Military AI by the Numbers: Defense AI Adoption Statistics (2026)

Last updated: October 26, 2026

Why it matters to your mission. The debate about AI in defense is loud, but the decisions get made on numbers: how many programs are fielded, how much is funded, how often projects fail, and how much faster a staff can actually act. This page gathers those numbers from primary government, oversight, and service-research sources, with each figure linked to the document it came from. Read together, they tell one story: the Department of War's constraint is less the model and more the fragmented data and manual workflows underneath it.

Definition

Defense AI adoption is the degree to which military organizations field artificial intelligence in real missions, measured by programs, spending, data readiness, decision speed, and governance rather than by demonstrations alone. The statistics below are organized around those five dimensions.

Key takeaways

• Adoption is wide but shallow. Federal AI use cases nearly doubled in a year, and the Pentagon's internal generative-AI tool reached 1.7 million users, yet most catalogued military AI projects are still in development, not fielded.
• The labeled "AI budget" is small and opaque. The explicit AI request held flat at $1.8 billion across two years, while far larger sums move through data and autonomy lines.
• Data, not the algorithm, is the bottleneck. A RAND study puts AI project failure above 80 percent; oversight and service research repeatedly trace the problem to fragmented data and manual "swivel chair" workflows.
• When the data problem is solved, decisions speed up measurably. Service experiments show AI compressing planning and sensor-to-shooter timelines by meaningful margins.
• Governance and trust lag the technology. The office that governs DoD AI was found overdue on its own policy, and models remain vulnerable to manipulated inputs.

Methodology and sourcing

Every statistic on this page is verified against a primary or authoritative source (GAO, DoD Inspector General, the DoD Comptroller, CBO, service war colleges and research institutes, FFRDCs such as RAND, and USAspending), and each is linked to that source. Figures we computed ourselves from the government's own spending data are labeled "Torch.AI analysis of USAspending data" so they can be reproduced. Where a figure is scoped to a single study or experiment, we keep it scoped rather than presenting it as department-wide. We do not publish numbers we could not confirm to a primary source.

How widely is AI being adopted across the Department of War?

Adoption is broad in appetite and uneven in fielding.

• Reported federal AI use cases nearly doubled from 571 in 2023 to 1,110 in 2024 - and the Department of War is exempt from that public inventory, so the government's largest AI user is also its least visible. (GAO-25-107653, 2025)
• The Pentagon's internal generative-AI platform reached 1.7 million active users, more than 42 percent of the total defense population, with 100,000 user-built AI agents and over 11 million documents processed, going from launch order to debut in 54 days. (CNAS, 2026)
• But the first public DoD AI inventory was mostly immature: 88 percent of its 685 identified AI projects came from research-and-development budget documents, meaning the capability was not yet fielded. (GAO-22-104765, 2022)
• Only 17 of the 88 reported major weapon systems had an AI project clearly identified in that inventory. (GAO-22-104765, 2022)
• For FY2026, the Pentagon for the first time broke out autonomy as its own budget line, requesting $13.4 billion for autonomous and remotely operated systems across all domains. (DoD FY2026 Budget briefing, 2025)

What does the Department of War spend on AI?

The money is large in aggregate and hard to pin to a single "AI" line.

• DoD requested over $1.4 billion for CJADC2 in its FY2025 budget, yet GAO found no comprehensive framework guiding those investments. (GAO-25-106454, 2025)
• The explicit, stand-alone AI request was $1.8 billion in both FY2024 and FY2025, and officials could not immediately enumerate all the programs the figure covers. (DoD FY2025 budget, via DefenseScoop, 2024)
• In FY2024, DoD obligated about $187.5 million across 244 prime contract actions that reference "artificial intelligence." (Torch.AI analysis of USAspending data, FY2024)
• Those AI-keyword contracts grew roughly 6.4x in five years, from 38 in FY2019 to 244 in FY2024. (Torch.AI analysis of USAspending data)
• Spending is concentrated: the single largest AI-keyword contract in FY2024 was about 64 percent of all AI-labeled contract dollars. (Torch.AI analysis of USAspending data)
• The Defense Innovation Unit's funding jumped nearly tenfold in one year, from $191 million enacted in FY2023 to $983 million in FY2024. (GAO-25-106856 + DoD Appropriations Act 2024)
• CBO found DoD spends over $300 million a year on counter-small-UAS research and development, and estimated that fielding layered counter-drone defenses at 100 U.S. installations would cost about $7.4 billion plus $500 million a year to operate. (CBO, pub. 62262, 2026)

Why do defense AI projects struggle? The data bottleneck

This is the through-line across oversight, FFRDC, and service research: the hard part is the data, not the model.

