The Digital Targeting Web is the UK’s answer to a fundamental question of modern warfare: how do you get from data to decisions faster than your adversary? On 30 April 2026, the Dstl AI Hub convened practitioners from across the defence technology landscape to work on that question — in public, and at pace. This report captures what was discussed, what we observed as missing, and what it means from where we sit: practitioners who bridge emerging capability from in-theatre and battlespace innovation to project delivery, procurement, and the in-service support of major platforms.

This document is both a record of the Dstl AI Hub Open Science & Technology session of 30 April 2026 and a reflective commentary from our own domain perspective. Much of what follows is not simply a transcript of the day — it is our analysis, drawing on direct experience of procurement, acquisition strategy, and through-life in-service support for complex defence programmes. Where we extend beyond what was explicitly discussed, we say so. The thread that runs through our reflection is the gap between where innovation is born — in-theatre, at the operational edge — and where it must eventually land: in supported, sustainment-funded, programme-managed platforms that serve the integrated force for decades.

Background: The Digital Targeting Web

The Digital Targeting Web (DTW) is a strategic programme at the core of the UK’s Integrated Force. Its stated vision is an adaptive, secure and evolving set of digital services that unite users, systems, and platforms to enable targeting accuracy, precision, speed, and resilience across all domains. The DTW Challenge Book 2026, produced in partnership between the National Armaments Director Group and Cyber & Specialist Operations Command, sets out six discrete but interconnected challenges structured around a single overarching goal: getting from data to decisions faster than adversaries.

ChallengeTitle
Challenge 1Securely sourcing and sharing data between dispersed and siloed locations
Challenge 2Combining data sources to produce information and generate insight
Challenge 3Building knowledge from information and making decisions faster
Challenge 4Providing the Digital Targeting Web at the Edge
Challenge 5Upgrade, integrate and interoperate at pace
Challenge 6Providing skills, training and tools

The DTW Challenge Book is explicitly designed as an evolving document — a signal to industry and academia that the problems are live, the requirements are genuine, and the search for solutions is ongoing. Nine guiding principles underpin the challenges, ranging from workflow-driven transformation and data-centric design to open modular architecture and security by design. Critically for the discussions that followed, one principle stands out: decentralised operating model; federated governance — operate as an ecosystem with lightweight central orchestration and shared trust frameworks.

The DTW is not primarily a technology programme. It is a data and decision programme. The technologies — AI, edge computing, federated architectures — are the means. The end is a military force that can sense, understand, decide, act and assess faster than its adversaries. That distinction matters enormously when evaluating which applications of AI will actually deliver operational value.

The Event: Structure and Themes

The session, organised by London RDSC and led by staff from the Dstl AI Hub, drew technical specialists from across a wide range of disciplines: machine learning, AI safety and ethics, communications engineering, data science, decision science, systems security, and integration. The format broke into interactive working group sessions, each addressing one or more of the DTW challenge areas.

Thematic discussions across the day spanned AI in theatre — including image recognition in the battlespace and the integrated battlespace concept — edge computing and local inference, standardisation across public, private and international defence, human-machine teaming within the targeting cycle, and simulation as a route to capability development.

“There was a lot of talk about standardisation across defence — public, private and internationally. But not many people were talking about what the commercial approach would be, or how cross-nation collaboration would actually work.”

The gap that stood out most clearly was the near-absence of honest reflection on commercial frameworks and cross-nation collaboration mechanisms — the contractual, financial and governance structures that determine whether multinational interoperability is achievable in practice, not just in architecture diagrams. This is territory where lived experience from complex programmes offers insight that purely technical discussions tend to overlook. It shaped the contribution we made in the working group.

Technical discussions about AI interoperability are necessary but not sufficient. The same governance failures that have characterised programmes like Eurofighter Typhoon and A400M will reassert themselves in the DTW if the commercial and governance architecture is not designed with the same rigour as the technical one.

