Why this tour exists
Most writing about "federal AI" is written at thirty thousand feet — AI strategy, AI principles, AI guardrails. That writing is useful for policy, useless for a small business trying to find its first contract. The question that matters to a founder is concrete: which agency, which office, which mission, which contract vehicle. This article is that tour. It walks through fifteen-plus agencies, with specific programs and specific patterns, drawn from the public AI use case inventories agencies are required to publish under OMB M-24-10 and the AI in Government Act, plus award data from SAM.gov and USAspending.gov.
The goal is not completeness. Each of these agencies publishes dozens to hundreds of use cases. The goal is to give a small AI firm a starting map — where the density is, what the patterns look like, and which doors are worth knocking on.
Department of Defense components
U.S. Army — logistics, predictive maintenance, intelligence
The Army's AI portfolio is unusually concrete because it is tied to specific programs of record. Predictive maintenance for ground vehicles and aviation (Bradley, Stryker, Apache, Blackhawk) dominates — classical ML applied to sensor streams and maintenance logs to predict faults before mission-impact. Logistics optimization at the theater level is another major thread. On the intelligence side, the Army's DCGS and associated PED cells are consuming increasingly sophisticated NLP and vision tools.
- Predictive maintenance on ground and aviation platforms
- Theater-level logistics optimization (ammunition, fuel, spare parts)
- Geospatial intelligence processing (EO, SAR, full-motion video)
- Medical readiness forecasting for AMEDD
- Recruitment and retention analytics at HRC
Small-business entry is well-paved: xTechSearch, SBIR topics through DSIP (typically 80 to 120 Army topics per cycle), and OTA consortia like AMTC and S2MARTS.
U.S. Navy — undersea, EW, fleet maintenance
Navy AI is split across NAVAIR, NAVSEA, NAVWAR, ONR, and the Marine Corps. NAVAIR's predictive maintenance on aviation platforms is the highest-dollar production use. NAVWAR runs the largest AI-adjacent portfolio in the Navy around electromagnetic spectrum operations, network defense, and command and control. Undersea autonomy is a major ONR push — classifiers for sonar, autonomous mission planning for unmanned undersea vehicles.
- Sonar classification and undersea target recognition
- EW and spectrum-awareness ML at NAVWAR
- Predictive maintenance across aviation and surface fleet
- Autonomy for unmanned surface and undersea vehicles
- Logistics and fleet readiness forecasting
Navy SBIR is the most volume-heavy of any service, with some of the clearest Phase III transition paths.
Department of the Air Force — air ops, space domain awareness, targeting
AFWERX and SpaceWERX drive the visible portion of Air Force AI work. Behind them, specific programs — Kessel Run for air operations software, Platform One for DevSecOps, Rogue Squadron and Big Safari for capability integration — are the real consumers. Space Force is the fastest-growing AI customer in the department, with space domain awareness (tracking, characterization, behavior analysis of space objects) as its dominant use case.
- Air operations center software and decision support (Kessel Run)
- Space domain awareness and satellite behavior analytics
- Targeting and sensor fusion pilots (careful scoping required)
- Aircraft predictive maintenance (F-16, B-52, KC-46)
- Personnel and financial management ML
CDAO and joint programs — data, test & evaluation, AI safety
The Chief Digital and AI Office is the department-wide function for AI policy, responsible AI, AI test and evaluation, and data governance. CDAO runs Task Force Lima (generative AI) and coordinates the DoD AI use case inventory. The dollar volume is smaller than service programs, but CDAO's T&E methodology work often sets the pattern the services adopt.
Civilian agencies with heavy AI production use
Department of Veterans Affairs — clinical, imaging, claims
VA publishes the largest AI use case inventory of any federal department. The portfolio leans clinical: imaging classifiers for radiology and pathology, clinical decision support for triage and risk stratification, suicide-risk prediction models (REACH VET is the best-known), claims processing and evidence extraction. VA's data environment is uniquely rich — it is a direct nationwide integrated health system — and that makes clinical ML tractable at scale.
- Imaging classifiers across radiology, pathology, dermatology
- Clinical risk stratification (REACH VET, heart failure, sepsis)
- Claims evidence extraction and rating automation
- Scheduling and no-show prediction
- Veteran-facing virtual assistants at VA.gov
VA procurement is split between T4NG, VECTOR, and set-aside vehicles. SBIR is limited. The realistic path for a small firm is a teaming seat on a prime's T4NG task order.
