11 hours ago
2
Recent discussion of artificial intelligence in intelligence analysis has consistently framed the technology as a means of accelerating an existing process. The intelligence cycle (collection, processing, analysis, production, dissemination) remains the implicit organizing schema, with AI cast as something used to drive its stages faster, increase resource efficiency, or widen its scope. We find this framing inadequate. It leaves the cycle itself intact, treating it as a sound structure merely in need of added speed, when the more consequential present opportunity is to reconsider the structure altogether. The intelligence cycle is an industrial-era artifact, popularized by Sherman Kent in the immediate post-war period, when information was scarce, expert labor concentrated, and the consumer a narrowly-defined institutional decision-maker. None of these conditions still holds. Recent work by Gartin, Schlickenmaier and by Reed and Szylkiewicz has argued for updating intelligence with agile, information-technology methods, and for shifting delivery toward a services-centric rather than goods-centric model. These arguments address the outdated production cycle, but neither fully anticipates the extent to which AI permits the cycle to be displaced wholesale rather than merely modernized.
As practitioners building intelligence programs in this environment, we observe that the prevailing conversation remains bounded by traditional conceptions of what analytic work is. This paper proposes a different framework, organized around a single assertion: that AI enables a scale and rigor in cognitive information work that were previously unavailable, and that this in turn dissolves several of the assumptions on which the cycle previously depended. The argument rests on a particular architectural premise that analytic reasoning can be captured as a structured data schema rather than compressed into an overly-simplified finished narrative. From this premise follow five ruptures with the traditional cycle:
First, the core value unit of intelligence shifts away from the finished assessment toward a more complex artifact that contains the entire decision architecture by which the assessment was reached.
Second, AI-enabled analysis becomes continuous and ongoing rather than fixed to a single publication date.
Third, analytic accuracy becomes measurable, and therefore improvable, for the first time in history.
Fourth, the relationship between provider and decision-maker narrows, and can be scaled to the needs of the individual consumer rather than to a generic reporting requirement.
And fifth, source handling and judgment collapse into a single operation rather than appearing as separate steps.
Taken together, these constitute a revolutionary rather than evolutionary departure from the manual methods that have governed intelligence analysis for a century, and they open new ground for rigor, accountability, and accuracy in global risk forecasting.
From Finished Product to Assessment Process
Modern intelligence analysis tradecraft treats the finished product as its core deliverable. The product serves both as the vehicle of value to the consumer and as the measure of organisational output, and it is the terminus toward which collection, refinement, and assessment are all directed. This arrangement was never optimal. Reliance on a single artifact collapses a complex analytic process into one compressed object, in which the judgments, and biases, included along the way are flattened into a single deliverable. Tradecraft notes and caveats have occasionally preserved fragments of this reasoning, but the product standing alone has never fully represented the value chain that produced it. The most important steps in the analyst's work, the alternative hypotheses entertained, the source biases weighed, and the contingencies sketched against one another, do not travel with the document, forming a lost layer of metadata that usually remains behind.
Artificial intelligence relieves the scale pressure that heretofore forced this compression. An analyst working with well-designed AI tools can now meaningfully execute and record each of the steps in an intelligence workflow rapidly and at scale, which permits the end product to change, away from the finished product deliverable to a searchable, indexable, and auditable log of expertise that has produced a range of potentially useful data through its work. In this process, the end value unit of analysis becomes the actual analysis itself, rather than an artificial summary of that analysis compressed to finished product size. By deprioritizing the focus on the product as the end goal of all analytic work, more of the valuable decisions and information which informed its creation become accessible to both the analyst and the consumer, which may audit and explore these dynamically to enhance their own understanding.
For this process to compound rather than merely accumulate, the analytic representation of judgments must be persistent and structured. Judgments of this type include the reliability of the source, the credibility of the information it contains, the weight that information should play in contributing to a view of the world, how it might interplay with other events and trends, and so on. These expert judgments are collected as structured data and recorded as they are formed, so that they can later be reviewed against real-world outcomes as those outcomes resolve. This foundation of analysis permits auditability and recursive improvement in judgment, source collection, and analytic framing. AI tools used correctly should permit a thoroughness which enhances judgment rather than eroding it, because they help create a massive record of analytic work that persists and can rapidly be revisited, rather than a sequence of keyhole snapshots of reality which age into irrelevance from the moment they are completed.
Because our representation of analyst judgment is structured across various data points rather than as a single loose narrative, it supports more than retrieval. A sufficiently large corpus of analytic judgments can be used to generate predictive assessments based on prior weighting and modal relationships, to trace multi-path higher-order consequences that human reasoning follows poorly, to identify which forces carry the most systemic weight, and to express conclusions as calibrated, quantified probability rather than verbal estimate.
