Reinventing Intelligence To Operate At Machine Speed

For almost a decade, the U.S. Intelligence Community (IC) has lived with a quiet anxiety: that the proliferation of open-source information, machine learning, and artificial intelligence would one day reshape its analytic enterprise the way the internet reshaped the newspaper industry. Senior civilian policymakers, military commanders, and national security staff are not passive consumers. They need decision support and have an incentive to build their own if national agencies do not meet their needs. That day is arriving.

The IC’s leadership insists it is seizing the AI moment. At the Special Competitive Studies Project (SCSP) AI+ Intelligence Summit in April 2026, CIA Deputy Director Michael Ellis reported that the Agency ran more than 300 AI projects in 2025 and had, for the first time in its history, used AI to generate an intelligence report. He described plans to embed AI “coworkers” in the Agency’s analytic platforms—tools that will help prepare assessments, edit for clarity, and check drafts against tradecraft standards. AI, in his words, “won’t do the thinking for our analysts, but it will help draft key judgments.”

This is real progress, and it deserves credit. The near-term vision positions AI as a valuable assistant to the human analyst, a tool that drafts, edits, triages, and flags while the existing workforce, workflow, and organizational structure stay largely in place, even as CIA leaders describe mixed human-machine teams over the next decade. This is a sensible starting point, but there is a tremendous opportunity to build on it. Today’s agentic AI systems, which can plan, research, reason, draft, and iterate with minimal human intervention, are increasingly capable of performing core analytic tasks themselves, faster and at a fraction of the cost. That progress opens the door for IC leadership to move beyond the question of whether to adapt and focus on how best to shape that adaptation. By taking deliberate steps now, leaders can redesign workflows and structures on their own terms, keeping the IC’s analytic enterprise ahead of shifting customer expectations rather than reacting to them.

In its groundbreaking 2021 final report, the National Security Commission on Artificial Intelligence envisioned AI supporting human analysts in an AI-ready Intelligence Community by 2025. Subsequent assessments by SCSP, including our 2024 report, Intelligence Innovation, urged enterprise-scale AI adoption, preparing for an AI-ready workforce, and expanding the IC’s use of unclassified, open source information as the “INT of first resort” to provide the fuel for future AI-powered analytic work. While the trend toward better and better AI was clear to us, we assessed that AI systems were still too prone to hallucinations, weak reasoning, and limited access to outside tools and were not yet capable of reliably producing sophisticated intelligence assessments. We could sense the future, but assessed the IC would have plenty of time to prepare and adapt.

What Agentic AI Can Actually Do Today

What has changed in the last two years is the rapid pace of growth in AI capabilities. The frontier models produced by Anthropic, OpenAI, Google, and xAI today are very capable platforms for building autonomous analytic workflows. An orchestrated stack of AI agents can now perform, in parallel and with minimal human prompting: open-source data scanning and collection; source credibility assessments; pattern recognition and anomaly detection; hypothesis generation and structured analytic techniques; drafting, editing, and tradecraft compliance checking; and dissemination tailored to specific customer profiles, formats, and languages—continuously, around the clock, in any theater.

The benchmarks bear this out. Stanford’s 2026 AI Index found that AI capability is now outpacing the benchmarks designed to measure it: frontier models gained 30 percentage points in a single year on Humanity’s Last Exam, an expert-level test built to be hard for AI. Leading models scored under 10 percent when the test was released in early 2025; several models exceeded 60 percent in Stanford’s early 2026 results. ForecastBench, run by the Forecasting Research Institute, compares the accuracy of AI systems in forecasting real-world events with human superforecasters using a difficulty-adjusted Brier score, which rewards being right and being well calibrated. In January 2026, superforecasters still led the best AI entry (xAI’s Grok 4.20, at about 0.102) by 0.017 Brier points, and the Forecasting Research Institute (FRI) projected parity for around November 2026. In its July 2026 update, the institute reported that several systems, including entries from Cassi AI, xAI, and Google DeepMind, were statistically indistinguishable from superforecasters on the tournament leaderboard. OpenAI’s GDPval benchmark measures whether AI can produce real professional deliverables, such as reports, spreadsheets, and presentations, across 44 occupations. Industry experts compare the AI’s work with that of human professionals without knowing which is which. In September 2025, the best model (Claude Opus 4.1) matched or beat the expert 47.6 percent of the time. In its March 2026 results, OpenAI reported rates of 70.9 percent for GPT-5.2 and 83.0 percent for GPT-5.4.