HHS — CMS, NIH, FDA, CDC, SAMHSA
HHS is the other giant in federal AI. CMS runs large fraud-and-abuse ML systems on claims data. NIH uses AI for research portfolio analysis (NIH RePORTER), grant triage, and scientific literature search. FDA is using ML for adverse event detection and post-market surveillance. CDC deploys forecasting models for communicable disease. SAMHSA has a long-running analytics portfolio around the treatment locator, the behavioral health surveys, and the opioid overdose surveillance data (see our ML capability for how we apply this pattern; HHS agency page for full portfolio).
Treasury & IRS — fraud detection, compliance, forecasting
Treasury's production AI is heavily concentrated in fraud and non-compliance detection. IRS uses ML to prioritize audits, detect identity theft in refund fraud, and flag tax preparer anomalies. FinCEN deploys ML on BSA filings. The OCC, FDIC, and Federal Reserve use ML for bank supervision and stress testing. Dollar volume is significant and procurement tends to be vehicle-driven (IRS has its own large vehicles plus GSA Schedule).
Social Security Administration — disability claims, fraud
SSA has one of the longest-running federal ML programs, predating the current AI wave by more than a decade. The headline use case is disability claims processing — ML-assisted medical evidence extraction and quality review for DDS decisions. Fraud and identity theft detection is the other large production program. SSA is a conservative customer that rewards vendors with documented model governance and explainability.
Department of Homeland Security — TSA, CBP, CISA, USCIS, FEMA
DHS AI is diverse because the department is diverse. TSA uses computer vision for checkpoint imaging and risk-based passenger screening. CBP runs ML on targeting and cargo screening and an extensive facial-recognition program. CISA has growing AI portfolios around network defense, vulnerability prioritization, and threat intelligence. USCIS uses NLP on immigration benefit adjudication. FEMA uses ML for disaster damage assessment (often fused with commercial satellite imagery).
Science, energy, and research agencies
Department of Energy — national labs, grid, fusion, science AI
DOE is the most technically sophisticated AI buyer in the civilian government, because its seventeen national labs are themselves producers of foundation-class AI research. ORNL, LBNL, ANL, LLNL, PNNL, and SNL all run major AI programs. Grid AI (load forecasting, phasor data analytics, grid resilience), materials discovery, fusion diagnostics, and scientific foundation models are the dominant threads. Small-business paths include SBIR (DOE runs a well-organized cycle), direct lab subcontracts, and CRADAs.
NASA — mission data, space operations, Earth science
NASA's AI work lives in three clusters: ground operations ML (anomaly detection on spacecraft telemetry), science data processing (Earth science, heliophysics, planetary), and mission autonomy (on-board processing for rovers, edge inference for small satellites). JPL, Goddard, Ames, and Langley each host distinct AI portfolios. SBIR is the cleanest small-business entry.
NSF — research funding and its own portfolio analytics
NSF is primarily a funder of AI research rather than a consumer, but it does run internal AI for proposal analytics, portfolio mapping, and merit review workflow. NSF SBIR/STTR is one of the best-funded small-business programs in the civilian government for early-stage AI companies.
NIH — biomedical AI, literature, grants
Inside HHS but worth calling out separately. NIH's AI is substantially biomedical: NCI's Cancer Research Data Commons, NLM's literature and PubMed tools, NHLBI imaging, and the All of Us genomics program. NIH SBIR is generous and well-run — three-month I/II bridges, clear Phase III paths, and program officers who genuinely engage with small firms.
NIST — AI measurement, testing, standards
NIST is not a major operational consumer of AI but it is the measurement authority for the federal government. NIST's AI Risk Management Framework, AI Safety Institute (AISI), and evaluation programs (like ASPEN and the AI Pilot Evaluation) shape how every other agency ends up buying AI. Small firms working on evaluation, red-teaming, and model auditing often end up in NIST's orbit.
Law enforcement, intelligence, and justice
FBI and the Department of Justice — case analytics, OSINT, forensics
FBI's AI work spans case triage, digital forensics (OCR, audio-visual processing, link analysis across seized devices), OSINT collection, and counterintelligence analytics. DEA, ATF, and USMS have smaller but similar portfolios. Entry is typically via an FBI BAA or subcontract to an incumbent. The bar for operational security is high and rewards vendors with documented control discipline.
