Grant Writing for AI Startups: What SBIR Reviewers Actually Look For

SBIR reviewers score your proposal against published criteria, usually three of them: technical merit and feasibility, qualifications of the team, and commercial potential. A reviewer with a stack of proposals and a scoring rubric is not reading for vision. They are looking for the specific sentence that tells them whether the thing you propose can be shown to work or not work inside the Phase I period.

Most technical founders lose points not because the work is weak but because they wrote the proposal like a pitch deck. A pitch deck removes risk. A Phase I proposal is a document about risk. Those are opposite genres, and the reviewer is reading in the second one.

Phase I exists to de-risk, so say what the risk is

The program structure is public on sbir.gov: Phase I is a small feasibility award, typically six to twelve months, intended to establish technical merit and feasibility. Phase II is the larger development award, and eligibility for it is generally conditional on having completed a Phase I. Phase III is commercialization with non-SBIR money.

Read that again with the emphasis on feasibility. The government is buying an answer to a question it does not currently have. If your proposal implies the question is already answered, you have argued yourself out of the award: there is nothing left to fund. Founders do this constantly, because every other document they write is optimized to project certainty.

A reviewer wants to see a named technical risk, a stated mitigation, and a measurable milestone whose failure would be visible. Admitting a hard problem and describing the experiment that resolves it scores higher than claiming there is no hard problem. The second version reads as either naive or evasive, and both are score reductions.

Rewrites

Here is the shape of the edit, applied to claims of the kind that appear in real AI proposals.

Technical claim

Weak: "Our novel architecture will dramatically improve retrieval accuracy for enterprise AI applications through advanced semantic understanding."

There is no proposition here. Nothing in that sentence can come out false, which means nothing in it can come out true, which means a reviewer cannot score it. Also "dramatically" and "advanced" are doing all the work and neither is a quantity.

Strong: "We hypothesize that adding bitemporal validity intervals to a vector retrieval index will raise recall@10 on time-sensitive queries from a measured baseline of 0.48 to at least 0.70. Risk: the interval filter may degrade approximate-nearest-neighbor recall by forcing post-filter over-fetch. Mitigation: partial HNSW indexes per validity epoch, evaluated against a pre-filter baseline. Phase I milestone: a 500-query labeled benchmark and a measured recall figure for both approaches, reported whether or not the hypothesis holds."

That version can fail. That is the point. It names the metric, the baseline, the target, the thing that could go wrong, and what gets delivered either way.

Market claim

Weak: "The AI memory market is projected to reach $12B by 2030, and capturing even 1% would represent $120M in revenue."

The unsupported market claim is one of the most reliable ways to lose commercial-potential points. The "even 1%" move is a known tell, because it substitutes arithmetic for a go-to-market argument. It also usually cites a number with no source, or a source that is a press release about a report nobody on the panel can read.

Strong: "Our initial segment is the estimated 40,000 organizations that have deployed at least one MCP-compatible assistant internally, identified via public integration directories. We have conducted 22 discovery interviews; 14 described re-explaining project context across tools as a recurring weekly cost, and 6 estimated it at over two hours per engineer per week. Two have signed letters of intent, attached. Our Phase II commercialization target is 40 paid organizational deployments, derived from a 5% conversion assumption on our current 800-person waitlist."

Smaller numbers, cited, with the assumption stated as an assumption. A reviewer can check it. Checkable beats large.

The unstated-assumption kill

The other common failure is subtler and it happens inside the work plan. A milestone that silently depends on something you do not control is a milestone the reviewer cannot believe. Three recurring versions:

• A hire."Month 2: our ML engineer implements the retrieval evaluation." Is that person employed today? If the plan requires recruiting a specialist in a competitive market during a six-month award, say so and state the fallback: a named consultant, a letter of commitment, a reduced-scope path.
• A partner."We will validate on production data from a design partner." If there is no executed agreement, the milestone is contingent on a negotiation. Attach the letter or describe the synthetic-data alternative.
• A dependency shipping."We will use the forthcoming spec revision." Standards slip. Specify the version you will build against today.

Write the work plan and then, for each milestone, ask what has to be true for this to happen on this date. Every answer that is not fully inside your control goes in the proposal explicitly, with a mitigation. Reviewers do not penalize acknowledged dependencies. They penalize the ones they find themselves.

Mechanics that cost people awards

Read the specific solicitation, not general advice about SBIR, including this post. The topic areas are narrow and the agencies differ substantially: NSF runs a project-pitch prescreen before you may submit a full proposal, DoD topics are tied to stated mission needs and a proposal that does not map onto one is out regardless of quality, NIH has its own study-section culture. Registration alone (SAM.gov, an assigned UEI, agency portals) takes weeks and the deadline does not move for you.

Write for a reviewer who knows the field but not your product, has limited time per proposal, and is filling in a numeric score per criterion. Put the falsifiable claim, the risk, and the milestone in the first page. If the abstract does not contain a number, rewrite the abstract.

We are working through this process for Unimatrix ourselves, which is the only reason this post exists. The prompts we used to draft and stress-test proposal sections, including one that just hunts for unstated assumptions in a work plan, are in the free prompt library.