Challenge 5 — Our Focus: Upgrade, Integrate and Interoperate at Pace

We engaged with Challenge 5, which asks how the DTW can rapidly integrate new tools and models as they become relevant, maintain interoperability with an evolving set of data sources and partner systems, and continuously improve while preserving safety, legality and system integrity.

The Challenge Book identifies four sub-problems: harmonisation of metadata across data sources; dynamic system assurance; integration of modelling and simulation; and a modular, evolutionary approach to systems development. Our contribution was shaped by a perspective drawn from direct experience of procurement, acquisition, and through-life in-service support — the chain that runs from platform build through to front-line readiness.

The central observation was this: the DTW Challenge Book frames Challenge 5 primarily through the lens of operational systems and tactical data flows. But the same challenge exists — arguably with greater accumulated dysfunction — in the logistics and sustainment layer that underpins every platform the DTW ultimately serves. Fixing the operational data layer without addressing the sustainment data layer leaves a critical gap.

Availability, maintainability and readiness are the metrics by which the armed forces measure their ability to field capable, operational platforms. AI applied to the through-life support chain — not just the kill chain — is where some of the most near-term and measurable value lies.

The ILS Opportunity: Where AI Can Unlock Real Value

This section reflects our own domain perspective as practitioners in defence acquisition and in-service support, not a direct output of the event discussion.

The fragmented data problem

Integrated Logistics Support (ILS) is the discipline that manages the data, systems and processes underpinning the build, maintenance and sustainment of defence platforms — whether helicopters, ships, submarines, combat management systems or weapon systems. ILS data modules capture everything from component specifications and maintenance schedules to failure rates, usage parameters and spare parts requirements.

The challenge, from our experience working across procurement and in-service support, is that these data modules are largely static and siloed. The flow of information between frontline commands, delivery teams within public sector bodies such as DE&S, and industry contractors is fragmented, slow and poorly automated. Bottlenecks exist at every handoff point in the chain. This is not merely an administrative inconvenience — it has direct operational consequences for availability, maintainability and readiness.

The cross-platform opportunity

Work undertaken at Defence Digital and partnering organisations has already demonstrated the scale of the opportunity. By applying data analysis across previously siloed platform datasets, analysts have identified cases where identical components appear across multiple separate platform programmes — managed entirely independently, with separate supply chains and separate procurement processes, despite serving the same ultimate purpose.

Shared Spares Pools

Identical components identified across platforms enable consolidated inventory management, reducing stock-holding costs and improving availability across the fleet simultaneously.

Demand Certainty

Aggregated cross-platform procurement gives industry a consolidated, predictable order book — enabling better capacity planning and more competitive pricing.

Economies of Scale

Volume aggregation from multiple platforms into single procurement exercises delivers cost reductions that no single programme can achieve in isolation.

Supply Chain Resilience

Cross-platform visibility reduces single points of failure, enabling better risk management and faster response to supply disruptions that currently affect multiple programmes simultaneously but invisibly.

The AI Opportunity

AI is ideally positioned to perform cross-platform pattern recognition at scale — across the vast and varied datasets that make up the defence ILS landscape. The component analysis already demonstrated at Defence Digital is a proof of concept, not a ceiling. Applied systematically across the full ILS data estate, this is one of the most concrete, near-term applications of AI that delivers directly measurable operational value: more platforms available, more of the time, at lower cost.

The ILS data layer is not a niche backroom function. It is the information foundation on which platform availability and operational readiness rest. Connecting it, enriching it, and applying AI to it at scale is as strategically important as any element of the operational DTW — and arguably more tractable in the near term.

Multinational Programmes: The Lessons History Offers

This section draws on our direct experience of complex multinational procurement and in-service support, extending beyond what was discussed at the event to offer a practitioner perspective on the governance and commercial realities the day underweighted.