Intelligence Community — NGA, NSA, CIA, DIA, ODNI
The IC is the largest non-DoD AI customer, but its spending is largely invisible. NGA is the most prolific AI consumer — geospatial analytics, automated imagery exploitation, change detection, pattern-of-life analysis. NSA's AI work is largely classified. CIA In-Q-Tel invests in dual-use AI companies. DIA runs the Machine-assisted Analytic Rapid-repository System (MARS) program. Small-business entry requires clearance or a cleared prime partner and is the slowest path of any listed here.
Regulatory and independent agencies
Financial and consumer regulators
SEC runs ML on EDGAR filings for anomaly detection and market surveillance. CFTC applies similar techniques to derivatives markets. FTC has growing AI programs around consumer protection and deceptive practices detection. Procurement tends to run through GSA vehicles and small contract ceilings favorable to small firms.
Commerce, USPTO, Census
USPTO uses AI heavily in patent classification and prior-art search. Census runs ML across survey quality, imputation, and response prediction. NOAA has one of the largest environmental ML portfolios in the government — weather forecasting foundation models, fisheries assessment, climate modeling. NOAA's AI work in particular is growing fast and has a dedicated NOAA AI Center in Reston.
Patterns across the portfolio
Step back from the specifics and a few patterns emerge:
- Classical ML is the production backbone. Across almost every agency, the dollars and operational impact sit on classification, forecasting, anomaly detection, and triage — not on generative AI. Generative and agentic AI are growing but are still mostly in pilot.
- Mission data is the moat. Agencies with rich, domain-specific data (VA clinical, IRS tax, NGA imagery, CMS claims) run the most sophisticated ML. Small firms that bring mission-aware modeling, not just generic model plumbing, get repeat business.
- Compliance is a capability, not a tax. Every agency listed here has strict data-handling requirements. Vendors that treat FedRAMP, ATO, and audit logging as table stakes (see our FedRAMP map) ship faster and win more. Vendors that treat compliance as an afterthought burn months and lose awards.
- Transition is the killer. Pilots are easy to start and hard to scale. The winners are the firms that design for Phase III — integration into an existing program of record — from Phase I forward.
Where a small business should actually start
With 400+ public AI use cases across the federal inventory, the risk is analysis paralysis. A simple filter:
- Pick two or three agencies where your technical depth maps to the mission.
- Read the last two cycles of their SBIR topics. That is the cleanest signal of what program offices are willing to fund.
- Pull the last twelve months of awards on relevant NAICS (541511, 541512, 541519, 541715) and see who the incumbents are.
- Find three sources-sought notices or RFIs on your mission from the last six months. Respond to all three. That is free positioning.
- Write a one-page capabilities statement aimed at the specific program office, not at the agency generically.
The federal AI market is enormous. It is also legible. The firms that win are not the ones who chase every opportunity — they are the ones who commit to a small number of agencies and go deep on mission.
Frequently asked questions
Under OMB M-24-10 and the AI in Government Act, CFO Act agencies publish annual AI use case inventories to ai.gov and their own agency sites. DoD components publish separately under CDAO guidance. Those inventories are the cleanest public signal of what is in production or pilot.
VA leads on raw count. HHS leads collectively (VA, CMS, NIH, FDA, CDC). DoD dollar volume is larger than any civilian department but is reported differently. Raw count is a weak proxy for dollar size — a single large DoD capability can dwarf dozens of small civilian pilots.
The majority are tagged 'deployed,' but 'deployed' often means 'narrow decision support in a single office.' Agentic AI is mostly in pilot. Classical ML is where production density is highest.
Army SBIR (via xTechSearch and DSIP), Navy SBIR (via NAVWAR and NAVAIR), NIH SBIR (via NCI and NLM), and DOE SBIR are the highest-leverage entry points. Each combines dollar volume, predictable topic cadence, and willingness to work with first-time small businesses.
Combine the inventory entry's listed program office with SAM.gov award history on related NAICS, USAspending.gov award descriptions, and FedBizOpps sources-sought on the same mission. Three signals triangulating on the same office justify a capabilities meeting request.
No. Classified programs are explicitly excluded. The IC and classified DoD portfolios are substantially larger than what shows in the public inventories. For those, the signal comes from BAAs, industry days, and cleared-prime relationships rather than published inventories.