Security intelligence agencies exist to find the threats that are not obvious. The obvious threats — known subjects on watchlists, entities flagged through formal intelligence channels, individuals with prior adverse histories — are handled through existing screening and alert systems. The harder problem is the threat that is present in the data but not yet individually flagged: the person of interest who has not yet generated a formal record, the network whose members each look innocent in isolation, the pattern of activity that becomes significant only when connected across multiple sources.
Finding that kind of threat in the volume of intelligence, referrals, and open source information that flows through a major security agency requires applying analyst judgment to the right items in the right sequence. That is the prioritisation problem. And it is the problem that most security agencies are not solving at scale.
The referral volume that reaches a counter-terrorism or serious organised crime function on any given day exceeds what the available analysts can individually assess with the depth the material warrants. The response to that volume pressure is triage — conscious or unconscious — that determines which items receive attention and which are queued. The quality of that triage determines the outcome. A critical threat that is queued below lower-priority items and reaches an analyst after the window for action has closed is not a detection failure. It is a prioritisation failure, with the same operational consequence.
The asymmetric cost of prioritisation failure
The investment case for AI-assisted threat triage in border and security intelligence is categorically different from the investment case in almost any other government function. In customs, a missed duty fraud costs revenue. In immigration, a missed overstay costs enforcement resource. In counter-terrorism and serious organised crime intelligence, a missed escalation can cost lives. The asymmetry of that consequence means that the question of whether to invest in better triage cannot be answered through a marginal efficiency calculation. It must be answered through a judgment about whether the current approach is adequate to the risk environment.
The evidence that manual triage of high-volume intelligence flows is inadequate is not primarily an analytical finding. It is visible in the operational record of every major security failure that was preceded by intelligence that was present but not acted on in time. The pattern — threat signal present in the data, not escalated before the window closed — is consistent enough to be structural rather than incidental. It reflects the mathematical impossibility of human analysts maintaining consistent triage quality across volumes that have grown beyond the capacity of the approach.
What AI-assisted threat triage looks like
The triage model does not assess whether a specific intelligence item represents a real threat. That determination requires human judgment and contextual understanding that AI models cannot reliably provide in high-stakes security contexts. What the model provides is a structured assessment of the priority with which analyst judgment should be applied.
The inputs that support triage scoring include the credibility and sourcing of the intelligence item, its consistency with known threat typologies and indicators, the severity implied by the information if accurate, the time sensitivity of any actionable elements, and the relationship of the item to other current intelligence items in the agency’s picture. A counter-terrorism referral scoring model combines these inputs to produce a triage priority that places the item in the analyst queue at the right position — not at the top of every queue, which would defeat the purpose of prioritisation, but at the position its characteristics warrant relative to the other items competing for analyst attention at the same time.
Network and entity link analysis addresses the distributed nature of organised threat networks. A model that maintains an operational entity graph — a continuously updated representation of the relationships between individuals, organisations, and activities that are relevant to the agency’s threat picture — can identify when an incoming intelligence item connects to an existing network of concern, elevating its priority beyond what its individual characteristics would suggest. The connection that makes an otherwise ordinary item significant is invisible without the network context. It is precisely visible with it.
The OSINT dimension
Open source intelligence volumes have grown to a level that no agency can monitor through manual analyst effort. Social media platforms, dark web forums, public financial disclosures, and open-source databases produce signals relevant to security operations at a volume that requires automated collection, filtering, and triage before human analysts see any of it. NLP classification models that assess the relevance and credibility of open source items, and that connect them to the agency’s existing intelligence picture, convert raw OSINT volume into a structured feed of assessed items that analyst capacity can meaningfully engage with.
The governance requirement around AI-assisted OSINT analysis is significant. The evidentiary and civil liberties implications of AI models that process open source information about individuals require careful design of the scope, retention, and auditability of the analysis. This is not an argument against AI-assisted OSINT. It is an argument for building it with the governance architecture that the operational environment requires from the outset.
What success looks like
The metrics are threat prioritisation accuracy — the proportion of escalated items that are confirmed as warranting investigation — analyst caseload per analyst per period, response timeliness for items that are subsequently confirmed as high-priority, and the missed escalation rate on items that were de-prioritised and subsequently identified as significant. The last of these is the hardest to measure and the most important. A programme that tracks only the items it acted on correctly is measuring its successes. A programme that also tracks the items it prioritised incorrectly is measuring its failure rate — which is the information required to improve.
