AI in Navy R&D: What ONR’s New Chief Should Fix

A decade is an eternity in AI.

That’s why the most alarming detail in recent discussions about U.S. defense research isn’t a particular model, drone, or sensor—it’s the timeline. Multiple government assessments and practitioners have flagged that Defense Department tech programs can take 10+ years to move from idea to fielded capability. In the time it takes to push a traditional program across the finish line, the commercial AI ecosystem will have shipped several major model generations, rebuilt toolchains, and shifted the entire market.

Against that reality, leadership changes at the Office of Naval Research (ONR) matter more than they usually get credit for. ONR sits upstream of the Navy’s future: autonomy, undersea sensing, cyber, electronic warfare, human performance, and the enabling infrastructure that makes AI in national security more than a slide deck.

Rachel Riley—reported as a former Department of Government Efficiency (DOGE) staffer and ex-McKinsey consultant—has been tapped as acting chief of naval research. Observers quoted in the source article see her as someone willing to cut through bureaucracy and force sharper accountability. That’s exactly the kind of disruptive posture Navy AI modernization needs. It’s also exactly the kind of posture that can go wrong if “efficiency” becomes a synonym for “indiscriminate cuts.”

The real constraint on Navy AI isn’t talent—it’s throughput

The Navy doesn’t lack smart people. It lacks throughput: the ability to translate promising research into deployable systems at a pace that matches operational urgency.

This matters because AI-enabled modernization is not one program. It’s a supply chain of capabilities:

• Data pipelines that survive contested networks
• Model development, evaluation, and security hardening
• Edge compute and power on ships, subs, and unmanned platforms
• Human-machine interfaces that fit real watchstanding workflows
• Rapid feedback loops from fleet operators back to developers

When that pipeline stalls—because requirements freeze too early, testing environments are scarce, approvals stack up, and funding locks into multi-year commitments—AI becomes “research” instead of “advantage.”

A blunt but accurate way to say it: an AI capability delivered two years late is often a different capability than the one you needed. Threats, sensors, and tactics don’t wait for milestone reviews.

Why ONR is uniquely positioned to help (and to hurt)

ONR’s charter and culture make it a hinge point. Done well, ONR can:

• Fund high-risk research that industry won’t touch
• Transition prototypes into programs of record (or into rapid fielding)
• Shape standards for evaluation, safety, and mission performance

Done poorly, ONR becomes a holding pattern—well-funded, busy, and decoupled from fleet outcomes.

The source article includes a critique you hear often in defense innovation circles: some research portfolios persist for years without producing fielded capability, while commercial firms build comparable tech faster because they’re forced to ship.

That critique isn’t anti-research. It’s anti-research without consequence.

A DOGE-style mandate can speed AI adoption—if it’s aimed correctly

A “DOGE veteran” label signals something specific: aggressive scrutiny of staffing, spend, and processes. In the Navy R&D context, that can be healthy—because many delays are self-inflicted.

But efficiency in defense R&D can’t be measured like a corporate back-office reorg. The Navy isn’t optimizing for quarterly earnings; it’s optimizing for wartime resilience.

So the right question isn’t “How do we cut?” It’s:

What should ONR stop doing so it can do more of what actually reaches the fleet?

What “strong leadership” should look like in ONR

If Riley (or any ONR leader) wants faster AI fielding, the moves that matter are structural:

1. Make timelines a first-class requirement. Every funded effort should have an explicit transition path and a time-bound decision point: scale, pivot, or stop.
2. Force “commercial-first” checks. Before starting bespoke development, require proof that the capability isn’t already available commercially—or that a commercial solution can’t be adapted securely.
3. Treat prototypes as a product, not a paper. If a program can’t be exercised in realistic conditions—at sea, in degraded comms, with fleet operators—it’s not ready for serious funding growth.

This aligns with the broader “AI in Defense & National Security” reality: modern AI progress comes from fast iteration, data feedback, and continuous evaluation. Long cycles break that.

The hardest part: separating sacred cows from strategic research

One risk in “speed and efficiency” campaigns is collateral damage: cutting exactly the research the nation can’t afford to lose.

The source article makes an essential point: ONR funds areas that won’t reliably attract commercial investment but are strategically vital—encryption, physical oceanography, marine geosciences, climate-related ocean science, and other domains tied to undersea dominance.

Commercial AI doesn’t naturally fund “undersea truth.” It funds ad targeting, logistics optimization, and consumer robotics.

So ONR needs a crisp portfolio split:

Bucket 1: Commercially available or dual-use (move fast)

These are areas where private capital and dual-use markets are already pushing the frontier:

• Autonomy stacks (planning, perception, multi-agent coordination)
• Computer vision for ISR
• Predictive maintenance and logistics
• Cyber analytics and anomaly detection
• Simulation tooling

For this bucket, ONR should act less like a lab and more like a disciplined integrator:

• Buy, adapt, and harden
• Run competitive bake-offs
• Push systems into fleet experimentation quickly

Bucket 2: National-security-unique (protect and fund patiently)

These are domains where America’s edge depends on sustained public investment:

• Undersea sensing physics and materials
• Anti-jam and low-probability-of-intercept communications
• Advanced cryptography and secure compute
• Ocean environment modeling tied to ASW outcomes

For this bucket, “efficiency” should mean better governance, not starvation.

