Show MoD you’ve got AI under control.

The buyer’s question. A programme ships an AI tool that helps a staff officer cross-correlate terrain, meteorological and force-structure data to produce a course-of-action option. Or an imagery-analyst triage system flags satellite images for review. Or an LLM-driven assistant drafts staff-officer prose. The AI shapes the output the human signs. What does the Playbook require?

Defence AI Playbook reference. p.11 frames AI Assisted Operational Planning as “decision support tools to automate elements of operational planning” — not decision-making tools. p.9 lands the canonical phrasing on Object Detection in Satellite Imagery: “the analyst remains responsible for interpreting what is seen to deliver actionable insight to the frontline.” p.10 (Analysis of RF Signals) requires the model’s confidence measure to flow through to the human reviewer. Together these anchor the meaningful-human-in-the-loop obligation.

What counts as compliant. A named human decision-owner per AI tool — staff officer, imagery analyst, signals analyst by role. The model’s confidence / ranking metadata preserved end-to-end so the reviewer is not flying blind. A record of where the AI was relied upon, where it was overridden, and the basis of the override. Operator training on the AI’s known failure modes before the tool is integrated into the workflow. Without an override record, the human-in-the-loop control is unauditable after the fact.

What Secruna ships for decision support. Detection of bespoke ML / NLP planning assistants, Microsoft Copilot and Azure OpenAI workflows used by staff officers, MODCloud-hosted decision-support pipelines, object-detection ML on imagery, and RF signal classifiers. A pre-filled decision-support obligation block in the Defence AI Use Statement. A confidence-preservation check that fires when an AI score is dropped before it reaches the human reviewer’s screen.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_human_oversight filtered to decision-support tools, with the obligation statement preview on each card.

The buyer’s question. A capability uses NLP and graph visualisation to expose hidden relationships across a classified document corpus. Or an ML classifier labels RF waveforms with a confidence score. Or change-detection runs across satellite imagery to flag activity to a Defence Intelligence analyst. The AI’s output shapes an intelligence product. What does the Playbook require the supplier to retain?

Defence AI Playbook reference. p.8 (Intelligent Search & Document Discovery, Public Perception passage) requires “a broad range of potential harms from AI in Defence” to be considered, with the assessment held under “constant review” as data sources and models evolve. p.10 (Analysis of RF Signals) is explicit that the AI emits a confidence measure on its classification — the obligation is to preserve that measure end-to-end so downstream consumers can weight the AI’s output rather than treat it as ground truth.

What counts as compliant. A provenance record per AI classification or extraction that flows into an intelligence product — model name, version, training-data caveats, runtime confidence. Confidence scores preserved end-to-end through the analyst’s workflow. The broad-range-harms assessment kept under constant review and re-run as new classified data sources or model revisions land. Analyst overrides logged with the basis of the override, so the audit trail captures both the AI’s and the human’s contribution.

What Secruna ships for intelligence. Detection of NLP / vector search systems on classified corpora, ML signal classifiers, change-detection pipelines on imagery, and ML-assisted HUMINT triage. A provenance-record schema that snaps onto the intelligence product and preserves the AI’s confidence score. A constant-review schedule that fires the broad-range-harms re-assessment when a model or data source changes.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_audit_trail filtered to intelligence-processing tools, with the provenance-record template available for one-click export.

The buyer’s question. A predictive algorithm forecasts spare-parts failure across a legacy fleet using a decade of historical service records. Or an ML pipeline optimises store holdings against operational demand. Or an autonomous-resupply RAS uses AI navigation to cross unknown terrain. The AI’s output drives a procurement, deployment, or operational decision. What gate does the Playbook impose before the model is allowed to do that?

Defence AI Playbook reference. p.13 (Last-Mile Resupply) is explicit, in capitals: “Test, Evaluation, Verification and Validation (TEVV) and assurance” — the canonical TEVV anchor in the Playbook. p.6 (Spare Parts Failure Prediction) flags the data-quality and accuracy review obligation when ML runs on legacy maintenance data. p.7 (AI at the Edge) sets the expected pipeline shape: model training, compression and testing before deployment to the platform.

What counts as compliant. A TEVV gate run before the model influences a procurement, deployment or operational decision. Documented data-quality limits — particularly acute on legacy platforms with manually recorded service records (p.6) and on edge platforms with constrained compute (p.7). Testing against the deployment environment, not just the training environment — the Playbook is explicit that deployment may differ significantly from training (p.13). A TEVV evidence pack kept alongside the capability so MoD assurance can re-run the gate at any point in the system’s life.

What Secruna ships for logistics. Detection of predictive-maintenance pipelines, autonomous- navigation stacks, and store-optimisation ML in MODCloud and supplier Azure / AWS accounts. A TEVV evidence-pack template aligned to the Playbook’s canonical anchors. A re-run trigger that fires when training data, model version, or deployment environment changes.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_tevv filtered to logistics-and-predictive tools, with the TEVV pack template attached to each capability record.