From Episodic to Continuous Analysis
The intelligence cycle was built around episodic production not because it produced the best analytic results but because scale challenges prevented anything more rigorous. Between products, the analyst's judgment existed only in their head, and even then, was a nebulous and ill-defined thing. Kent’s “Words of Estimative Probability” and Tetlock’s superforecasting projects both pointed toward a need for improved, continuous, and calibrated judgment, but neither could provide a way to operate such a system of rigor continuously at scale. Artificial intelligence changes this arithmetic. A well-trained model handles what human analysts struggle to achieve at scale, ingesting raw data, mapping it to analyst-defined areas of interest, and updating mathematical prediction models. This rapid processing enables the analyst to spend bandwidth on setting the scope of analytic questions, interrogating the quality and biases of sources, and defining the weights and relationships the models will assign to various real-world events. Far from the language of the factory assembly line, the modern discipline of intelligence we espouse more closely resembles the rhythm of a trading desk, where equities analysts mark positions to market continuously, forever adjusting expectations based on a never-ending flow of data.
In this framing, an equities analyst wouldn’t save up all their trading positions to be submitted in one package at the end of the day, and we propose that appropriately tooled intelligence analysts similarly no longer need to wait until a publication date to deliver analytic value. By connecting front-end AI summarization and chat systems to back-end analyst enrichment areas, customers are able to query the latest in analyst judgment on demand, creating an instant feedback loop in which customer queries inform and sharpen ongoing analytic priorities. This serves the analyst as much as the consumer. It removes the obligation to produce filler during quiet periods, and it lets analytic output follow the genuine cadence of a topic rather than an arbitrary calendar.
Measuring and Improving Accuracy
Intelligence consumers hold the analyst accountable not only for a judgment but also for the reasoning by which it was reached. Historically this accountability has been difficult to honor, because much analytic judgment was formed reflexively and poorly recorded. The methods now available for capturing and structuring reasoning make the problem tractable for the first time. Once reasoning is recorded as structure, it can be scored against outcomes as they resolve, using calibration methods such as Brier scoring. The essential property is that each judgment is preserved as it was made and is not revised afterward. That is what keeps the scoring honest: the analyst is measured against the call they actually made, not a version softened by hindsight.
We are deliberate about the strength of this claim. The architecture does not inherently make analysts more accurate. What it makes possible is the measurement of accuracy and the diagnosis of error. When a judgment proves wrong, the structure allows the failure to be traced to a specific weighting or relationship rather than absorbed into an unaccountable whole. It is this decomposability, sustained over time and across many resolved judgments, that creates the conditions for improvement, for the individual analyst and for the models their judgments inform. The data describing how and why an analyst reached a judgment is, in this respect, more valuable than the judgment itself, because it is the raw material of recursive refinement.
This is a meaningful departure. For most of its history, intelligence analysis has struggled to know whether it was improving in delivering decision advantage or predictive insight, because the record needed to properly audit this improvement was never systematically available. For the first time, a complete and inspectable record scored against reality is within reach, presenting the opportunity for true improvements in forecasting accuracy.
From Generic to Specified
A further constraint the cycle never escaped was the assumption that consumers were finite and institutionally legible. The analyst writing for a government agency in 1990 could reasonably picture a handful of senior officials whose interests were bounded by their roles in advancing the national interest. This model functions poorly in the wider modern intelligence context, in which the reader of any given report might vary widely based on their position and access. For intelligence teams working in today’s commercialized contexts, the reader of a report might be a CFO weighing currency exposure, an operations director routing freight around contested waterways, a general counsel mapping sanctions risk, or a fund manager modelling financial tail risk. Each actor is sufficiently distinct from the others that how information is presented to them, and what information is relevant to their decisions, is so different as to destroy the value of a single, universal intelligence report. Each actor makes a different decision against a different geometry of exposure to the same geopolitical environment. A generic product written to the centre of this readership delivers very little decision value to any specific stakeholder because it is intended for none of them.
