AI Can Fix Defense R&D’s Broken Small-Business Pipe

In the last two decades alone, the Defense Department awarded $27.5 billion in SBIR Phase I and Phase II funding to 8,945 companies. That sounds like a thriving innovation engine—until you look closer: 15% of that money ($4.1B) went to just 25 firms, and many of them behave like professional grant-catchers rather than technology transition partners.

That’s not a “startup problem.” It’s a systems problem. And it’s a national security problem.

In this AI in Defense & National Security series, we talk a lot about mission outcomes—faster sensing, better intelligence analysis, resilient cybersecurity, safer autonomous systems. But none of that matters if the defense R&D pipeline rewards the wrong behaviors. The uncomfortable truth is that the Small Business Innovation Research (SBIR) program—designed as an on-ramp for new entrants—often functions as a bureaucracy exam that favors incumbency.

Here’s the stance I’ll take: AI won’t fix SBIR by itself, but AI can make SBIR measurable, harder to game, and dramatically more aligned to operational outcomes.

The SBIR bottleneck isn’t funding—it’s incentives

The core issue isn’t that SBIR lacks money. It’s that the program’s structure creates predictable winners.

SBIR is supposed to work like a pipeline:

• Phase I: small awards (often $40K–$150K) to prove feasibility
• Phase II: larger awards to build and test prototypes
• Transition: follow-on procurement (the part that actually matters)

On paper, it’s rational. In practice, too many projects stall after Phase II, and too many awards go to firms optimized for proposal throughput instead of fielded capability.

The data highlighted in the source article is the kind that should trigger an immediate redesign conversation:

• Less than 4% of DoD Phase I funding over ~20 years went to companies with no prior federal experience.
• Some “SBIR mills” win across dozens of unrelated topics, which is a red flag for depth, not a badge of genius.
• Several large award recipients show weak transition performance (e.g., high SBIR totals with relatively small subsequent procurement).

If you want AI advantage in defense, you need a pipeline that promotes technical merit and adoption—not bureaucratic fluency.

Why this gets worse in 2025–2026

SBIR lapsed on Sept. 30, 2025, and reauthorization proposals are on the table. That timing matters. Congress is being asked—right now—to decide whether SBIR remains a “spread the money” program or becomes a “ship real capability” program.

And as defense AI becomes more central (from target recognition to cyber defense to logistics optimization), SBIR will increasingly be a feeder into AI-related mission systems. If the feeder is biased toward incumbents and repeat winners, the downstream AI ecosystem will be too.

“Small business” rules that invite large-firm behavior

A program can’t meet its mission if its definitions don’t match reality.

The article points out a common-sense mismatch: firms can qualify as “small” based on employee count (often under 500 employees) even if their revenue is massive. Many reasonable observers would not call a company worth hundreds of millions “small,” especially in a program explicitly meant to help true small firms get started.

This is where policy and technology need to meet. Policy sets the rules; AI gives you the ability to enforce the intent of the rules at scale.

What AI can do immediately: eligibility and concentration monitoring

Even without changing a single statute, the government could apply AI-enabled analytics to detect concentration and eligibility anomalies earlier:

• Entity resolution: identify firms that appear separate but share ownership, executives, addresses, or subcontracting structures.
• Award concentration alerts: flag topic areas where awards cluster among repeat winners in a way that suppresses competition.
• Revenue/size risk scoring: combine public signals (and verified submissions where allowed) into a “likelihood of being truly small” score.

This isn’t about banning success. It’s about preventing a taxpayer-funded program from becoming a business model.

Snippet-worthy rule: If SBIR is the product, transition is the metric—and AI can make transition measurable.

The real “valley of death” is measurement failure

Defense innovators talk about the valley of death like it’s a law of physics: prototypes die between R&D and procurement. But the article makes a sharper point:

When program offices aren’t required to track post-award outcomes, the system naturally optimizes for compliance, not adoption.

If the only enforceable outputs are milestones and reporting, then the winning strategy is paperwork excellence. That’s how you get a program where some firms can repeatedly win Phase I/II awards across wildly different domains.

AI can turn transition into a first-class KPI

To make SBIR serve national security, program managers should be evaluated on outcome metrics that map to actual capability delivery.

AI helps because it can fuse messy, cross-system data into usable scorecards:

• Phase I/II awards (topic, amount, sponsor)
• Follow-on contracts and options (what got purchased, how often)
• Time-to-first-procurement after Phase II
• Operational uptake proxies (integration into programs of record, test ranges, operational exercises)
• Cyber and supply-chain risk signals (for AI systems especially)

Then you publish a clear, standardized set of metrics—internally at minimum, publicly where feasible.