• A RAND study drawing on interviews with 65 data scientists and engineers reports more than 80 percent of AI projects fail, about twice the failure rate of non-AI IT projects, with inadequate data and data infrastructure among the root causes. (RAND RR-A2680-1, 2024)
• RAND's congressionally mandated assessment found DoD's AI posture "significantly challenged across all dimensions," and that DoD data, where it exists, often lacks traceability, understandability, accessibility, and interoperability. (RAND RR-4229-OSD, 2019)
• Army leaders have historically spent as much as 80 percent of their time on data and information processing rather than higher-order thinking. (Modern War Institute, 2026)
• DoD still runs command and control largely through "swivel chair" analysis - analysts manually re-keying data from one system into another - which GAO says prevents decisions at the needed scale and tempo. (GAO-25-106454, 2025)
• After 12 Global Information Dominance Experiments, the flagship integration result was connecting just two large data systems so commanders could assess logistics for proposed courses of action. (GAO-25-106454, 2025)
• The Air Force's DCGS support to airborne ISR missions grew more than 1,900 percent from 2001 to 2015, outpacing the manual processes meant to handle it. (Air & Space Power Journal, 2017)
• As far back as 2013, one Air Force ISR system ingested 1.3 petabytes a month, about 1,000 hours of full-motion video every day. (Air & Space Power Journal, 2020)

How much does AI actually speed up decisions?

When the data is made usable, service experiments show measurable gains. These results are scoped to the studies that produced them.

• In a West Point systems-engineering simulation, AI decision aids cut the sensor-to-shooter interval from 11.0 to 7.7 minutes, nearly a 30 percent compression of the decision cycle. (Parameters, Army War College, 2025-26)
• In a JRTC field experiment, an AI-assisted brigade staff produced complete, doctrinally formatted operations orders in about 23 hours each. (Modern War Institute, 2026)
• A Naval Postgraduate School thesis showed transfer learning let an open-source model detect ships in satellite imagery at 1.00 recall and 0.98 precision, using as few as 200 training images. (NPS, 2018)
• Another NPS thesis predicted a vessel's future at-sea position from historical tracks with 94 percent accuracy, enabling automated anomaly flagging. (NPS, 2017)

What are the trust and governance gaps?

Adoption is outrunning governance, and models remain brittle on unverified data.

• The DoD Inspector General found that the office governing all DoD AI was overdue issuing the AI policy it was directed to write in 2022, and its roles for DoD data, analytics, and AI were not clearly defined to stakeholders. (DODIG-2025-039)
• Air Force researchers showed imperceptible "universal adversarial perturbations" could make AI target-recognition systems misclassify military vehicles such as tanks up to 64 percent of the time - a direct argument for provenance-controlled data. (Aether, Air University, 2024)
• The Air Force stood up TITAN, an $18 million government-owned, 7.6-petabyte hybrid cloud, to give its AI and ISR data a secure, government-controlled home. (Air & Space Power Journal, 2020)

How Torch.AI reads these numbers

Read together, the figures point away from "buy a better model" and toward "make the data usable and the reasoning traceable." Projects fail mostly on data, not algorithms; the largest time cost is manual data handling; and the biggest fielded wins are small relative to how fragmented the underlying systems are.

Torch.AI builds government-owned reasoning infrastructure for exactly that gap. It connects to existing systems of record rather than replacing them, turns fragmented, multi-source data into coherent understanding at machine speed, and keeps every result traceable to its source. NEXUS applies semantic representation so unstructured data becomes something you can query and correlate. HALO fuses resolved entities into a graph an analyst can trust and trace. ORCUS is the data-availability layer that connects to the sources and preserves provenance end to end. Because the reasoning layer runs on top of the authoritative systems and can be deployed government-owned, the customer keeps control of the layer where data becomes decision. You can see how this is packaged on the Torch.AI software page, or request a technical walkthrough against a sample of your own data.

Sources

All figures above link to their primary source inline. Primary and authoritative sources used on this page include the U.S. Government Accountability Office (GAO-22-104765, GAO-25-106454, GAO-25-106856, GAO-25-107653), the DoD Inspector General (DODIG-2025-039), the DoD Comptroller FY2025 and FY2026 budget materials, the Congressional Budget Office (publication 62262), USAspending.gov (for Torch.AI-computed contract figures), RAND (RR-4229-OSD, RR-A2680-1), the Center for a New American Security, the U.S. Army War College (Parameters) and Modern War Institute, the Naval Postgraduate School, and Air University (Air & Space Power Journal and Aether).

How much does the U.S. military spend on AI?

The explicit stand-alone DoD AI budget request was $1.8 billion in both FY2024 and FY2025, but far larger sums move through adjacent data and autonomy lines, including a first-ever $13.4 billion FY2026 autonomy line. The labeled AI figure understates total AI-related investment.

What percentage of AI projects fail?

A 2024 RAND study reports more than 80 percent of AI projects fail, roughly twice the failure rate of IT projects that do not involve AI, with data and data-infrastructure problems among the root causes.

Why is defense AI adoption so hard?

Oversight and service research consistently trace the difficulty to fragmented, hard-to-access data and manual workflows rather than to the AI models themselves. DoD data often lacks traceability, understandability, accessibility, and interoperability.

How much faster does AI make military decisions?

In controlled service experiments, AI has compressed a sensor-to-shooter interval from 11.0 to 7.7 minutes and helped a brigade staff produce operations orders in about 23 hours each. These results are scoped to the studies that produced them.