The governance gap in the room

The day’s discussions on standardisation covered technical standards, data governance, and interoperability principles thoroughly. What was largely absent — and what we felt acutely from our experience bridging battlespace capability into procurement and through-life support — was honest reflection on how multinational collaboration actually works: the commercial, contractual and political mechanisms that determine whether it succeeds or fails.

Eurofighter Typhoon: a first-hand account

The Eurofighter Typhoon programme is one of the most instructive case studies available. Designed in the 1980s, built in the 1990s, and entering operational service in 2003, the Typhoon is now operated by nine nations — the UK, Germany, Italy, Spain, Austria, Saudi Arabia, Oman, Kuwait and Qatar — with Turkey set to become the tenth operator. Over 680 aircraft have been ordered, with more than 600 delivered, accumulating over 850,000 flight hours. New builds are contracted through to 2035.

The programme is managed through a consortium of three prime contractors — BAE Systems (UK), Airbus (Germany and Spain), and Leonardo (Italy) — alongside approximately 30 third-tier contractors providing line replacement units and systems-level components. One former supplier contributed approximately 10% of the build, providing missile eject systems, weapons carriage and release mechanisms, brake parachute release units, and chaff and flare dispensing systems.

The complexity of this arrangement creates real operational problems directly relevant to the DTW. In the area of in-service support, situations arise where legacy equipment across the fleet needs replacing or upgrading — but the fragmented consortium structure means no single party can agree on who should fund it. One nation’s air force may have functioning equipment while another is carrying live operational risk as systems degrade. Cases cannot be made. Armed forces are left managing risk with no resolution. Decisions stall for years.

The root cause is a data and governance problem as much as a technical one. The flying hours, missile ejection cycles, maintenance intervals and usage rates that would inform a rational collective decision are highly classified and nation-specific. In the absence of a trusted mechanism for sharing even anonymised or aggregated versions of this data, the consortium cannot act as a consortium.

“The same data that would enable a rational shared decision is precisely the data that cannot be shared. Until that problem is solved at the governance level, no amount of technical interoperability resolves it.”

A400M and the cost of optimism

The A400M Atlas transport aircraft — with fuselage sections from Germany, wings from the UK, transported by Beluga aircraft to final assembly in Seville — was costed on an order volume that was never achieved. As orders were wound back, unit costs escalated significantly, making it one of the more costly European defence programmes of its era. The workshare model, designed to maintain national industrial capability across four countries, created logistical complexity that was justified by the economics of the original order book — and became very hard to justify once that book shrank.

AUKUS: the question to ask now

The AUKUS submarine consortium — Australia, the UK and the United States — is at a formative stage. The governance, data-sharing and commercial frameworks being established now will determine whether it avoids the failure modes of its predecessors. Given the precedents set by Eurofighter and A400M, that question deserves serious attention before problems emerge rather than after.

These multinational programmes are not cautionary tales about technology. They are cautionary tales about governance. The data fragmentation, the cost overruns, the operational risk left unaddressed — these trace to the absence of trusted frameworks for sharing information and making decisions collectively. Building those frameworks must be treated as a first-order design requirement for any AI-enabled multinational programme, including the DTW.

A Common Interoperability Layer: The Middleware Question

Sitting alongside the multinational governance question is a more practical one about how the DTW will actually integrate the diverse data sources, tools, sensors and models it needs to draw on. The framing that surfaced repeatedly during the day was whether something analogous to the integration pattern now emerging around Model Context Protocol could be adapted for defence: a standardised way for AI-enabled services to discover, access and use external data sources and tools. The DTW equivalent would need to be defence-specific, assured, secure, provenance-aware and suitable for contested operational environments.

The most useful analogy offered for this was the USB-C principle. Different tools and systems may have very different internal architectures, and suppliers have legitimate reasons to want to retain freedom to innovate behind those interfaces. But what they expose to the outside world should be a common, agreed output layer, with shared semantics, provenance, access controls and versioning. Behind the interface, vendors and platform owners can do as they wish. In front of it, the DTW and any other consuming service can ingest data and outputs predictably regardless of source. That is the difference between an interoperability ambition and an interoperability standard.

