When talking about AI in defense, the conversation quickly turns to algorithms. In practice, however, other factors are decisive: data availability, integration capability, release logic, and the question of whether people have good reason to trust the machine.

Thesis: AI integration is less a technology problem than a friction problem.

The points of friction that really slow down programs

Legacy infrastructure and architectural breaks

Military IT landscapes are robust, but have grown historically. AI systems, however, require continuous data flows, secure interfaces, monitoring, and updateability. Where these fundamentals are lacking, media breaks, manual workarounds, and delays arise.

Rule of thumb: If operation and updating are not taken into account, the result is a laboratory product.

Data silos and lack of data responsibility

Data silos arise from structure: organizations, sub-areas, service providers, classification logic. Metadata, provenance, and defined quality criteria are often missing. This makes scaling extremely difficult.

Rule of thumb: Without data ownership, AI will never be reliably reproducible.

Governance and procurement: AI needs lifecycle design

AI ages – due to drift, new threats, new data situations. Governance must therefore map the lifecycle: validation, versioning, approvals, incident response.

Rule of thumb: Governance is not the enemy of speed – it is a prerequisite for repeatable speed.

Trust as a prerequisite for use

Operators must be able to assess when AI helps and when it errs. This requires boundaries, test scenarios, and fallback rules.

Rule of thumb: Trust is not created through explanation, but through rules of use.

Interoperability as an alliance capability

In a multinational context, systems only work in combination. Standards and joint tests must be established early on, otherwise costly special solutions will arise.

The practical framework “Stabilize → Standardize → Scale”

This is where our HSC approach comes into play: We are not a standard consulting firm. We first stabilize execution (leadership, clarity, timing, responsibility) and then make sustainable improvements – lean DNA in complex projects.

Step 1: Define the mission thread

• What decision should AI support?
• Where does the data come from?
• Who uses the results?
• What escalation applies in case of uncertainty?

Step 2: Operationalize data stewardship and data quality

• Owner per source
• Metadata & provenance
• Quality criteria (“usable”)
• Access and classification

Step 3: Trust pack per AI function

Contains:

• Purpose and area of application
• Limitations, error patterns
• Test scenarios (close to deployment)
• Uncertainty logic
• Fallback/override/stop

Step 4: Model Ops as a mandatory component

• Versioning of data/models
• Drift monitoring
• Revalidation cycles
• Incident response process

Step 5: Enforce interoperability early on

• Standards for interfaces/formats
• Joint tests
• Compatible security assumptions

KPIs for rollout readiness

• Time-to-field
• Data readiness
• Override/acceptance + reasons
• Drift/incidents

Risks and trade-offs

• Governance can be paralyzing if it is not designed to be lean.
• Transparency can increase vulnerability.
• Secrecy and data usage must be actively balanced.
• A lack of interoperability leads to massive integration costs later on.

Conclusion

AI in the military becomes effective when it is managed as a capability: with operational logic, data responsibility, trust, and interoperability. Those who make friction visible and stabilize it first scale faster – and more securely.