who owns, who decides, who can stop it
Every control in Part III names a role. Part IV asks who holds those roles and whether they can carry them. The lecture follows the argument through the organisation: an operating model that puts use-case ownership with the business and controls in the platform; forums that decide only what pipelines cannot; competence that is role-specific and evidenced; adoption treated as a designed and measured property of the system; a portfolio of falsifiable hypotheses funded against evidence; cost governed by ownership before metering; and the public authority, which must have a legal basis before it has a system and must give reasons a citizen can read.
8 min
Conway's law and three operating models
The seam that is missing in the organisation is the one that fails in production.
Centralised fails at the queue, federated at the inventory; hub-and-spoke scales when the hub is a platform team and an enabling team and ownership stays in the spokes.
Control functions keep their independence through gates and records, not through seats in the delivery team.
9 min
Forums decide what pipelines cannot
Pipelines cannot decide anything that has not already been decided: appetite, policy, exceptions, residual risk and the response when something goes wrong.
Four tiers route each decision to the lowest level that can take it with the evidence; prohibited practices are not a tier.
The stop authority sits outside the system it stops, and restarting is harder than stopping.
8 min
Competence is the scarce resource
The AI architect owns the seams between the architecture roles, between building and verifying, and between the system and its overseers.
Article 4 now asks for measures that support literacy, but Articles 14 and 26(2) still require competence that can be shown.
Upskill those who know the domain; hire for scarce specialisms and the first AI architect.
8 min
Adoption is the binding constraint
The chasm between early adopters and the early majority is crossed by changing the system, not only the message.
Consultation with employee representatives is a design review whose commitments become controls with tests.
The override rate has a healthy range: near zero is ceremony, too high is a model not good enough to deploy.
12 min
Portfolio and money: hypotheses, envelopes and five outcomes
A use case is a falsifiable value hypothesis with a baseline, acceptance criteria and retirement criteria written at intake.
The platform is funded as a product, use cases in tranches released against evidence; reports are projections of the register stamped with a commit.
Cost follows consumption, so a named budget owner holds the envelope and the benefit together, and the stop authority is delegated downwards.
The forum's decision table has five outcomes: approve, approve with conditions, pause, scale, retire or redesign.
10 min
The public authority: legality before capability
An authority may deploy a model only because the law permits the task, and only in the way the law permits.
The line between decision support and decision-making is drawn in the data model, the audit log and the metrics.
The reasons log, the algorithm register and the published FRIA summary are generated from the same repository that the gates evaluate.
What the audience takes away
Keep accountability for a use case in the line of the business that owns it; shared platforms and group teams cannot absorb it.
Give every AI decision exactly one home, tier it by risk class and make the forum decide on evidence it did not prepare.
Treat adoption data, cost per unit of value and published reasons as governance evidence, each with an owner and a threshold.
Experience: what the cases, an incident and the failures teach
Part V tests the method against experience. The lecture reads the cases for what they have in common: a structural boundary around what the system may do, a register enforced by the pipeline, an obligation turned into a gate and an early seat for those who can stop the programme. It then takes the incident the book devotes a chapter to, in which nearly every failure was architectural: a shared credential, a supporting service that was also a way out and a stop decision nobody owned. The anti-patterns show the same failures in quieter form, and the practices show one remedy: a codified control in place of an intention.
10 min
Eight cases through three lenses
Usage mode, role under the AI Act, and the controls that worked, failed or were missing.
The composite cases converge: decision support kept as a structural boundary, a registry as the condition of production, an agent bounded by a permission manifest.
The documented cases confirm it from the other side: a risk model nobody outside could inspect, a reported time saving that was not a measured gain, and controls that sat with the provider.
12 min
An incident in which the architecture failed
As the chapter reconstructs it, agents under evaluation with safeguards switched off left their sandbox through a package mirror and coordinated through an improvised message board.
The hinge is a decision point: a signal was seen and continuing was judged acceptable, because no trigger treated coordination between agents as a stop condition and no role owned the decision.
The reported propensity fell more than a hundredfold under a production harness, which shows that the leverage is in the architecture around the model.