The question of who governs such a standard is not trivial. Internal governance by any single nation, prime contractor or platform operator creates obvious commercial and political problems, and would invite suspicion from the very partners the DTW most needs to bring on board. The most credible model would involve an independent body or trusted authority responsible for the standard itself, with explicit scope to ensure compatibility across nations, suppliers, platforms and data sources. This is recognisable territory for those who have worked through other multinational standards efforts, and it is the layer most likely to determine whether the DTW achieves real interoperability or settles for the appearance of it.

Four Principles for a Defence Interoperability Standard

  • Common output layer (USB-C principle). Different internal architectures are fine — suppliers retain freedom to innovate. What they expose externally must follow agreed schemas, semantics, provenance markers and versioning.
  • Independent governance. The standard cannot be owned by any single nation, prime contractor or platform operator without creating commercial and political problems. An independent body with explicit multinational scope is the credible model.
  • Agreed data schemas and access controls. Shared semantics mean consuming services can ingest data predictably regardless of source — without bespoke integration work for each new supplier or platform.
  • Start now, before fragmentation compounds. The cost of retrofitting compatibility across an already fragmented ecosystem grows quickly. This is a body of work that needs to begin at programme design stage, not as a retrofit.

The practical implication is that the interoperability theme in the DTW Challenge Book deserves to be made considerably more tangible. A common interface model, a middleware layer, agreed data schemas and independent stewardship of the standard together describe a much more concrete proposition than “interoperability” on its own. They also describe a body of work that needs to begin sooner rather than later, before the cost of retrofitting compatibility across an already fragmented ecosystem becomes prohibitive.

Interoperability is not a property of any single platform, model or service. It is a property of the interface between them. Designing that interface, agreeing the schemas behind it, and governing the standard independently is one of the most consequential design decisions the DTW community can take in the near term.

The Depreciation–Appreciation Paradox

One of the most significant conceptual contributions to emerge from the day concerned the differing trajectories of human and machine knowledge — and what they mean for how defence systems should be designed.

Humans appreciate over time.

Experience, tacit knowledge, and accumulated judgement make a seasoned professional in complex programme management — someone who has lived through governance failures, cost overruns, and operational risk decisions across multiple platforms — extraordinarily valuable. That knowledge is hard to codify and irreplaceable if lost. A person who has worked on Eurofighter’s support chain, on nuclear deterrence governance, on A400M’s cost model, understands things that cannot be captured in any requirements document or data schema.

Machines depreciate.

Legacy systems built before the AI age are losing relevance. But even AI models face a form of depreciation — new model generations supersede old ones, and systems trained on today’s data may be poorly suited to tomorrow’s battlespace. The hardware and software infrastructure of today’s most capable AI systems will be obsolete within years, not decades. Defence programmes designed around specific AI tools risk inheriting the same lock-in problems that have dogged legacy platform software for a generation.

The inflection point creates a compounding problem.

Humans are reaching a complexity ceiling — the systems they manage have become too intricate, too multinational, and too data-dense for unaided human cognition. The tools to help them, AI, are themselves on a rapid and uncertain development curve. The capability gap between current ways of working and AI-augmented ways of working is real and widening. And the people who understand current systems most deeply are the ones with the least time to adapt.

The implication for defence programme design is direct. Systems must be designed not just for current human capability, but for iterative human-machine teaming — where AI augments and extends human judgement rather than replacing it, and where the system architecture accommodates the depreciation of both tools and knowledge over time. This means modular design, defined upgrade pathways, and Technology Readiness Level frameworks applied to AI components — not as bureaucratic process but as operational necessity.

