AI in Military Command and Control: CJADC2, Maven, and the Decision Advantage Race

The premise of modern military AI is not autonomous weapons. It is decision speed. If two forces have comparable sensors and weapons, the one that converts observation into action faster wins engagements — and the bottleneck in that conversion is no longer sensing or shooting. It is the human and organizational process in between.

Combined Joint All-Domain Command and Control is the Department of Defense's concept for connecting sensors, weapons, and decision-makers across air, land, sea, space, and cyberspace, and moving data seamlessly across services and coalition partners. AI is the mechanism intended to make that data usable at the speed the concept requires.

The scale of commitment is now visible in budget documents. The Pentagon's fiscal 2027 budget materials include a request for more than $2 billion for command-and-control technology licenses and engineering support, with more than $1.5 billion of that to expand access to Palantir's Maven Smart System in support of the Joint Force AI-Enabled Headquarters initiative.

What CJADC2 actually is

CJADC2 is frequently described as a network or a system. It is neither. It is a set of data standards, interfaces, and architectural principles intended to allow disparate systems to interoperate.

The problem it addresses is real and long-standing. A U.S. Navy destroyer, an Air Force fighter, an Army fires battery, and a Space Force sensor were procured separately, use different data formats, communicate over different networks, and were never designed to work together at machine speed. Coordinating them historically required humans translating between systems by voice and manual entry — a process measured in minutes to hours.

Against a target that moves, or a threat that arrives in seconds, that timeline fails.

CJADC2 aims to establish common data standards so any sensor can pass data to any shooter, a resilient network fabric that functions when degraded, and a decision layer that presents fused information and recommends courses of action.

The first two are engineering problems of considerable difficulty. The third is where AI enters, and where both the promise and the risk concentrate.

Maven Smart System

Maven Smart System has become the most visible instantiation of AI-enabled command and control in the U.S. military. It provides a graphical interface for intelligence and targeting alongside a growing set of CJADC2 functions.

The adoption trajectory is unusual for a defense software program. As of March 2026, Maven Smart System had over 20,000 active users across 35 military service and combatant command tools operating across three security classification domains — a user base that had doubled since January 2026.

The contract ceiling has expanded accordingly, from an initial $480 million to nearly $1.3 billion through 2029, with the FY2027 request adding substantially more.

In April 2026 a directive from the Pentagon's acquisition leadership established AI-enabled decision-making as the cornerstone of CJADC2, formalizing what had been an emergent trend.

Two observations about this trajectory are worth making.

First, the adoption pattern — organic growth driven by user demand rather than mandated fielding — is genuinely unusual in defense software and suggests the tool addresses a real operational need rather than a programmatic one.

Second, the concentration risk is significant. A single commercial vendor supplying the decision layer across services and combatant commands creates dependency of a kind the department has historically tried to avoid. The Department has signaled awareness of this, mapping out plans for a new enterprise command-and-control program office and C2 suite, which suggests an intent to establish government architecture ownership around the capability.

Where AI actually helps

Separating substantiated value from vendor claims is useful.

Data fusion and correlation. Multiple sensors observing the same area produce overlapping, contradictory, and incomplete reports. Correlating these into a single track picture is computationally intensive and error-prone for humans at scale. Machine assistance here is well-proven and represents the majority of realized value.

Object detection and classification. Identifying vehicles, aircraft, vessels, and structures in imagery and full-motion video. This was Project Maven's original purpose. Mature, valuable, and dramatically reduces analyst workload on a task humans perform slowly and inconsistently.

Pattern-of-life analysis. Detecting deviations from established patterns across large datasets. Well-suited to machine analysis and difficult for humans across the data volumes involved.

Course of action generation. Producing and evaluating options against constraints. Useful for enumerating possibilities and identifying overlooked options, though the quality of evaluation depends heavily on how well the model captures the actual operational situation.

Resource allocation and deconfliction. Assigning sensors to targets, weapons to threats, and aircraft to airspace. A well-structured optimization problem where machine assistance clearly outperforms manual methods.

Logistics and predictive maintenance. Less discussed and possibly higher value than the targeting applications. Forecasting demand, predicting failures, and optimizing distribution deliver measurable readiness improvements.

Where AI helps less than claimed

Equally worth stating plainly.

Intent inference. Determining what an adversary intends from observed behavior requires contextual and cultural understanding that current systems do not possess. Confident machine assessments of adversary intent should be treated with substantial skepticism.

Novel situations. Machine learning systems perform well on distributions resembling their training data and poorly outside them. Military operations are characterized by adversaries deliberately creating novel situations. This is a structural limitation, not a temporary one.

Adversarial robustness. Any system whose behavior can be inferred can be manipulated. Adversaries will use camouflage, decoys, and deliberately crafted inputs to induce errors. This is an active and unresolved research area.

Explanation. When a system recommends an action, understanding why is essential for a commander deciding whether to accept it. Current explainability techniques provide limited insight, particularly for deep learning systems. This is a serious operational limitation, not an academic concern.

Data quality dependency. These systems inherit the quality of the data they consume. Military data is frequently incomplete, inconsistently formatted, and stale. The unglamorous work of data engineering determines outcomes more than algorithm selection does.

The acceleration problem

The strategic case for AI-enabled command and control is decision speed. That case has an uncomfortable corollary.

If both sides field systems that accelerate decision cycles, the result is not that one side gains advantage — it is that both operate at a tempo that compresses the space for human judgment. Crisis management, de-escalation, and the deliberate pause that has historically prevented miscalculation all require time that automated systems remove.

