Fixing Navy R&D: A Faster Path for Defense AI

A decade is an eternity in military technology.

Yet it’s still normal for Defense Department research programs to take 10+ years to become something sailors can actually use—if they make it that far. Government watchdogs have been pointing at the same culprits for years: layered oversight, risk aversion, “pilot purgatory,” and contracting models that reward activity more than outcomes.

That’s why the Navy’s recent leadership change at the Office of Naval Research (ONR) matters to anyone tracking AI in defense & national security. The Navy has put Rachel Riley, a 33-year-old former McKinsey partner and Department of Government Efficiency (DOGE) alum, in the acting chief role. The reaction from observers is telling: a mix of optimism that someone will finally say “no” to stalled projects, and concern that reform energy could accidentally gut the kind of science no venture capitalist will fund.

Here’s the stance I’ll take: ONR doesn’t need “more innovation theater.” It needs ruthless focus on fielding—while protecting the research that keeps the U.S. competitive against China in the long run. If Riley can balance those two, it could materially change how the Navy adopts autonomy, AI-enabled sensing, and decision support.

Why ONR leadership matters for defense AI outcomes

ONR’s leadership matters because it decides what gets funded, how fast it’s expected to mature, and whether “research” ever becomes a deployable capability. In practical terms, ONR influences the Navy’s pipeline for autonomy, machine learning, human-machine teaming, undersea sensing, electronic warfare, and the data infrastructure that makes modern AI possible.

What’s different about this moment is the surrounding pressure. As of late 2025, the defense ecosystem is being squeezed from both sides:

• Commercial pace is accelerating. Venture-backed defense and dual-use companies are shipping autonomy stacks, sensing payloads, and edge compute faster than traditional timelines.
• Federal science funding is under stress. As broader research budgets tighten, DoD becomes a backstop for strategically vital science.

That creates a leadership test: can the Navy buy what already exists, build what must be unique, and stop paying for “perpetual science projects” that don’t deliver?

The real constraint isn’t AI talent—it’s time-to-field

Defense teams often frame the problem as a shortage of AI engineers or data scientists. It’s not.

The binding constraint is the system around them: requirements that mutate, contracting that can’t keep up, integration that’s treated as someone else’s problem, and test authorities that arrive late. When programs take a decade, two predictable things happen:

1. The commercial market moves on, so the government fields something already obsolete.
2. Teams spend more time in meetings and compliance cycles than building and validating models.

Riley’s background—reform work, bureaucracy reduction, and organizational restructuring—signals the Navy is at least trying to attack the “system problem,” not just the “model problem.”

What a DOGE-style reform mindset could (and should) change

The useful part of DOGE’s ethos is speed, clarity, and measurable outcomes—applied carefully. The article notes DOGE’s broader controversies, but also highlights that the Pentagon’s DOGE elements have operated differently than some of the more chaotic stories elsewhere.

If Riley brings the right version of that mindset to ONR, three shifts would have outsized impact on AI and autonomy programs.

1) Make “commercial first” the default for autonomy and AI tooling

The fastest way to modernize military AI is to stop reinventing what the private sector already sells. For many AI components—data labeling workflows, MLOps platforms, edge inference runtimes, model monitoring, synthetic data generation—there’s a mature commercial market.

ONR should treat “commercially available” as a formal gate:

• If a capability exists commercially, buy and adapt.
• If it doesn’t, fund research with a clear transition plan.

This approach matters in autonomy. Commercial autonomy is being driven by logistics, maritime shipping, mining, agriculture, and industrial robotics—fields that push perception and navigation forward quickly because there’s money on the line.

Defense-specific work should focus on what’s actually unique: contested comms, adversarial deception, emissions control, cyber resilience, kinetic safety constraints, and mission assurance.

2) Replace research “activity metrics” with fielding metrics

Most companies get this wrong when they modernize R&D: they track outputs that are easy to count, not outcomes that matter. ONR should move programs toward a small set of fielding metrics that are hard to fake:

• Time from award to first operational experiment (in months)
• Number of test events with real users (operators, not just engineers)
• Integration readiness (data interfaces, compute footprint, security path)
• Cost per test iteration
• Demonstrated performance under degraded conditions (GPS loss, comms denial, sensor noise)

For AI systems, “performance” can’t be a single accuracy number. It has to include robustness, drift behavior, and failure modes in operationally realistic settings.

A blunt but effective rule: if nobody can explain how the work gets onto a ship, submarine, aircraft, or maritime drone, it’s not a program—it’s a seminar series.

3) Kill sacred cows—and then prove you didn’t kill the wrong ones

The source quotes a former defense official arguing that some ONR enterprises haven’t produced deployable outcomes and survive because the organization is uncomfortable saying “no.” That rings true across government R&D.

But here’s the trap: cutting is easy; cutting correctly is hard. If ONR swings a scythe without a taxonomy of what it funds, it risks gutting strategically vital research that doesn’t look like a near-term product.

