Grantee due diligence that quotes its source, every flag

Deterministic checks run identically for every grantee

Legal status, registration, required attachments, and budget arithmetic get the same pass for every prospective grantee. Ineligible or non-compliant status is flagged with the evidence beside it. The queue order and the award size do not change the rubric. You get a consistent baseline before any human reads a word.

Every red flag quotes the document it came from

A flag is not an AI paraphrase of a filing. It is the quoted passage from the Internal Revenue Service (IRS) Form 990, the registration record, or the application, checked against the source. When a record does not address a diligence criterion, the gap is flagged, not guessed. You can trace each finding back to the page it came from.

Governance and reputational judgment goes to a person

Subjective calls are the ones that get a foundation in trouble. Cemented AI does not decide them alone. Any governance, conflict-of-interest, or reputational assessment routes to a reviewer with the criterion and the quoted text attached. The model gathers and quotes the evidence. A person makes the call. Every non-deterministic decision in the packet carries a name.

A packet that is defensible to a board and to declined applicants

When a grant is declined or a flag is questioned, you can show the passage behind every finding. There is no black box to explain away. The same evidence trail that satisfies your board also answers a declined organization that asks why.

How a review runs

1. Step 1

   Add the prospective grantee

   Upload the application, the IRS Form 990, and any registration or governance records you hold for the organization.

2. Step 2

   Run the deterministic checks

   Legal status, good-standing, completeness, attachments, and budget arithmetic get the same pass every grantee gets, with evidence attached to each result.

3. Step 3

   Route the judgment calls

   Governance, conflict-of-interest, and reputational questions go to a reviewer with the criterion and the quoted source text attached. A person makes the call.

4. Step 4

   Deliver the packet

   Each finding quotes the filing or record behind it. The packet is ready to defend to your board and to declined applicants.

Common questions

What does a deterministic check actually cover?

Legal and tax status, registration and good-standing, required attachments, completeness, and budget arithmetic. These run the same way on every prospective grantee, with the evidence attached to each result. Anything that calls for judgment routes to a reviewer instead.

Does an AI decide whether to fund an organization?

No. Cemented AI gathers and verifies evidence. It never decides a subjective governance or reputational call on its own. Every judgment call routes to a person with the criterion and the quoted source attached. A human owns each non-deterministic decision.

How do I know a flagged risk is real and not a hallucination?

Every flag quotes the source filing or record, and the quote is checked against that source. If a record does not address a criterion, that gap is flagged as missing rather than guessed. You verify each finding by reading the passage behind it.

Can it pull records from a grants management system?

Connectors to grants management and intake systems are an Enterprise conversation. Ask about connectors for Enterprise and the scope your foundation needs. Uploaded content stays scoped to your account and is not used to train frontier models.

How do I get started, and what does it cost?

Start on the Free plan to run diligence on a few prospective grantees and read the quoted findings yourself. Pro and Max add more volume. Enterprise covers connectors and larger teams. See the pricing page for current plan details.

My team has no dedicated quality-assurance staff. Is this still workable?

Yes. The deterministic checks give a mid-sized team a consistent baseline without added headcount. The human-in-the-loop step puts judgment calls in front of an existing program officer, with the evidence already gathered. The reviewer reads quoted text, not raw filings.