This is already being tested. Dr. Christopher Porter, head of international security cooperation at Google Cloud, has set large language models loose on the National Intelligence Council’s archive of declassified National Intelligence Estimates to dissect how the IC’s most senior analysts built their long-range judgments. Work like this suggests that the tacit knowledge analysts long believed made them irreplaceable may be increasingly capturable and deployable at scale. The moat for human-produced intelligence analysis is narrowing.

Customers Are Already Building Around the IC

Perhaps the greatest long-term threat to the IC analytic enterprise is not replacement from within—it is disintermediation from without. As frontier AI systems become accessible through commercial platforms and enterprise deployments, sophisticated consumers of intelligence are discovering that they can put their questions directly to commercial AI agents, or build their own systems that are tuned to their specific information needs, decision timelines, organizational vocabulary, and risk tolerances.

The appeal is easy to understand. These tools answer in minutes rather than on the 24-hour turnaround the IC offers most senior customers. They tailor every response to the person asking, they require no clearances, and they never close. And they are not bound by the IC’s coordination cycles or institutional risk aversion. During the 2026 Iran campaign, The Washington Post* reported that the Pentagon’s Maven Smart System supported near-real-time targeting. The United States struck more than 1,000 targets in the first 24 hours, according to an analysis by the Carnegie Endowment for International Peace.

The IC’s traditional rejoinder—that commercial tools lack access to classified intelligence and decades of hard-won expertise and tradecraft—describes a moat that is draining, for two reasons. First, the volume of open-source and commercially available data—satellite imagery, ship and aircraft tracking, social media, trade and financial records—is growing exponentially, steadily eroding the marginal value of secret collection on a widening range of questions. SCSP warned in its 2024 report that many open-source assessments from media and think tanks rival what the IC produces using mostly classified information. Commercial AI does not tap the IC’s classified holdings, but on a growing range of questions its assessments may rival the IC’s. Second, as Porter’s work shows, AI systems can be taught the IC’s analytic tradecraft.

The Promise of Agentic AI for Intelligence Analysis

The disruption is real, but so is the opportunity. Properly built and governed, agentic AI offers the IC’s analytic enterprise the chance to deliver more intelligence, to more customers, faster and in greater depth than the Community has ever managed—providing the “decision advantage” the IC must achieve. The goal is not simply speed to an answer. It is speed to defensible judgment: analysis that is timely, sourced, traceable, caveated, and fit for the decision it supports.

Agentic systems expand both the scope and the pace of analytic production. Where today’s analytic cadre must ration its attention across a limited number of priority accounts, AI agents can monitor every country, issue, and actor at once; update assessments as new reporting arrives rather than on a production calendar; provide persistent indications and warning around the clock, as decision support rather than prediction, since no current system reliably predicts geopolitical flashpoints and outputs still need analyst validation; and surge instantly when a crisis breaks—without pulling analysts off other accounts to do it.

They can also end the IC’s chronic triage of “global coverage” accounts. Low-priority and episodic issues—a fragile mid-sized state, a regional resource dispute, an obscure arms transfer—rarely command dedicated analysts until they erupt into crises, at which point the IC starts from a cold baseline. Agentic systems can sustain continuous, credible coverage of these accounts at marginal cost, freeing the Community’s scarce human experts to concentrate on the truly complex, high-stakes, and highly compartmented problems where their judgment matters most. SCSP’s 2025 intelligence memo to the president proposed a related step: a National Open Source Center to take on day-to-day collection for lower-priority global coverage topics.

Agentic AI can also widen access to the IC’s best analysis. An analytic engine is not constrained by the size of a briefing staff. It can give sub-Cabinet decisionmakers—assistant secretaries, combatant commanders, ambassadors, and their staff—access to the best available national intelligence, tailored to their portfolios, decision calendars, and preferred formats. Today, that kind of dedicated support is reserved for the President and a handful of Cabinet principals.

It can also make sharing far less burdensome. Sanitizing an assessment for release is slow, line-by-line work—one reason so little IC analysis reaches audiences beyond the national security community. Agents that generate releasable versions alongside the classified original—with humans reviewing what is released—would let the IC extend timely support at scale to departments and agencies outside the national security community; to state, local, tribal, and territorial governments; to foreign allies and partners; and, where appropriate, to the American people. This extends the “write for release” default that SCSP proposed in Intelligence Innovation.