The buyer’s question. A capability ranks candidates for a recruitment shortlist, scores service personnel for a posting decision, or recommends promotion candidates from workforce data. The Playbook lists Career Management in the AI Now band on p.4 — the tooling is in service today. Defence personnel are stakeholders. What does the supplier owe the affected individual, the chain of command, and MoD assurance?

Defence AI Playbook reference. p.3 (Cdre Singleton introduction) sets the “ambitious, safe and responsible” principle that is the spine of the Playbook. p.8 (Public Perception) is explicit: “we carefully consider a broad range of potential harms from AI in Defence, in order to maintain confidence of all stakeholders, including the public.” Defence personnel are stakeholders by name. The substantive obligation is meaningful human oversight of every personnel-affecting decision, with disclosure to the affected individual.

What counts as compliant. A meaningful human decision-owner on every posting, promotion, or recruitment outcome — AI ranks, shortlists, or flags; the chain of command decides. Disclosure of AI involvement to the affected personnel through the relevant grievance or appeal channel. A documented training-data, model-version, and fitness-for-purpose record (with particular attention to representative coverage of protected characteristics across the Defence workforce). Recurring adverse-impact audits, not just a one-shot deployment review. Where the EU AI Act Annex III §4 (employment / HR) overlaps, the Secruna matcher fires both rules in parallel.

What Secruna ships for personnel. Detection of HR / personnel ML pipelines, recruitment shortlisting tools, career-management analytics, and workforce-planning models that touch Defence personnel data. A disclosure block tailored for the chain-of-command context. A recurring adverse-impact audit schedule attached to the system record.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_human_oversight filtered to personnel systems, with the disclosure block ready for the human-resources policy team to adopt.

The buyer’s question. A cyber capability runs ML threat detection on network telemetry, automates incident response, classifies malware samples, or uses an LLM to triage threat-intelligence feeds. The model’s output drives a control decision — block / allow / quarantine / escalate. A wrong call has operational consequences inside the Defence enterprise. What gate does the Playbook impose?

Defence AI Playbook reference. p.13 lands the TEVV anchor that applies symmetrically here — a cyber-defence model influencing a control decision sits in the same slot as a safety-critical autonomous system. p.7 (AI at the Edge) sets the training, compression, testing pipeline expected before deployment. p.12 (LLMs for Defence — Unknown Risks) flags the information-security, IP, and reputational risks of cloud-hosted AI on cyber-relevant data.

What counts as compliant. A TEVV gate before the model influences a control decision. Testing against adversarial and evasion conditions, not just clean test traffic. The model’s confidence or score preserved on every decision so a cyber analyst can weight the automated recommendation. Cleared-environment data kept inside cleared environments — the Playbook is explicit on the information-security risk of cloud-hosted AI on Defence data (p.12). A re-test cadence that fires when the threat surface or the model changes.

What Secruna ships for cyber. Detection of ML threat-detection platforms, automated incident-response pipelines, malware classifiers, and LLM-augmented threat-intel triage tools. A cyber-specific TEVV evidence-pack template covering the adversarial-test cases the Playbook anchors require. A data-residency check that flags cloud-hosted AI calls against the supplier’s declared cleared-environment boundary.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_tevv filtered to cyber-defence tools, with the adversarial- testing checklist attached to each capability record.

The buyer’s question. A chartered firm uses CoStar AI on a valuation of MoD estate. A tier-2 supplier uses Copilot to draft technical narrative inside an MoD deliverable. An SME partner runs an ML anomaly check on telemetry feeding a prime’s MoD-funded programme. The AI is a commercial product the supplier already owned; the deliverable touches MoD. What inherits down the chain?

Defence AI Playbook reference. p.3 (Cdre Singleton introduction) is explicit that the Defence AI ecosystem extends to industry, non-traditional suppliers, SMEs, and AUKUS partners: “We are building an AI ecosystem to strengthen our relationship with industry, establish new links with non-traditional suppliers and small and medium enterprises”. The Playbook’s substantive obligations propagate down the supplier chain accordingly. p.8 (Public Perception) reinforces that broad-range-harms review applies wherever Defence AI lands, including via supplier deliverables.

What counts as compliant. Register the AI tool against the MoD contract or programme it supports, with artifact-level metadata that lets the prime audit downstream AI involvement. Retain a provenance record — model, version, training-data caveats, runtime confidence — for every AI-assisted deliverable that flows into MoD. Comply with the relevant Playbook obligation pattern of the underlying use case (decision support, intelligence, logistics, personnel, cyber); the Secruna matcher fires the relevant rules in parallel where the supplier-side AI activity matches. Preserve the audit trail through the chain — the prime’s assurance question can land on a sub-supplier’s AI tool at any point.

What Secruna ships for suppliers. MoD-context tagging on any AI tool used on a Defence deliverable (so a CoStar AVM on MoD estate is tagged differently from a CoStar AVM on a private-sector job). A supplier-side provenance-record export designed to satisfy a prime’s assurance question. Inheritance logic that fires the substantive Playbook rules in parallel where the supplier-side AI activity matches.

See this in your dashboard at: /inventory?framework=uk_defence_ai_playbook&category=requires_audit_trail filtered to supplier-context tools, with the MoD-context tag visible per instruction record.