Fix SBIR to Field Defense AI Faster (Not More Demos)

Defense AI doesn’t fail because we can’t build models. It fails because the system that’s supposed to move useful tech from a small business into a program of record is built to admire prototypes, not deploy capability.

That’s why the recent call for a “clean sheet” redesign of the Small Business Innovation Research (SBIR) program lands so hard. The argument isn’t “SBIR is bad.” It’s sharper: SBIR’s current plumbing creates a predictable valley of death where promising AI, autonomy, and cyber tools stall—right when operators need them most.

This post is part of our AI in Government & Public Sector series, where we focus on the real-world mechanics of getting AI adopted responsibly inside government. Here, the mechanics are the story: budgets, incentives, topic-writing, contracting pathways, and who is accountable when a pilot never becomes a product.

The real SBIR problem: transition is nobody’s job

SBIR’s biggest flaw is that it funds invention while treating transition like a nice-to-have. In defense acquisition, “transition” means a program office budgets for, contracts, integrates, tests, and sustains a capability. If nobody owns that work, your Phase II success becomes a Phase II museum piece.

The source article uses a software metaphor—too many patches, too much technical debt. That’s accurate. But the deeper issue is governance: SBIR is structured so that the people who control long-term outcomes (program managers) can ignore it, while the people who run SBIR cycles can declare success without fielding anything.

Why the valley of death hits AI harder than other tech

AI in national security is unusually sensitive to this funding and ownership gap because deployment isn’t a single event.

For many AI defense applications—intelligence fusion, cyber analytics, targeting support, logistics optimization, autonomous systems—fielding requires:

• Data access agreements and authority to operate (ATO) pathways
• Integration with mission systems and networks (often classified, often brittle)
• Human-in-the-loop workflows and training
• Monitoring, drift detection, and model updates
• Sustainment funding (compute, labeling, red-teaming, re-validation)

A program that can fund a promising model but can’t fund integration and sustainment is basically funding a demo. Most SBIR “wins” die in the handoff from R&D money to procurement/operations money.

What a clean-sheet SBIR redesign gets right

A clean-sheet SBIR redesign should keep the one thing SBIR does uniquely well: create an on-ramp for non-traditional small businesses. Defense is still dominated by programs of record and incumbent primes. SBIR remains one of the few paths where a small team with a real idea can get a serious at-bat.

The article proposes keeping the statutory Phase III authority (the practical contracting mechanism that allows sole-source follow-on awards). That’s smart. Phase III is the closest thing SBIR has to a “deploy now” button.

But the redesign’s most important move is this:

Put transition money inside the program offices—where delivery is accountable.

The “10% transition fund” is the forcing function

The proposal: each acquisition program creates a dedicated, “colorless” transition line equal to 10% of its R&D budget to pull SBIR technologies into operational use.

This is a big deal for three reasons:

1. It creates a receiving dock. Transition fails when there’s nowhere for tech to land. A ring-fenced transition line is a landing zone with real dollars.
2. It changes program manager behavior. Program managers optimize for cost/schedule/performance. Give them a pot of money they must execute on tech insertion and they’ll suddenly care about market research and integration.
3. It turns “pilot culture” into portfolio culture. Instead of celebrating how many Phase I awards were made, success becomes “how many got inserted into programs that matter.”

From a defense AI angle, a program-side transition budget is how you pay for the boring necessities that make AI real: data pipelines, security controls, integration engineering, test harnesses, and sustainment planning.

AI can improve SBIR topic selection—but only if incentives change

Using AI to automate topic development sounds great until you remember: topics are politics in sentence form. If topic-writing remains a closed ritual, automation will just accelerate bias and vendor capture.

The clean-sheet proposal argues for AI-assisted topic development with human review, using plain-language problem statements and end-user involvement (including open voting/crowdsourcing).

That direction is right, but it needs guardrails.

What “AI-assisted topics” should look like in practice

A practical, defensible workflow for AI in government topic generation would include:

1. Problem intake in operator language (what’s breaking, where, and at what tempo)
2. AI clustering to group similar pain points across units and commands
3. De-biasing checks to detect overly specific phrasing that maps to a known vendor’s solution
4. Human adjudication by a mixed panel (operators + program office + contracting + security)
5. Public traceability: why a topic exists, who owns it, what “transition” means for it

Here’s the standard most organizations avoid saying out loud: if a topic can’t name its transition owner and a plausible system where it will live, it’s not a topic—it’s a science fair prompt.

Why end-user voting helps, and where it can go wrong

End-user voting pushes SBIR toward real mission pain. It also risks becoming a popularity contest.