None of this requires the IC to lower its standards. It lets the IC stop rationing its best analysis.

The Anatomy of an AI Analytic Engine

The purpose of all-source analysis is not to produce reports. It is to reduce uncertainty for decisionmakers, warn them of danger and opportunity, and help frame their choices. Many of today’s products and processes—the daily publication cycle, the fully coordinated finished assessment, the layered review—were designed for a paper-driven era serving a small number of senior readers. Automating them as they stand would simply produce the old outputs faster. The better question is which functions the IC must perform for its customers, and how humans and machines should share them. The layers described below are a near-term engine that can be built from today’s workflows. The destination is more ambitious: a system in which collection, analysis, and delivery happen continuously rather than in sequence, and in which the report itself may no longer be the primary product.

A functional AI Analytic Engine would operate across four integrated layers:

Data Acquisition and Scanning

Automated agents continuously monitor designated data environments—open-source, commercial, and (where authorized) classified holdings—flagging items of relevance against standing collection requirements and emerging triggers. When a human analyst can read only a fraction of an account’s daily reporting, these agents ingest it all—translating, deduplicating, and tagging every cable, report, and foreign-language source against the analytic line before a human opens an inbox. The same agents can generate intelligence requirements automatically and in real time—turning a customer’s query, or an upcoming summit, negotiation, or sanctions decision, into specific collection and analytic tasking within minutes rather than waiting for requirements to work their way through a formal cycle that can take weeks.

Analytic Rigor

A second tier of agents applies structured analytic techniques: key assumption checks, indicator monitoring, alternative hypothesis generation, confidence calibration. These processes can be encoded against established tradecraft standards and applied consistently, without the fatigue that affects human analysts, though models carry biases of their own. Every judgment can be logged with supporting evidence, sources, and a confidence rationale, creating an auditable record of inputs, sources, and judgments. A log does not by itself fully explain how a model reached a judgment, but few human-authored products can match this level of accountability.

Production and Tailoring

Production agents generate, format, and tailor draft assessments to customer specifications—length, classification level, language, visualization preferences, and dissemination channel—using validated templates. The more important change, however, may be to the product itself. A static report, written once and read once, fits poorly in a world that changes by the hour. The engine makes possible continuously updated assessments, dynamic warning displays, and interactive simulations that let a customer test how an adversary might respond to a given move.

Customer Interface and Feedback

A final layer manages the customer relationship: tracking outstanding requirements, soliciting feedback on delivered products, and routing that feedback to refine future tasking. Over time, the engine learns each customer’s priorities, preferred formats, and decision calendar, anticipating questions before they are asked. As customers deploy AI agents of their own, this interface will increasingly run machine-to-machine—a policymaker’s agents querying the IC’s directly—with analysts setting the rules of engagement rather than handling each request.

The architecture is not simple to build. It requires accurate workflow mapping, robust data integration, and carefully constructed tradecraft ontologies—formal representations of what “good analysis” looks like in each analytic domain. The hardest part is neither the models nor the engineering. It is converting the tacit expertise of experienced analysts—how they break down a question, which sources they trust and why, how much evidence is enough, when a judgment is ready—into explicit, documented workflows that AI systems can execute and humans can test. But it is buildable. Several private-sector firms are already building versions of it. One, DreamPort, says its team includes former CIA officials and has built a multi-agent platform based on CIA analytic tradecraft. Others deliver AI-enabled analytic products modeled on the President’s Daily Brief to C-suite decisionmakers; still others are deploying “digital intelligence officers”—tradecraft-based autonomous agents designed to work inside business units.

Rebuilding the Analytic Organization

The critical insight for IC leadership is this: such an engine does not require a large human analytic workforce to operate. It requires a smaller, differently skilled workforce to build, maintain, govern, and improve it. The organizational design implications are profound.

Today’s analytic organization is a pyramid: a broad base of analysts producing drafts, topped by successive layers of reviewers, editors, and managers whose principal function is quality control and coordination. An engine-centered organization flattens that pyramid. It looks less like a production line and more like a software product team—small, cross-functional cells of domain experts, data engineers, and AI engineers, each responsible for a mission area and the agents that serve it.