Once transition metrics are visible, it becomes much harder for “high-volume proposal shops” to hide behind activity.

A practical metric set (simple enough to adopt)

If you’re looking for a realistic starting point, I’d use three numbers per vendor:

1. Transition ratio = non-SBIR procurement dollars within 5 years ÷ SBIR Phase I/II dollars
2. Transition speed = months from Phase II award to first follow-on procurement
3. Fielding signal = count of deployments/integrations (even if partial), validated by program offices

You don’t need perfect measurement to get major behavior change—you just need measurement that affects awards.

Where AI fits in defense R&D modernization (beyond dashboards)

AI can do more than score performance. It can reduce the “bureaucratic literacy test” that keeps new entrants out.

The article describes solicitations as dense, contradictory, and not written in plain English. That’s a solvable problem.

1) AI-assisted solicitations that are readable and testable

Program offices could maintain a “solicitation twin”:

• Plain-language summaries generated from the official text
• Automatic checks for contradictions and missing evaluation criteria
• Examples of what “good” looks like (redacted, generalized)

This reduces accidental exclusion of capable nontraditional companies—especially AI startups that aren’t staffed to decode federal prose.

2) AI-enabled source selection that rewards substance

Source selection should still be human-led. But AI can help evaluators focus on technical merit by:

• Detecting template recycling across proposals
• Highlighting unsupported claims (“we will achieve X”) vs evidence (“we achieved X in test Y”)
• Mapping proposals to prior results and transition history

Used correctly, this doesn’t automate judgment—it improves it.

3) AI-ready transition paths for AI systems

AI projects die after Phase II partly because AI doesn’t “transition” like hardware. You need:

• Data access agreements
• Model evaluation plans
• Security testing
• MLOps pipelines
• Continuous monitoring

A better SBIR approach for AI in national security is to require a Transition Package by late Phase II:

• Target system and integration owner identified
• Data rights and model cards documented
• Test & evaluation plan (including red-teaming)
• Cyber posture and supply-chain risk assessment

If you want mission planning AI, cyber defense AI, or autonomous targeting support, this is the unglamorous work that makes it real.

A reform agenda that’s tough, fair, and measurable

The source article argues for three structural reforms:

• Treat “small” as under $40M annual revenue
• Cap lifetime Phase I/II funding at $75M per firm
• Reserve a portion of Phase I awards for new entrants

I agree with the direction. It’s hard to defend a “small business” pipeline where a handful of repeat winners can dominate for decades.

But I’d add a fourth reform that’s specifically relevant to AI in defense:

Make “transition intent” part of award eligibility

A Phase I proposal should be non-starter unless it includes:

• A named government problem owner (not just a topic)
• A plausible path to procurement (contract vehicle, program office, or operational unit)
• Clear data and integration assumptions

This doesn’t punish early research. It forces honesty.

Memorable line: If nobody can buy it, the government shouldn’t fund it—especially for AI.

What leaders can do next (even before Congress acts)

If you’re a program executive, innovation office leader, or prime responsible for partnering, you can improve outcomes without waiting for a perfect reauthorization.

Here are five moves that consistently work:

1. Publish a transition scoreboard for your portfolio (even internally). If a vendor wins repeatedly but doesn’t transition, stop pretending that’s success.
2. Create a new-entrant lane with simplified Phase I applications, shorter proposals, and evaluator training that resists “familiarity bias.”
3. Standardize AI transition artifacts (model cards, evaluation reports, security documentation) so Phase II doesn’t end in ambiguity.
4. Adopt concentration thresholds (topic-level and vendor-level) that trigger review when the same names keep winning.
5. Use AI to reduce paperwork load, not add to it—auto-fill compliance, summarize requirements, and help small teams submit strong proposals.

These are unsexy changes. They’re also the kind that shift billions in outcomes over time.

Defense AI advantage depends on fixing the on-ramp

The U.S. doesn’t have an “idea shortage” in defense technology. It has a throughput and adoption shortage—and the SBIR pipeline is one of the biggest levers available to fix it.

If SBIR reauthorization keeps rewarding proposal factories, we’ll keep getting prototypes that never field. If SBIR becomes outcome-measured, new-entrant-friendly, and designed for AI-era transition realities, it can produce what the program promised decades ago: a steady flow of operationally relevant innovation.

This series is about AI in Defense & National Security, but the real theme is institutional competence. AI systems don’t matter if the institutions that fund them can’t distinguish progress from paperwork.

So here’s the forward-looking question worth sitting with: When the next conflict tests our speed of adaptation, will our R&D programs behave like a learning system—or like a grant distribution machine?