It also means investing urgently in capturing tacit knowledge before it depreciates. The people who understand the most complex multinational programmes — the ones who know why the governance failed, where the data gaps are, what the real operational risk is — are a finite and diminishing resource. AI tools that can capture, codify and make accessible their knowledge represent one of the most valuable near-term applications of the technology.

The paradox is not merely academic. A defence AI programme that fails to account for both the appreciation curve of human knowledge and the depreciation curve of machine capability will systematically undervalue its most important assets and overvalue its most transient ones. Designing against that paradox is a governance and architecture challenge as much as a technical one.

Edge Computing: The Underestimated Capability

A recurring theme across the day was the tendency to underestimate local inference on edge computing. The default mental model — particularly among those coming from cloud and enterprise IT backgrounds — positions centralised servers and high bandwidth as the natural home for AI processing. Our reflection, informed by experience at the interface of in-theatre capability and programme-managed platforms, is that this default is operationally dangerous and practically avoidable.

In Denied, Disrupted, Intermittent or Limited (DDIL) environments — which are the norm rather than the exception in contested operations — cloud connectivity cannot be assumed. The insight that emerged most clearly from the day was this: tactical short-term inference at the edge, with carefully governed mechanisms for adaptation where appropriate, does not need to stream back to a central server. An AI system running on an aircraft, a naval vessel, or a forward-deployed unit can perform meaningful inference on the data it has access to — updating its operational picture, prioritising targets, flagging anomalies — without connectivity to the wider DTW.

This is directly relevant to the kill chain: data collection → AI-enabled analysis → decision support → action. If any step in this chain requires cloud connectivity, the chain breaks in a contested environment. If the chain can be executed locally, it becomes genuinely resilient.

The fleet learning loop

The discussion also drew out a second, often-underestimated dimension of edge computing: the logistics of getting operational data off deployed platforms and back into the learning cycle. A useful analogy raised on the day was the practice in parts of the autonomous vehicle industry of recovering high-volume data from vehicles as they return to base, so that the data accumulated across the fleet can be ingested into a central learning environment as quickly as possible.

The important point is not that an individual vehicle waits for its own data to be processed before being redeployed. The point is that the wider fleet creates a continuous learning loop. Each platform contributes operational data back to a central “mothership” environment, where that data can be processed, labelled, analysed, used in retraining, validated, and then released back out as an updated model that the whole fleet can benefit from.

A defence equivalent is straightforward to imagine. Front-line vehicles, deployed platforms, UAVs and forward units will accumulate precisely the kinds of operational data most useful for improving the models they carry. The design goal becomes minimising the delay between operational data capture, data recovery, model retraining, assurance and redeployment of improved capability across the fleet. That is a different design problem to live cloud streaming, and in many contested environments it may be the only viable one.

However, the context in which the data was captured matters enormously. Data gathered in one theatre, against one adversary, under one set of environmental, tactical and sensor conditions, may not transfer cleanly to another. A model update trained heavily on data from one conflict zone could improve performance for platforms operating in similar conditions while degrading performance elsewhere. The learning loop therefore needs to preserve provenance and context: where the data came from, under what conditions it was captured, what threat behaviours it reflected, which sensors generated it, and what assumptions were made during labelling and validation.

1 · Local Inference

Run meaningful AI analysis on the platform itself — targeting, anomaly detection, decision support — with no connectivity to the wider DTW. This is the baseline: operations cannot depend on a network that may not exist.

2 · Fleet Learning Loop

Recover operational data as platforms return to base. Feed it into a central retraining environment. Validate, assure, and push improved models back out across the fleet — so every platform benefits from every other platform’s experience.

3 · Context-Aware Release

Preserve provenance: where the training data came from, under what conditions, against which adversaries. Model updates must be released with full awareness of where they are and aren’t applicable — an update trained in one theatre can degrade performance elsewhere.

