Shared knowledge · put into practice
What we are
learning.
The network's practice tracks connect hands-on experience with standards, sources and actionable approaches. Use them to improve decisions on governance, security, development and operations.
01Governance
AI becomes an operational capability
Jesper Schmidt · Sønderborg
Copilot Credits are a shared unit of consumption for Microsoft’s AI services. Pre-purchased capacity can be combined with pay-as-you-go, and consumption can be managed through budgets, alerts and limits. Credits should therefore be treated as an operational resource: every use case needs an owner, an environment, a budget and an agreed response to deviations. Microsoft describes the Default environment as suitable for experimentation and lightweight development, not business-critical solutions. Governance consequently spans finance, security, law, product ownership and continuous change management.
- Measure AI consumption and assign a budget, alert and accountable owner to every solution.
- Separate experimentation, development, testing and production in governed environments.
- Bring product, finance, operations, security and legal expertise into shared decisions.
02Development
Vibe coding with a production gate
Daniel Corydon-Petersen · Gladsaxe
Vibe coding can shorten the route from need to prototype, but responsibility for software in operation remains unchanged. GitHub recommends checking function, context, security, maintainability and dependencies, supported by automated tests and human review. A responsible municipal model therefore gives low-risk prototypes an isolated experimentation track, followed by a clear production gate. Otherwise, the maintenance burden and technical debt appear only after the impressive demonstration.
- Give low-risk prototypes a fast but isolated experimentation track.
- Require a repository, system owner, tests, scanning and qualified review before production deployment.
- Plan patching, monitoring, backups and decommissioning from the outset.
03Development
Agentic engineering as a discipline
Kim Stannov Søvsø · Aarhus
A coding agent can inspect a repository, make a plan, change several files and run tests, documentation and review tools. This makes agentic engineering more than code generation. Versioned Agent Skills and repository instructions can make organisational standards repeatable, but people still own the process: acceptance criteria, bounded implementation, automated controls, independent review and approved release.
- Store shared standards as versioned instructions and skills in repositories.
- Give the agent clear acceptance criteria and something against which it can verify its work.
- Keep people accountable for design, review, release, risk and operations.
04Security
Claude Code, security and AI-assisted interpreting
Morten Ingemann Zeiner · Hillerød
Claude Code can access files, commands and external tools. A security assessment must therefore cover permissions and data flows, not only the model. For AI-assisted interpreting in schools and early-years settings, a lawful basis for processing does not come with the product. The authority must document purpose, necessity, proportionality and legal basis; children’s data require especially strict data minimisation and risk controls.
- Combine managed permissions with file-system and network sandboxing.
- Assess the purpose, data and public task in the specific context.
- Clarify the DPIA, data-processing agreement, retention, transfers and human oversight.
05Architecture
A modular public-sector AI platform
Astrid Vestergård Greulich · Copenhagen
A modular platform is not one large chatbot, but a separation between shared capabilities and specific service solutions: identity, integrations, search, model gateway, logging and a governed release path. A use-case register can connect every product with its purpose, owner, data, risk, cost and service level. Open standards, documented data exports and replaceable components reduce vendor lock-in.
- Build reusable, loosely coupled capabilities with clear interfaces.
- Connect every use case with its purpose, owner, data, risk and costs.
- Preserve the ability to change models and suppliers through portability and standards.
06Security
Agent Harness: control outside the model
Parthee Vijayamohan · Kalundborg
Agent Harness moves enforcement out of the model and into a shared, vendor-neutral control framework. The proposal is not an adopted standard; the model must not itself be the security enforcement point. A signed execution manifest binds identity, purpose, data, versions, tools, network access and approvals. Execution is isolated, while access to models, tools, secrets and networks passes through external gateways using deny-by-default and least privilege.
- Place enforcement outside the model and keep it independent of the agent engine.
- Bind each run to a manifest and enforce access through isolation and gateways.
- Retain evidence covering code, builds, policies, tests, approvals and release.