Bespoke intelligence tailored to individual stakeholders is rare, because it is cost-prohibitive. Examples like the President’s Daily Brief show just how complex and difficult the process is to tailor an intelligence report to even one customer, let alone many hundreds or thousands. Today, AI makes this feasible, because it permits a single body of robust analytic work to be expressed differently for each consumer according to their specific exposure. The assessment surfaced to a Nordic manufacturer with significant Strait of Hormuz exposure differs significantly from the one surfaced to a Latin American agribusiness with none, though both can draw on the same underlying analysis in order to inform a wider geopolitical frame. This approach keeps client-specific context separate from the shared analytic base rather than absorbing it permanently, which matters as much for data governance as for scale. In other words, by keeping intelligence about the threat environment separate from context about the user’s potential impact until the last possible moment, delivery of truly tailored insights is permissible at a scale that humans alone cannot match. Delivering this well still requires human guidance, because the object is to inform human decisions, but it is reachable by a useful number of consumers only through automated composition and delivery. In practice it increasingly resembles data layers, dashboards, and conversational interfaces rather than documents and slide decks, which are inherently static and cannot respond to unique and specific customer interrogation. AI’s ability to handle mass data sets and rapidly synthesize them for human engagement is the key which unlocks these dynamic product offerings.
Source Handling as Judgment
One fiction the cycle's imagery sustained was that a clean separation existed between collection and analysis. In the logic of the assembly line, collection produced sources, processing ordered them, and analysis applied judgment. Practitioners have long known this separation rarely held in practice. Deciding which information to credit, and how heavily, is itself an analytic act, one frequently practiced by collectors but only sporadically preserved in the finished product in the form of sometimes feeble source reliability statements. An AI-enabled team can make this categorization a continuous and systematic piece of the analysis rather than a burdensome and occasional addendum to it. High-volume collection and tagging let analysts reach and index relevant information by reliability far faster, and automated tooling lets them record, in real time, which signals they judge useful, to what degree, and for which questions.
Two disciplines give this its force. The first is continuity: signals attach to persistent, identified subjects rather than floating as unlinked text, so that a judgment made today accrues to the same subject a judgment made months earlier addressed. The second is provenance carried as structure. Each catalogued signal carries its source, the system action that surfaced it, and the analyst decisions that touched it, so that the basis of a judgment travels with the judgment rather than being reconstructed after the fact. In our architecture the analyst encodes meaning into collection from first contact through to the point at which a signal is connected to the wider analytic framework. The system performs the high-volume triage and flagging; the analyst accepts, challenges, or supplies the context the system cannot; and the system then does the durable work of attaching that judgment to analysis where it carries lasting weight. The provenance this produces is more than an audit trail, and becomes part of what the consumer can interrogate. It also forms the basis for learning, over time, about collection gaps and the reliability of sources, serving as an internal collection management architecture.
After the Intelligence Cycle
Building an intelligence team that is AI-native from the outset, at a moment when most established intelligence institutions predate AI and are captured by institutional cultures which inhibit profound change, has shaped our thinking profoundly. The most valuable applications we find for AI push beyond legacy tradecraft, and concentrate on the high-volume work of collection, structuring, and presentation of data. Critically, we do not use AI to replace human judgment. The reason is not that models cannot produce reasoning, because they can, often fluently. It is that a model's account of its own reasoning cannot be relied upon as a faithful record of why it actually reached a conclusion. Auditable, attributable judgment of exactly that kind is what our architecture is built to capture from human analysts. Throughout our experimentation we have found success in a consistent division of labour: the system handles scale, the analyst supplies judgment, and the system records and surfaces that judgment rather than manufacturing it. Attempts to use AI to replace the analytic steps of the cycle risk producing analysis that sounds authoritative but cannot be held to account, and that is most dangerous when it is wrong. Any technology that amplifies human reasoning inherits its errors along with its strengths, which is why the core work of judgment must remain human and auditable.
The process changes we describe are early in their lifecycle, and the work of demonstrating them against a long track record remains ahead of us. Still, the process has taught us that significant changes to the discipline of intelligence analysis are almost certainly on the horizon, particularly as technological advances in model sophistication render traditional information-work delivery obsolete. Human analysts may defend the old ways of conducting analysis on nostalgic grounds, but the truth is that intelligence analysis conducted in this way has a poor track record of success, and disruptions which pose the opportunity for step improvements should be welcomed. These improvements should proceed from the end goal of intelligence analysis - to provide sustainable, responsible, and accurate forecasts about the future that enable decision advantage - rather than from a reactive defense of the previous normal process. To integrate AI in intelligence analysis in responsible ways requires abandoning many of the bad habits and basic assumptions that limited intelligence work in the preceding era. It also requires reconceiving the notion of the value and role of the human analyst in providing insight, and an audacity to believe that what has historically been unknowably complex can be rendered intelligible through sufficiently sophisticated modeling. One hundred years ago, humans struggled to predict the weather with any reliability; today, they expect a device in the palm of their hand to predict rain down to the minute. Similar changes are coming to the world of intelligence analysis. But they will require leaving behind the archaic tools of a previous era in order to reach their full potential.
