The CIA plans to put AI-powered “coworkers” into its analysts’ workflows in the coming years, Deputy Director Michael Ellis said on April 9 at a Special Competitive Studies Project event on AI and the intelligence community. Nextgov/FCW reported that the tools would sit inside the agency’s analytic platforms and help with basic tasks, with humans still involved, and that the agency recently used AI to generate an intelligence report for the first time in its history.
Over a longer horizon, Ellis said, the agency expects within a decade to treat AI as an “autonomous mission partner,” with officers managing teams of AI agents in a hybrid model.
/ 01What Ellis said
On the limits of the near-term tools, Ellis was direct: “It won’t do the thinking for our analysts.” According to Nextgov, the AI would help draft key judgments, edit for clarity, and check drafts against tradecraft standards. Defense One adds that the coworkers would help with tasks such as triaging information and flagging trends for human review.
Ellis said the agency had more than 300 AI projects last year. He also said the CIA has doubled its technology-related foreign intelligence reporting, which tracks how adversaries such as China use AI and other technologies, and has elevated its Center for Cyber Intelligence to a full mission center. “The battle of cybersecurity will be a battle of artificial intelligence,” he said.
Ellis did not name any AI vendor, but said the CIA “cannot allow the whims of a single company” to constrain its use of AI and that it is diversifying across multiple vendors. Both outlets set that remark against the dispute earlier this year between the Defense Department and Anthropic.
/ 02Why it matters
The intelligence community has used machine learning for years in collection and processing — imagery, signals, translation. What Ellis described is different: AI inside the analytic process itself, helping draft the judgments that go to policymakers. That is the part of the intelligence cycle where errors are most consequential and hardest to spot, because a finished assessment is read as the agency’s considered view.
The plan to have AI check drafts against tradecraft standards is notable. U.S. analytic standards require, among other things, that analysts describe the quality and credibility of their sources, express uncertainty, and distinguish what they know from what they assume. Those are exactly the properties that generative models do not naturally provide. A model can write fluent prose about sourcing; it cannot, by itself, show which report each sentence rests on.
The vendor-diversification point also matters for architecture. An agency that intends to swap or combine models from several companies needs its knowledge, and its record of how conclusions were reached, to live outside any single model.
/ 03What tradecraft asks of AI
For an AI coworker to be useful to an analyst, its contributions have to be checkable at the speed analysts work. That means every factual statement it offers should point to the specific reporting behind it; it should keep sources of different reliability distinct rather than blending them; and it should give the same answer when asked the same question of the same evidence, so that a reviewer’s check means something.
Structured knowledge is a practical way to get there. When incoming reporting is mapped to an ontology — people, organisations, places, events and the relationships between them — with each statement tied to its source document, an AI system can reason over that graph and return conclusions together with the evidence path that produced them. An analyst can follow the path, challenge a link, or see that two sources disagree. Swapping the underlying language model does not erase that record, because it lives in the graph, not in the model.
An AI coworker earns an analyst’s trust the same way a colleague does: by showing its sources.
/ 04What to watch
- How the first AI-generated report was produced. Details on its sourcing, review and use would say a lot about the agency’s standard for AI-drafted products.
- Tradecraft checks. Whether AI review against analytic standards becomes a formal step, and how its findings are recorded.
- Agent supervision. The decade-long vision of officers managing teams of agents will need ways to audit what each agent did and why. Expect that to become a requirement long before the agents are autonomous.
Ellis drew the right line: the coworker should not do the thinking. The real test is whether every sentence it drafts can be walked back to reporting an analyst can grade.
Analytic tradecraft already describes the graph. Key judgments rest on assessments, assessments rest on reports, and each report carries a source, a reliability rating and a date of information. If imagery, signals and open-source reporting are fused into one ontology with those links intact, a drafting agent can propose a judgment only along paths that exist, and a tradecraft check becomes a query: which judgments rest on a single source, which sources conflict, what changed since the last assessment. In our regional-conflict build, six independent sources fused into 720 records, each still tied to the document that produced it. Held in the graph rather than in any model, that record survives a multi-vendor strategy, and the platform is designed to run entirely inside a perimeter.
How we’d model it
Key judgment—restsOn→AssessmentAssessment—derivedFrom→Intelligence reportIntelligence report—hasReliability→Source ratingIntelligence report—conflictsWith→Intelligence report
- CIA plans for 'AI coworkers', deputy director says — Nextgov/FCW nextgov.com ↗
- CIA employees will get AI 'coworkers'—and eventually run teams of AI agents, deputy says — Defense One defenseone.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.