A clean rule I’ve found useful: if a project’s payoff depends on classified data, specialized test ranges, or undersea platforms, you won’t “market your way” into it. That’s ONR territory.

AI modernization needs new contract mechanics, not just new org charts

Org charts are visible. Contract structures are where time goes to die.

If ONR wants to accelerate AI-enabled modernization, it needs contracting patterns that match how AI is built and maintained.

What to change: from one-time deliverables to continuous performance

Traditional R&D contracts often reward deliverables like reports, documentation, or isolated demos. AI systems don’t behave like that in the real world. They drift. Data changes. Threats adapt.

ONR can set a higher bar by structuring work around measurable model performance and operational constraints, such as:

• Detection/false alarm rates under specified conditions
• Robustness to environmental shifts (sea state, clutter, sensor degradation)
• Compute and power limits on target platforms
• Resilience to jamming and spoofing
• Time-to-update models and revalidate them

If a contractor can’t show improvements against those metrics on a regular cadence, funding shouldn’t auto-renew.

The “Anduril test” for research programs

A line from the source article captures a frustration in the defense innovation community: some research efforts look like things a venture-backed defense firm would happily build into a product line.

That suggests a simple governance check:

• If a credible commercial defense company could ship a version in 12–24 months, ONR shouldn’t run it as a decade-long research program.

Instead, ONR should:

• Fund rapid prototypes
• Buy initial units
• Pay for integration and test
• Transition to a program office once utility is proven

That’s how you get AI into the fleet without turning every capability into a science project.

What success looks like in 12 months (practical milestones)

Leadership conversations get abstract fast. Here are concrete signals that ONR is accelerating AI adoption rather than just reorganizing it.

1) A published “stop list” and “scale list”

ONR shouldn’t be shy about ending efforts that don’t transition. A public-facing version (sanitized as needed) creates credibility.

• Stop list: programs ended due to lack of operational pull, redundant commercial alternatives, or stalled progress
• Scale list: programs graduating to larger funding because they met performance thresholds

2) A standard AI evaluation pipeline for the Navy

The Navy needs repeatable evaluation, not one-off demos. Within a year, ONR can establish baseline practices:

• Red-team testing for AI systems (adversarial conditions)
• Data provenance and labeling standards
• Reproducible test harnesses
• Model cards and risk documentation suitable for operators

This is the unglamorous part of AI in national security, and it’s where serious organizations separate themselves from hype.

3) Fleet-facing experimentation capacity, not just lab capacity

If sailors and Marines can’t try systems early, feedback arrives too late. ONR should expand pathways where prototypes are exercised by real units—especially in autonomy and decision-support.

When experimentation is routine, the “requirements” conversation changes. Operators stop asking for fantasy specs and start asking for version 0.3 improvements.

What this means for defense tech companies (and lead buyers)

If ONR leadership brings sharper scrutiny and tighter transition expectations, vendors will feel it immediately. That’s good news for companies that can ship real capability—and bad news for anyone relying on vague R&D deliverables.

Here’s how to position well:

• Bring evidence, not promises. Show performance in realistic conditions, including failure modes.
• Design for the edge. If your AI needs a pristine cloud connection, it’s not a Navy system yet.
• Offer upgrade paths. Make it easy to retrain, patch, and revalidate models.
• Speak “mission outcomes.” “Accuracy” is less compelling than “reduced time to classify contacts” or “fewer false alarms per watch.”

For government stakeholders, the buying posture should shift too: insist on frequent, measurable progress, and stop paying for inertia.

The bet: faster AI adoption without sacrificing strategic science

ONR’s new leadership moment is really a stress test for the whole Navy modernization story. Everyone says they want AI for decision advantage, autonomous systems, and resilient operations. The question is whether the R&D enterprise can move at the speed the strategy assumes.

A DOGE-aligned reform mindset can help—if it’s applied with precision. Cut duplicative efforts. Force commercial-first checks. Demand measurable progress. Then protect the research that only the nation will fund.

If you’re building, buying, or integrating AI in defense and national security, now is the time to pressure-test your assumptions about timelines, evaluation, and transition. The Navy doesn’t need more AI pilots that never leave the lab. It needs capabilities sailors trust when networks degrade and stakes rise.

If your organization is trying to move an AI capability from prototype to program—or struggling with evaluation, data readiness, or edge deployment—I’m happy to compare notes on what a practical transition plan looks like. What would you change first: contracting, test infrastructure, or portfolio governance?