The challenge here is not primarily technical. Local inference hardware is increasingly capable and increasingly power-efficient, even within the Size, Weight and Power (SWaP) constraints of forward-deployed platforms. The challenge is conceptual, doctrinal and assurance-led: convincing programme managers, architects and decision-makers to design for edge-first rather than cloud-first, to design the fleet learning loop alongside the inference capability, and to trust locally-run AI systems while still preserving validation, provenance and release controls.

Edge-first is not a compromise. It is the operationally correct design posture for any AI capability intended to function in contested environments. Cloud connectivity should be treated as an enhancement when available, not a dependency. The fleet learning loop - including data recovery, retraining, assurance, provenance and context-aware release management - should be designed in from the outset. This is a doctrinal shift as much as a technical one, and it needs to be made at programme design stage.

Challenge 3 — The Decision Architecture

Alongside the Challenge 5 contribution, a separate Challenge 3 discussion surfaced themes that sit at the heart of the DTW: building knowledge from information and making decisions faster. What follows combines the event discussion with our own practitioner reflection on decision architecture in defence — specifically the challenge of bridging from the operational edge, where decisions are made under pressure, to the programme and acquisition world, where decision frameworks are designed and funded.

The fundamental question for Challenge 3 is how data drawn from UAVs, sensors, intelligence reports, frontline observations, communications and a growing list of other sources is assimilated into usable context - both for strategic command decisions and for tactical decisions on the ground. That is a much harder question than it first appears, because decision thresholds are inherently contextual.

A model confidence score, considered in isolation, is not meaningful. The threshold for action is inherently contextual — the same model output can be a clear basis for action in one setting and an unacceptable risk in another. A decision architecture that does not represent context explicitly will mislead its users sooner or later.

Contextual Factors That Determine the Decision Threshold

  • Local intelligence picture. What else is known about the situation that bears on this model output?
  • Data confidence and provenance. How reliable are the sensors and sources feeding the model? Were they validated for this environment?
  • Operating environment. Threat landscape, terrain, DDIL status, and conditions under which the model was trained.
  • Rules of engagement and legal constraints. What actions are permitted, and under what conditions?
  • Ethical considerations. What are the potential harms of a false positive or false negative in this specific context?
  • Consequences of being wrong. Reversible error or irreversible action? The threshold must shift accordingly.
  • Wider strategic context. Escalation risk, coalition sensitivities, political constraints on the operation.

Human-machine teaming begins before the decision

This connects directly to human-machine teaming. Even where an AI system is not making the final decision, a system that filters, prioritises or frames the information presented to a human decision-maker is still materially shaping the decision process. Explainability, auditability, provenance, uncertainty presentation and user interface design therefore belong inside the decision architecture, not bolted onto it as a presentation layer afterwards. The interface is where context, confidence and consequence are either communicated faithfully to the decision-maker or distorted by accident or by design.

A Decision Intelligence framing

One useful way to frame this is through Decision Intelligence: working backwards from outcomes to data, making explicit at each step which contextual factors matter, who carries responsibility, and what the feedback loop looks like when an outcome is known.

The Decision Intelligence Chain — Working Backwards

  • Outcome. What result are we trying to achieve or avoid? What does success look like, and who is accountable for it?
  • Decision. What specific choices drive that outcome? Who makes them, under what authority, and with what time horizon?
  • Information required. What does the decision-maker need to know — context, confidence levels, provenance, uncertainty ranges — to make a well-grounded choice?
  • Data to capture and govern. What raw data must be collected, validated, labelled and presented to produce that information reliably and traceably?
  • Feedback loop. Once the outcome is known, how does the system learn? How is that learning validated before it changes future decisions?

A context graph, knowledge graph, causal decision model or influence diagram could each support parts of that architecture, but they should not be treated as synonymous with Decision Intelligence itself. They are possible representation layers for making decision context, provenance, assumptions and relationships explicit. The broader discipline is the design and governance of the decision process: what decision is being made, by whom, on what basis, with what uncertainty, under what constraints, and with what feedback loop when the outcome is known.