Because tradecraft standards, style guidance, and classification rules can be encoded and checked automatically, much of the layered human review that now governs production—line editing, formatting, sourcing checks, and basic tradecraft compliance—can fall away. Human reviewers remain, but their attention shifts from polishing individual products to auditing the system: validating its judgments, catching systematic errors, and refining the standards it applies. Fewer layers mean faster delivery. The coordination cycle that today takes days could take hours.

What does this mean for an individual analytic unit? Consider, illustratively, a branch that today covers a regional account with fifteen to twenty analysts, a branch chief, and layers of reviewers and editors above it. In an AI engine-centered model, a team of six to eight might cover the same account. Perhaps half would be substantive experts—senior analysts who own the analytic line, set requirements, and judge the engine’s output. The other half would run the machinery: data engineers who acquire new sources and maintain the pipelines that feed the engine; AI engineers who build, tune, and retrain the team’s agents as the account evolves; and an evaluation specialist who monitors the agents’ performance, tests them against known outcomes, and flags drift, bias, or signs of adversary manipulation. These figures illustrate the direction of change and are not an estimate; agencies should test them, as the recommendations below propose.

The team’s center of gravity shifts accordingly. Fewer people will write; more will build, measure, and steer. And the team’s performance will be measured not by the number of products it publishes, but by the quality, timeliness, and reach of the analysis its engine delivers.

What This Means for the All-Source Analyst

For the individual all-source analyst, the shift is unsettling, but it is not a dead end. Intellectual honesty requires acknowledging what AI systems cannot yet do reliably, and where human expertise retains a genuine comparative advantage. Six roles stand out.

First, analysts bring what no model has: lived experience. Years on an account—the travel, the liaison relationships, the memory of how a leader behaved in the last crisis, the instinct for when a source is too good to be true—add up to tacit knowledge of both substance and tradecraft that is hard to codify and slow to acquire. That knowledge is what lets an analyst recognize when the AI engine is confidently wrong.

Second, analysts will remain the bridge to the most sensitive intelligence, at least for a while. Operational elements within CIA, NSA, DIA, and other collection agencies will resist—in part legitimately—connecting their most sensitive compartmented holdings to agentic AI systems. The security architecture required to do so safely does not yet exist at IC-wide scale, and may not for some time. Until it does, analysts cleared for those streams will hold a genuine—though temporary—moat.

Third, analysts will advocate for the customer. They know what a policymaker is actually trying to decide, what that official already believes, and what will land. They will translate those needs into tasking for the engine and push back when its output answers the wrong question.

Fourth, analysts will become stewards and mentors of the analytic system. Someone must decide what the engine should try to do: which questions it pursues, what evidence standards it must meet, which heuristics it applies, which failure modes it watches for, and when to escalate, caveat, or withhold its outputs. And just as senior analysts train new hires today, they will coach the engine—correcting its reasoning, supplying the context it lacks, and feeding those lessons back into the standards and examples it learns from.

Fifth, analysts will manage teams of AI agents. A single analyst may direct several agents at once—one scanning collection, another running alternative hypotheses, a third drafting tailored products—assigning tasks, reconciling outputs, and spotting when one has gone astray. The skill is closer to running a small team than to writing a paper.

Sixth, analysts will decide what is finished—and answer for it. The engine can generate endless drafts; someone must judge when an assessment is sound enough, sourced well enough, and relevant enough to put in front of a decisionmaker. When a high-profile intelligence failure occurs—as failures inevitably do—a customer who relied on an AI system alone has no one to summon before a congressional committee. This is consistent with the principle of Meaningful Human Control in SCSP’s 2025 report The Digital Case Officer, which calls for a named officer responsible for an AI system’s operations and outcomes.

That accountability matters even more because AI systems fail differently than humans. They can fabricate sources, state weak judgments with unwarranted confidence, and be manipulated by adversaries at every layer—poisoning the data they ingest, compromising the models they run on, and manipulating the prompts, retrieval tools, and agents that tie them together. The danger compounds when agents draw on the same corrupted source and reinforce one another’s errors. Analysts are also prone to automation bias, the tendency to trust machine output too readily, which becomes harder to resist as outputs improve. Analytic integrity can no longer be judged by the finished product alone. It requires visibility into the data, models, prompts, and agent interactions that produced it. Accountability by itself, however, is not a stable value proposition for a workforce of tens of thousands. That is why the other five roles matter.