The fix is to treat voting as signal, not selection. The selection criteria should weight:

• Operational urgency
• Integration feasibility (systems, networks, security)
• Transition ownership (named program office)
• Sustainment plan (who pays in years 2–5)

In other words: let operators steer the spotlight, but force program offices to sign the receipt.

Consolidate “innovation offices” and judge them by fielded capability

Defense innovation organizations are often measured by activity, not outcomes. Number of topics published. Number of pitch days run. Number of prototypes funded.

The clean-sheet redesign calls this what it is: innovation theater. It also proposes consolidating fragmented innovation entities into a single institution with a clear mandate to invest and transition.

I’m opinionated here: consolidation is less important than accountability. If you merge ten organizations into one but still measure success by throughput, you just created a larger machine for producing pilots.

Better metrics for innovation offices (especially for AI)

If the goal is leads and outcomes—capability delivered, not vibes—use metrics that are hard to fake:

• Transition rate: percentage of Phase II efforts that reach Phase III and operational use
• Time-to-Phase III: median months from award to initial operational contract
• Integration success: number of systems with validated integration (not “demoed”)
• Sustainment adoption: funding committed for year-2 operations (compute + updates)
• Operator satisfaction: post-fielding feedback tied to mission outcomes

The article proposes disqualifying companies and government technical points of contact that fail to maintain a 10% transition rate. That’s aggressive—and it will be controversial—but the basic idea is right: if nobody fears consequences for non-transition, non-transition becomes the default.

A Phase III marketplace could be SBIR’s missing product layer

Most program managers don’t ignore SBIR because they hate innovation. They ignore it because contracting is slow and risky. If choosing an SBIR-derived solution feels like choosing paperwork and schedule pain, they’ll choose the incumbent.

The proposed answer: a centralized, AI-enabled Phase III marketplace that lists awarded projects and helps acquisition teams generate Phase III contracts quickly.

Done well, this can fix a real bottleneck:

• Discovery: program offices can’t easily find what already exists
• Due diligence: they can’t quickly assess fit, security posture, integration requirements
• Contract generation: they drown in templates, reviews, and justifications

What the marketplace must include for defense AI

For AI in national security, a marketplace can’t be a glossy catalog. It should behave like an acquisition-grade “app store” with compliance built in:

• Security artifacts (baseline ATO package components, control mappings)
• Data requirements (what data is needed, classification, access model)
• Integration profile (APIs, mission systems, network boundaries)
• Model facts (training data lineage, evaluation results, known limits)
• Sustainment plan (update cadence, monitoring, compute footprint)

If you want AI adoption in government, reduce the friction of picking a credible tool while increasing the friction of buying another forever-contract from the usual supplier.

Practical guidance: how leaders can act now (without waiting for Congress)

Even if SBIR policy changes take time, organizations can start operating as if the clean-sheet model already exists.

If you’re a program manager

Treat transition like a budget line, not a wish. Build a mini “10% transition fund” internally and use it to create insertion pathways.

• Require every SBIR effort you touch to have an integration plan and named target system
• Budget for security, data access, and sustainment from day one
• Run market research early; don’t wait for the “perfect contract vehicle”

If you run an innovation office

Stop optimizing for awards. Optimize for removals of barriers.

• Publish plain-language problem statements tied to a program of record
• Bring contracting and cybersecurity into topic shaping, not just award execution
• Track transition metrics publicly inside your organization

If you’re a small business building defense AI

Design your SBIR strategy around Phase III from the first slide.

• Show how you’ll integrate into existing systems and workflows
• Package security and documentation as a product feature
• Price sustainment honestly (models aren’t “done” after delivery)

A sentence I’ve found useful: “We’re not selling a model. We’re selling a maintained capability under operational constraints.”

Where this lands for AI in defense and national security

A clean-sheet SBIR redesign matters because the U.S. can’t afford an innovation pipeline that ends in prototypes. Near-peer competition, ongoing cyber pressure, and fast-moving autonomy trends don’t reward bureaucracy that confuses activity with readiness.

The best part of the proposal is also the simplest: move transition funding and responsibility to program offices, then build systems (like an AI-enabled Phase III marketplace) that make it easy to buy and field what works. That’s how you get from “interesting demo” to “capability operators rely on.”

If you’re working on defense AI—whether you’re inside government or supplying it—ask yourself one blunt question: Who has the budget and authority to keep this capability alive 18 months after the pilot ends? If the answer is fuzzy, the project is already sliding into the valley of death.