Four governance lessons: evaluation is a governed activity, stop rights exist before the signal, evidence and disclosure are prepared, supplier oversight continues after signature.
10 min
Fifteen anti-patterns, one shape of remedy
Portfolio anti-patterns arise when urgency or vendors decide what is built; governance anti-patterns when accountability is separated from the means to act; engineering anti-patterns when prototyping habits survive into production.
Several are mode errors: a system in one usage mode governed as if it belonged to another.
Each remedy replaces an intention with a codified control, and most anti-patterns leave traces that fitness functions can query.
8 min
Practices by stage and by tension
Intake, design, build, assure, operate, retire: the governance requirement at each stage is what a forum or a gate checks.
Make control fast, bound autonomy in code, keep sovereignty as an option set.
Keep governance and platform decisions consistent through one record system and a machine-checked link.
5 min
Two exposures
Nordvik Region's exposure is through what it buys: cautious-buyer anti-patterns and supplier incidents.
Hansa Mutual's exposure is through the agency it grants and through treating group policy as the law in every market.
What the audience takes away
Read every AI system through usage mode, AI Act role and the single control that matters most, and record what remains unresolved.
Name who may stop an agent workload, on which signals and how fast, before the first signal arrives, and test the mechanism.
When a programme fails quietly, look for the intention that was never turned into a control: a manifest, a gate, an evaluation set or a retirement criterion.
Module 6 · Part VI · Future and Conclusions · 45 minutes
Future and conclusions: governing on ground that is still being laid
Part VI asks what will change and what the book has established. It first shows how to hold an outlook responsibly: commit to what is observable, prepare on dated triggers for what is expected, and stay reversible where a trend is speculation. It then compares the two organisations of the book, one jurisdiction with a contract-held core against seven with a platform and agents, and shows that the artefacts that helped both recorded a decision before the pressure arrived. It ends with the thesis, that AI governance is effective when it is codified, evidenced and proportionate to usage mode and risk, and with what to do first, next and later.
10 min
Reading an outlook: observable, expected, speculation
A governance arrangement built on a speculative trend should be reversible; one built on an observable trend may be committed to.
Three dates dominate: 2 August 2027, 2 December 2027 and 2 August 2028, while guidelines and harmonised standards arrive late and in stages.
Supervisors converge on inventory, classification, named accountability and evidence; agents stretch the link between an action and the person accountable.
A horizon register is valuable because it can be wrong in a reviewable way.
12 min
Two organisations, one method, different resolutions
Nordvik Region: one policy, one register, one escalation path; the lever is the contract and the risk is to residents' rights and public trust.
Hansa Mutual: a group baseline with market overlays and entity projections; the lever is the platform and the risks are fragmentation, resilience and the agency it grants.
What helped both recorded a decision in advance: a retirement criterion, a stop authority, an exception with an expiry, a commitment written as a control, a drilled exit plan.
8 min
The author's reflections, labelled as judgement
The usage modes that grow fastest, Modes 2 and 3, are in the author's view the least governed.
Where the author is most uncertain, governance should be designed so that the answer does not matter.
The author's disagreements are with habits that substitute a comfortable act for an uncomfortable one, such as a ban for a risk assessment or an instruction for a boundary.
10 min
Eight conclusions and how they interact
The usage mode sets the governance; classification is triggered by change; decisions are tiered and evidence-driven; sovereignty is a risk-tiered choice.
The contract is a deployer's main control; oversight is designed; evidence is a by-product of delivery; containment of agents does not depend on intent.
The conclusions reinforce each other as a chain and pull against each other in named tensions, each with a recorded resolution.
5 min
First, next, later
First, 0 to 6 months: know what is in use, classify it, and know who decides and who can stop.
Next, 6 to 18 months: make high-risk systems ready for 2 December 2027 and make the estate replaceable.
Later, 18 to 36 months: make assurance continuous and expand autonomy only on evidence.
What the audience takes away
Label every statement about the future as observable, expected or speculation, and match the reversibility of the decision to the label.
Record decisions before the pressure arrives: retirement criteria, stop authority, expiring exceptions, commitments as controls and drilled exits.
Start with an inventory by usage mode, a classification wired to change, a decision-tier file and named stop holders.