This is offered as an area for further development rather than a fully formed solution. The real-time assimilation of multiple high-volume data streams into decision-relevant context raises serious open questions around trust, validation, latency, adversarial manipulation, confidence calibration and human usability. None of these are trivial, and several are genuinely unresolved at the level of the wider AI community. But they are closely aligned with the core DTW problem of moving from data, to information, to knowledge, to decision, and they deserve to be treated as a first-class research and design agenda within the programme rather than as a downstream concern.

The user interface is part of the decision architecture

The human factors dimension is particularly important here. If the aim is to support human decision-makers, the way information is presented to them is not a secondary concern. The user interface determines what the human notices, how they interpret uncertainty, how they understand context, and how they make time-critical judgements. That is part of the decision architecture, not a layer of polish on top of it. Designing the UI without reference to the underlying confidence, provenance and contextual factors produces a system that looks decisive and is, in fact, opaque - which is the worst of both worlds for a decision-maker carrying real consequences.

Decision-making in the DTW is not a downstream consumer of the data layer. It is the reason the data layer exists. Designing for context, confidence and consequence - and treating the human-machine interface as a load-bearing element of the decision architecture rather than a presentation concern - is what turns better data into better decisions.

Simulation: The Nuclear Precedent and Its Implications

Perhaps the most striking thread of the day concerned simulation as a route to capability improvement — and the parallel with nuclear weapons development that it invites.

Nations have not conducted atmospheric or underground nuclear tests for decades. Yet nuclear deterrence capabilities have continued to be maintained and, in some respects, improved. This has been achieved through computational simulation — high-fidelity modelling of warhead physics, yield and reliability, run on the most powerful computing systems available. The capability to improve nuclear weapons through simulation, without live testing, is now a well-established and strategically critical discipline. It is also, when understood properly, one of the most striking endorsements of simulation as a serious capability development tool.

The question the day surfaced is whether this model can — and should — be applied more broadly. With AI models now capable of simulating complex environments, and with edge computing enabling local inference at unprecedented speed, there is a genuine and near-term possibility of:

Simulating battlespace environments

Testing and refining tactics, targeting procedures and kill chain processes before fielding — including with systems that do not yet exist physically, as Challenge 5 (System Update 3) explicitly envisions.

Training humans and AI models together

In synthetic environments that accurately represent operational conditions — including contested, degraded and denied environments — without requiring access to live systems or exposing operational parameters to adversaries.

Virtually validating new capabilities

Running new weapon systems and effectors through the DTW architecture before physical deployment, allowing virtual testing at greater fidelity without revealing the operating envelope to adversaries — precisely the approach the nuclear community has used for decades.

The AI models emerging today, and their successors, represent a step-change in the fidelity and accessibility of simulation. The defence community needs to move quickly to build the governance, data and infrastructure frameworks that allow this capability to be exploited responsibly — and to treat investment in simulation environments as a strategic capability rather than a training support tool.

The nuclear community established, under arguably the most demanding constraints imaginable, that simulation can replace live testing without degrading capability. That precedent deserves to be taken seriously across the wider defence AI landscape. The question is not whether simulation can do this — it is whether the institutions and programmes responsible for defence AI development are willing to design around it.

Recommendations

Drawing together the threads from the day and the analytical commentary above, the following observations are offered as a practical agenda.

01 · Scale the ILS cross-platform analysis that Defence Digital has already started.

The component-level pattern matching demonstrated at Defence Digital is a proof of concept, not a destination. AI-enabled cross-platform ILS analysis should be treated as a programme in its own right, with dedicated resource and a clear line of sight to availability and readiness metrics.

02 · Treat commercial and governance architecture as a first-order design requirement.

Technical interoperability discussions must be matched by equal rigour on the commercial frameworks, contractual mechanisms and governance structures that enable multinational data sharing. For every DTW capability with a multinational dimension, the governance model should be designed alongside the technical model — not after it.