A better approach is a transparent portfolio model:

1. Near-term transition (6–24 months): prototypes that can enter programs of record, or be procured as capabilities.
2. Mid-term advantage (2–5 years): autonomy in complex environments, undersea sensing fusion, resilient navigation, EM spectrum AI.
3. Long-horizon science (5–15 years): oceanography, marine geoscience, cryptography, foundational sensing physics.

Cut hard in bucket #1 when it stalls. Be disciplined in bucket #2. Protect bucket #3 because it’s where strategic surprise comes from.

The AI angle: speed without breaking mission assurance

Defense AI fails in two predictable ways: either it ships too slowly, or it ships without trust. ONR’s job is to help the Navy avoid both.

Speed is not the same as recklessness. For AI-enabled national security systems, mission assurance has to be built into the development loop.

What “mission-ready AI” looks like in Navy programs

If you’re evaluating whether an AI project is actually on a path to fleet value, look for these characteristics:

• Edge-first design: can it run on constrained compute, power, and bandwidth?
• Degraded-mode operation: does it work when sensors fail or inputs are adversarial?
• Human-in-the-loop clarity: does the operator understand why the system recommends an action?
• Test evidence: does it have results from realistic trials, not just lab benchmarks?
• Security pathway: is there a plan for ATO, patching, and model updates over time?

A practical litmus test I’ve found helpful: if the team can’t describe how the model gets updated safely after deployment, they’re not building a fielded capability—they’re building a demo.

The Navy’s unique advantage: the ocean is a data problem

Navy missions are inseparable from the ocean environment: acoustics, salinity layers, seabed terrain, weather-ocean coupling, and undersea propagation. Those aren’t just “science topics.” They’re the priors that make undersea AI work.

That’s why the article’s warning is so important: some ONR-funded work—ocean climate science, physical oceanography, marine biology, high-end encryption—won’t attract enough private investment.

If ONR stops funding that science, the Navy doesn’t just lose papers. It loses operational advantage in undersea awareness and navigation.

A practical blueprint for modernizing ONR around AI

The fastest improvement comes from changing how programs transition from research to capability. Below is a concrete blueprint ONR leadership can use without waiting for a multi-year reorg.

1) Stand up “AI transition cells” tied to real fleets

Create small, empowered teams that sit between researchers and operators. Their job is simple: move prototypes into operational experiments fast.

Each transition cell should include:

• An operator rep (Surface, Sub, Aviation, Marines)
• A test/evaluation lead
• A security/compliance lead
• An integration engineer (data + interfaces)
• A contracting specialist who can execute quickly

This is how you prevent “research islands” that never plug into operational systems.

2) Use milestone-based funding that forces learning

For AI programs, milestone gates should include:

1. Data readiness: labeled data, governance, and access are solved.
2. Baseline model: a working model beats a slide deck.
3. Robustness tests: adversarial and degraded-input testing completed.
4. Operational experiment: trial with users in a realistic environment.

Funding should increase only when gates are met. That’s not punitive—it’s how you keep portfolios honest.

3) Normalize “buy + integrate” alongside “invent”

ONR should explicitly reward teams for integrating proven components rather than building everything from scratch. When teams are punished for using commercial tools (“not invented here”), the Navy pays more and gets less.

A strong policy statement from leadership can change this culture quickly: integration is engineering work, and engineering work is valuable.

4) Build an off-ramp for programs that don’t work

Every research organization needs a clean way to stop work. Without it, zombie programs accumulate.

A disciplined off-ramp includes:

• A documented stop decision
• Reusable artifacts (datasets, code, test results)
• A short “what we learned” memo that informs future efforts

This turns failure into institutional knowledge instead of institutional embarrassment.

What government and industry leaders should do next

If you lead in DoD, a prime, or a defense tech firm, this ONR leadership shift is a planning moment. The Navy is signaling it wants faster outcomes. Align accordingly.

For defense organizations

• Identify where you’re duplicating commercial AI capabilities and stop.
• Push for test events early; don’t wait for perfect requirements.
• Budget for integration, security, and sustainment from day one.

For defense tech and dual-use companies

• Package offerings around “time-to-operational-experiment,” not just features.
• Bring a security pathway and update model (patching, drift monitoring, retraining).
• Show how your autonomy stack behaves under comms loss and sensor noise.

For ONR and Navy program owners

• Treat “transition” as a first-class deliverable.
• Protect long-horizon science that underwrites undersea advantage.
• Make it socially acceptable to cancel programs that don’t meet gates.

A simple sentence can guide the entire portfolio: If it can’t transition, it’s not done.

Where this fits in the AI in Defense & National Security series

This post sits at the uncomfortable intersection of AI strategy and government execution. The models matter. The data matters. But the deciding factor is usually the machinery that turns R&D into operational capability.

The Navy’s ONR leadership change is a real opportunity to tighten that machinery—especially for autonomy and AI-enabled sensing—without hollowing out the science that keeps the U.S. ahead in the Indo-Pacific competition.

If you’re trying to modernize defense AI in 2026, the target isn’t “more pilots.” It’s shorter cycles, clearer gates, and better transitions.

If your team is working on AI for surveillance, autonomy, cyber defense, or mission planning—and you want a realistic path from prototype to fielded system—let’s talk about what that transition plan should look like before the next budget cycle locks.