This is a genuine strategic stability concern, raised consistently by analysts and inadequately addressed in acquisition. The competitive dynamic makes unilateral restraint unattractive: a force that deliberately decides more slowly than its adversary accepts operational disadvantage.

There is no clean answer. Partial mitigations include deliberate decision checkpoints for consequential actions, architectural separation between recommendation and execution, and clear delineation of which decisions require human deliberation regardless of time pressure. Whether these survive operational pressure in a crisis is untested.

A second-order effect deserves mention: automation bias. Operators presented with confident machine recommendations under time pressure tend to accept them. The formal presence of a human in the loop does not guarantee meaningful human judgment if the interface, the tempo, and the institutional expectation all push toward acceptance. Interface design is therefore a safety-critical function.

Architecture and industry structure

Several structural questions will determine how this market develops.

Government architecture ownership versus vendor platforms. The department's move toward an enterprise C2 program office suggests intent to own the architecture while integrating commercial capability. Whether this succeeds against the gravitational pull of an established, widely adopted vendor platform is uncertain and consequential.

Open standards and data portability. If data and models are portable across vendors, competition persists. If they are not, the incumbent's position becomes structural. This is the single most important policy question in the segment.

Classification boundaries. Operating across multiple security domains creates architectural complexity that favors vendors with existing accreditation — a meaningful barrier to entry that shapes competition.

Edge versus enterprise. Cloud-based analytics fail when connectivity fails. Distributed contested operations require capability at the tactical edge with intermittent connectivity. This drives fundamentally different technical approaches, and the enterprise-first architectures currently dominant may not extend well.

Coalition interoperability. CJADC2 explicitly includes coalition partners. Sharing data and models across national boundaries raises classification, sovereignty, and industrial policy questions that remain largely unresolved.

Practical guidance

For programs and organizations working in this space:

Invest in data engineering before algorithms. The dominant failure mode is not inadequate models. It is inconsistent, incomplete, and inaccessible data. Organizations that fix data infrastructure realize value; those that buy models without fixing data do not.

Design interfaces for calibrated trust. The interface must convey confidence and uncertainty accurately enough that operators trust the system when it is reliable and question it when it is not. Uniform confident presentation destroys the human check.

Require evaluation against adversarial conditions. Performance on curated test data predicts nothing about performance against an adversary actively inducing errors.

Preserve human deliberation for consequential decisions. Architecturally separate recommendation from execution, and define which decision classes require human judgment regardless of time pressure.

Plan for degraded operation. Systems assuming reliable high-bandwidth connectivity will fail in the environments they are procured for. Graceful degradation to local capability is a threshold requirement.

Insist on model and data portability. Vendor lock-in at the decision layer is a strategic risk, not merely a commercial one.

Outlook

AI-enabled command and control will expand, because the operational value in data fusion and analyst workload reduction is real and demonstrated. Maven's organic adoption growth is evidence of genuine utility rather than programmatic momentum.

Expect the department to pursue government architecture ownership while continuing to buy commercial capability — a difficult balance that has failed before and may fail again. Expect edge computing and degraded-operations requirements to reshape architectures currently designed around enterprise cloud assumptions. Expect coalition data sharing to remain a persistent friction point.

And expect the acceleration problem to remain unaddressed in acquisition while being raised repeatedly in analysis. The competitive dynamics that make decision speed valuable are the same dynamics that make deliberate slowness unattractive. That tension is unlikely to resolve through technology.

The realistic assessment is that AI in command and control delivers substantial, unglamorous value in data handling and analyst support, more limited value in decision recommendation, and carries strategic risks that are understood but not managed. That is a reasonable place for a maturing capability to be — provided the limitations are acknowledged in how it is fielded and employed.

There is also a workforce dimension that receives far less attention than the technology. Systems of this kind require operators who understand both the operational problem and the limitations of the tools — enough to recognize when a recommendation is outside the system's competence. That skill set does not currently exist in quantity, is not systematically trained, and has no established career path in most services.

The consequence is predictable. Capable software is fielded to users who either distrust it entirely and revert to manual methods, wasting the investment, or trust it uncritically and inherit its errors. Neither outcome delivers the intended value. Organizations that have invested in training a cadre of users who genuinely understand the tools report substantially better outcomes than those that fielded software and assumed adoption would follow. This is an unglamorous, sustained investment that competes poorly for attention against capability procurement, and it is probably the highest-return action available.

Frequently asked questions

What is CJADC2? Combined Joint All-Domain Command and Control — the Department of Defense concept for connecting sensors, weapons, and decision-makers across all domains and with coalition partners, moving data seamlessly at machine speed rather than through manual coordination.

What is the Maven Smart System? Palantir's AI-enabled software platform supporting CJADC2, providing interfaces for intelligence analysis and targeting. As of March 2026 it had over 20,000 active users across 35 tools and three classification domains, with the user base having doubled since January 2026.

Does AI make targeting decisions? In current fielded systems, AI assists with detection, classification, correlation, and recommendation while humans authorize engagements. The practical concern is automation bias — whether human authorization remains meaningful under time pressure and confident machine recommendations.

What is the main risk of AI in command and control? Two risks dominate: compression of decision timelines to the point that human deliberation and de-escalation become impractical, and automation bias in which operators accept machine recommendations without adequate scrutiny.

How much is the Pentagon spending on AI command and control? The fiscal 2027 budget request includes more than $2 billion for command-and-control technology licenses and engineering support, with more than $1.5 billion directed toward expanding access to the Maven Smart System.