What was broken
An AI-native startup needed early non-dilutive capital to take its platform to market, and had to justify the ask against a credible utilisation plan.
What we did
CapEasy structured the seed-fund application — a fund-utilisation plan tied to product and platform milestones, market sizing, and multi-year financial projections. We aligned the proposal to the program’s eligibility and permitted-spend rules.
Where it landed
The startup secured a public seed-fund approval. A seed-fund approval is the committee’s sanction; actual disbursement follows the program’s due-diligence and compliance milestones and is not guaranteed.
Where AI and automation startups actually find non-dilutive money in the US
SBIR/STTR is not one program — it is eleven federal agencies running their own solicitations under a shared statute, and the fit is sector, not buzzword. NSF’s America’s Seed Fund is technology-agnostic by design: NSF 26-510 states the program "funds across nearly all technology areas and market sectors" rather than gating on a named topic, which makes it the natural first stop for an AI or automation platform that does not map cleanly to a single mission agency. A defense-facing automation product should also be checking AFWERX/DoD SBIR topics; a health-data model belongs on NIH’s SBIR track; an energy-systems automation play fits DOE. sbir.gov’s solicitation search is the single index across all eleven — the mistake is applying to the first agency you have heard of instead of the one whose mission matches the product.
NSF’s current ceilings, per NSF 26-510: Phase I is capped at $305,000 in R&D funding for a 6–18 month project, inclusive of direct and indirect costs, the small-business fee, Technical and Business Assistance (TABA) funding, and an optional Innovation Corps (I-Corps) allocation. Phase II goes up to $1,250,000, typically over 24 months. Cost-sharing is prohibited — NSF pays the award as stated, not as a match.
What the budget justification actually has to show for an AI applicant
NSF SBIR budgets are itemised, not lump-summed: personnel, TABA (up to $6,500 for commercialisation activities per seedfund.nsf.gov’s Phase I budget guidance), and everything else falls into "other direct costs" — the category cloud and GPU compute lives in. A reviewer reading an AI proposal’s budget is looking for the same thing an incubator committee looks for in a utilisation plan: does the spend map to a milestone, or is it a round number attached to a vague deliverable. Compute is the line item that most exposes the difference. A schedule of instance-hours or GPU-hours against a rate, tied to which model-training or evaluation milestone it funds, reads as a controlled R&D expenditure. A single "cloud costs" figure with no unit basis reads as padding — and NSF’s own budgeting guidance for applicants warns that reviewers who have seen hundreds of proposals spot both padding and under-scoping on sight.
The same discipline runs the other market's seed-fund file: the fund-utilisation plan had to tie every line of spend to a product or platform milestone before the committee would sanction it. A grant proposal is read as a milestone map with a price attached to each step, not as a request for a sum.
The accounting system a Phase I applicant needs — and what Phase II demands after it
A Phase I award does not require an audited accounting system, but it does require one that can already do two things: separate this project’s costs from the rest of the business, and produce time-and-effort records for everyone charging time to the award — sbir.gov’s accounting-system tutorial is explicit that timesheets are required "even if they are the president or CEO" and must record actual hours, not estimated percentages. NSF’s own guidance says the same in grant language: time-and-effort records have to exist so that salary charges can be verified.
Phase II raises the bar sharply, because the contract typically moves from fixed-price toward cost-reimbursement, and DCAA evaluates the accounting system against ten standards before funds move — segregation of direct from indirect cost, cost accumulation by contract, a consistent indirect-rate allocation method, exclusion of unallowable costs, and a working timekeeping system among them. An AI startup that treated Phase I compute spend as one undifferentiated cloud bill has no way to answer that review; one that tracked GPU-hours by project and milestone from day one already has the ledger DCAA is asking for.
Approval is the start of the file, not the end of it
A Phase I award is not a wire transfer with no further questions — NSF requires a technical narrative and a project outcomes report when the award period ends, filed through Research.gov, and a second formal report if the company applies for Phase II. That is the same shape as the outcome in the engagement: an approval is the committee’s sanction, and disbursement still runs through the program’s own due-diligence and compliance milestones. The file that got the approval — reconciled, milestone-tied, permitted-spend-compliant — is the same file the program keeps reading afterward.
What to take from it
- SBIR/STTR is eleven agency programs under one statute, not one grant — match the mission agency (NSF for general tech, AFWERX for defense, NIH for health, DOE for energy) instead of defaulting to the most familiar name.
- NSF SBIR Phase I tops out at $305,000 over 6–18 months, inclusive of TABA and the small-business fee; Phase II runs to $1,250,000 over about 24 months — check NSF 26-510 for the current solicitation before budgeting against last cycle’s numbers.
- Cloud and GPU compute belongs in the budget as a unit-priced, milestone-tied line item — hours or instances against a rate — not a single "cloud costs" figure; that is the line reviewers use to judge whether a budget is real.
- Time-and-effort records for every person charging the award, kept in actual hours from day one, are a Phase I requirement in practice and the exact thing DCAA checks before Phase II cost-reimbursement funds move.
- A grant or seed-fund approval is a committee sanction, not disbursement — funding bodies keep reading the file through milestone and compliance checkpoints after the award letter.