The analysts who thrive will be those who stop measuring their worth by the products they author and start measuring it by the quality of the judgments they enable. The pen is passing to the machine. The judgment is not.

The Analytic Community’s Strategic Choice

The IC’s analytic enterprise stands at a strategic inflection point that its leadership has not yet fully internalized. Agentic AI systems are not on the horizon. They are in the market. IC customers are adopting them, private firms are building them, and adversary services, including the PRC’s, are pursuing them. The question is not whether the all-source analytic workflow will be substantially automated. It will be. The question is whether the IC will lead that transformation or be overtaken by it.

For the IC’s analytic enterprise, continued tinkering at the margins is no longer sufficient. IC analytic organizations must fundamentally transform how they are designed, staffed, and governed. That transformation must begin with the workforce. Leaders at CIA, DIA, and other IC components that play a significant role in all-source analysis should direct their analytic organizations to pause growth in traditional, production-oriented analyst positions until they complete an honest accounting of the skills the transformed analytic enterprise will actually require. The answer is not more people who can write finished intelligence. It is people who can build, modify, govern, and improve agentic AI systems. The current workforce is not a legacy asset to be managed down; it is the primary source of the expertise the engine must encode. The pause should therefore be paired with an immediate reskilling mandate: every analytic organization should identify a cohort of its best analysts and train them as workflow designers, AI evaluators, and capability owners. And before undertaking major restructuring, agencies should test alternative human-AI team models against traditional teams, measuring not only speed and output but analytic quality, reliability, and usefulness to decisionmakers.

Specifically, the enterprise will need substantially more data engineers who can identify, acquire, clean, and connect the data holdings on which AI agents operate. It will need AI/ML engineers who can build and optimize agent workflows. It will need software designers who can build the interfaces through which customers and internal users interact with the AI Analytic Engine. It will also need a small, highly experienced cadre of domain experts and tradecraft masters to encode analytic standards and evaluate outputs. Conversely, the enterprise will need far fewer of the roles the current production model rewards: line editors, production staff, and generalist drafters. Recruitment, training, promotion, and performance evaluation should reflect that shift now, not after the engine is built—rewarding officers for the quality of the systems they design, supervise, and improve rather than the number of reports they write.

Workforce changes alone are not enough. The IC should accelerate its AI infrastructure buildout, with the necessary security and governance architectures to support it, that empower experienced human analysts to create AI bots that work seamlessly across classified and unclassified data to deliver tailored analytic tools and assessments under human supervision. It should use Other Transaction Authority and shared enterprise licensing to shorten acquisition, as the Defense Intelligence Agency did through Task Force Sabre. DIA Chief AI Officer Maj. Gen. Robert Kinney said at the April 2026 summit that one award took 40 days from request for information to award. It should continually re-evaluate the Intelligence Community Directives that govern all-source analysis to enable AI-assisted and AI-produced analysis while maintaining the highest standards for transparency, interpretability, and analytic rigor and integrity. It should develop specific guidance on implementing Meaningful Human Control, such as a named responsible officer for each assessment, and regular red-teaming against data poisoning and adversary manipulation. It should create career series and a digital reserve for AI talent. And it should invite close foreign intelligence to share this approach.

None of these changes will be easy, or comfortable, for the IC to adopt. They will require IC leadership to make decisions that threaten established career paths, upset organizational equities, and challenge the self-conception of a workforce that takes justified pride in its expertise. But the alternative—a workforce that becomes progressively less relevant to its customers while its leaders maintain the appearance of human primacy—is worse for the mission and ultimately worse for the people in it. The IC’s analytic workforce contains some of the most capable people in the U.S. national security enterprise. They deserve leadership honest enough to tell them what is coming, and decisive enough to build the organizational structures that will let the best of them thrive in what comes next.

All statements of fact, opinion, or analysis expressed are those of the author and do not reflect the official positions or views of the U.S. Government. Nothing in the contents should be construed as asserting or implying U.S. Government authentication of information or endorsement of the author's views.

The Cipher Brief is committed to publishing a range of perspectives on national security issues submitted by deeply experienced national security professionals. Opinions expressed are those of the author and do not represent the views or opinions of The Cipher Brief.

Have a perspective to share based on your experience in the national security field? Send it to Editor@thecipherbrief.com for publication consideration.

Read more expert-driven national security insights, perspective and analysis in The Cipher Brief