The AI Budget Is Already Approved
This is Article 5 and the final one in the Connected or Exposed series, which examines why public sector organisations struggle to answer compound operational questions.
Two things are happening simultaneously across public sector.
The first is workforce pipelines under structural pressure. The OECD Employment Outlook 2025 documents the picture across member states – ageing workforces, shrinking talent pools and skills mismatches compounding faster than most organisations are planning for. In the US, the retirement wave accelerated sharply in 2025. Allied nations are scaling defence spending toward wartime levels, with some carrying more than a fifth of military positions unfilled.
The second is AI investment accelerating faster than the underlying data foundations can support. Budgets are committed, pilots are in production and the pressure to deploy is real across every tier of government.
Both trends are real. But are they being built to work together?
The human core
The frontline capacity that public sector operations depend on is specific people with specific skills in specific places: combat medicine, disease surveillance, nuclear safety, emergency management. These are not functions that automate cleanly. Academic literature on infectious disease management is unambiguous: human cross-jurisdictional intelligence during an active outbreak cannot be replicated by automated systems.
The scenarios that run through this series – the naval maintenance backlog that no system could map, the contractor dependency that only became visible when the contractor was gone, the access profile that outlasts the role it was granted for – are failures of workforce visibility, not failures of automation readiness.
AI doesn't change the irreducible human core of high-stakes operational functions. Miss that, and the investment is aimed at the wrong problem.
The force multiplier
A smaller specialist workforce with clear sight into what it has and what's at risk can do more than a larger one operating blind.
Instead of AI replacing the medevac nurse, how about AI helping the operations director answer, in seconds, which nurses are certified, available, and deployable without creating a gap elsewhere?
Or AI helping the agency head model with what breaks if the disease surveillance specialist is unavailable, before an outbreak (not during it).
The compound questions this series has examined — who can surge, what breaks if this person leaves, which contractor dependency cascades, whose access no longer matches their role — are questions that currently take days of manual work across multiple systems to answer badly. They should take seconds to answer well.
Most organisations already have the data that would support this. It exists in HR platforms, training records, clearance databases, procurement systems, scheduling tools. The problem is that it exists in pieces, each system using its own definitions and answering its own question.
A role in the HR system, a position in the scheduling tool, a function in the mission database all describe overlapping concepts with different vocabulary and no shared model of how they relate. Some describe this as a map of how the organisation works. The more precise term is ontology – a shared model that defines what concepts mean and how they relate, so that data from different systems can connect and be queried as one.
A smaller workforce asks more of the data, not less. If fewer people are being required to cover more ground, the systems supporting their decisions have to answer harder questions faster.
What makes it work
The DoD's AI Hierarchy of Needs (published in the Department's 2023 Data, Analytics, and AI Adoption Strategy) is a pyramid with quality data at its base, insightful analytics in the middle, and Responsible AI at the top. The sequence is deliberate and the strategy is explicit that "all analytic and AI capabilities require trusted, high-quality data." That same strategy defines "linked" as one of the required data quality properties – data that allows users to "exploit complementary data elements through innate relationships." Relationships, in other words, are first class citizens. They need to be stored and treated as such.
The foundation is a knowledge layer. Not a dashboard. Not a better HR system. A connected model where people, roles, skills, functions, missions, and the dependencies between them are stored as data and are not siloed across systems with incompatible definitions. A knowledge graph at its core captures what the organisation has. A context graph extends it with what's changing: who's available, what's at risk, what decisions have been made and why. Continuously updated and traversable in real time, this gives AI access to the actual operating picture.
GraphRAG then enables an agent to retrieve a relevant slice of the graph and reason over it in natural language, producing a grounded, traceable response. But the compound operational questions in this series require something more specific: multi-hop reasoning, which means following chains of relationships within the graph across entities that aren't directly linked. Person to role to mission to dependency — not a query against a table, but a traversal that provides that visibility only by following the whole chain. This is the capability that graph databases do best, and it's what makes the compound question answerable at all.
The people who remain spend less time assembling answers and more time acting on them.
The sequence
The AI budget is already approved across the public sector organisations discussed in this series. That's not the decision.
The decision is whether the knowledge layer goes in before the AI or after. Organisations that invert the sequence deploy AI onto fragmented data that doesn't reflect operational reality. They discover the gap when something breaks, and pay to rebuild the foundation they should have built first.
Connected organisations (the ones that have built the knowledge layer before the AI) give the specialists who remain clearer visibility and faster answers. They can manage the compound pressures that fragmented data, and fragmented AI on top of it, cannot.
Connected or exposed. The AI budget is already approved. The question is whether the knowledge layer is going in first or last.