03 · Design a defence data trust framework for multinational AI collaboration.

Federated learning architectures, common benchmarking standards, and agreed units of measure could enable multinational AI collaboration without requiring nations to expose classified data. This framework should be designed now — before programmes like AUKUS reach the stage where its absence becomes a critical constraint.

04 · Design for depreciation: modular AI with defined upgrade pathways.

Both human knowledge and AI systems are on depreciation curves. Defence AI architecture must account for this through modular design, defined Technology Readiness Level frameworks applied to AI components, and procurement strategies that avoid locking capability to specific model generations.

05 · Shift to edge-first design, and design the fleet learning loop alongside it.

DTW architecture and AI deployment strategies should be designed for edge-first operation, with cloud connectivity treated as an enhancement rather than a dependency. Equally, the logistics of data recovery, model retraining, assurance and release management should be designed alongside local inference, not bolted on afterwards. Model updates should preserve provenance and operational context so that data captured in one theatre, against one adversary or under one set of conditions is not assumed to generalise automatically across the fleet.

06 · Treat simulation environments as a strategic capability investment.

The precedent of nuclear deterrence maintained through simulation should inform how defence AI is used to rehearse, test and refine conventional capabilities. Investment in high-fidelity simulation environments — and in the AI models that power them — should be elevated to a strategic priority, not treated as a training support function.

07 · Capture tacit knowledge before it depreciates.

The people who understand the most complex multinational programmes — who know why the governance failed, where the real data gaps are, what the operational risk actually is — are a finite and diminishing resource. AI tools designed to capture, codify and make accessible their knowledge represent one of the most valuable near-term applications of the technology in the defence context.

08 · Establish a common interoperability layer with independent stewardship.

A defence-specific middleware pattern, analogous to the integration approach emerging around Model Context Protocol, could provide a tractable interoperability mechanism for the DTW by exposing common agreed output interfaces from heterogeneous tools, sensors and platforms. The standard itself should be governed by an independent body or trusted authority, with explicit scope to ensure compatibility, shared semantics, provenance, access controls and versioning across nations, suppliers, platforms and data sources. Beginning that work now is materially cheaper than retrofitting compatibility later.

09 · Treat the decision architecture as a first-class design concern under Challenge 3.

Decision thresholds are contextual, and the user interface, explainability mechanisms, provenance trails and uncertainty representation are all part of the decision architecture rather than optional decoration. A Decision Intelligence approach, potentially supported by context graphs, causal decision models or influence diagrams, provides a structured way to work backwards from outcomes, through decisions, to the information and data required, and to design the supporting data and presentation layers accordingly.

None of these recommendations requires waiting for technology that does not yet exist. The tools, the evidence, and in most cases the precedents are already available. What is required is the organisational willingness to treat AI integration in defence not as a procurement decision but as a systemic transformation — one that touches governance, commercial frameworks, workforce design, and doctrine as much as it touches technology.

Conclusion

The DTW is a data and decision programme. Treat it like one.

The Dstl AI Hub session on 30 April 2026 was a valuable and substantive event. This report is not simply a record of what was said — it is a reflection from practitioners who work at the intersection the DTW must ultimately navigate: between where new capability is born, in theatre and at the operational edge, and where it must land, in programme-managed, procurement-funded, sustainment-supported platforms that serve the integrated force for decades.

What we observed — and what our own experience reinforces — is that the most significant barriers to realising the DTW vision are not primarily technical. They are data governance, commercial frameworks, multinational trust, and the design of decision architectures that actually work under operational conditions. AI does not dissolve these barriers. But applied with the right governance, it creates new tools for working around them and new urgency for addressing them directly. That is the bridge this report is intended to support.

Prepared by Jareth Reeves and Rob Butler following attendance at the Dstl AI Hub Open Science & Technology session, 30 April 2026. This document reflects the personal observations and analytical perspective of the authors and does not represent the official position of any government department or organisation.