There are sections of a grant application AI should not write. Here's how we decide.

Compose can draft any section of any application. The interesting product question is not what it can do. It is where we decided to stop it, and how we decided.

Our rule is a risk ratio, and we apply it section by section: how much time does auto-drafting save, versus how much damage does a plausible error do, weighted by how likely that error is to be caught before submission?

That third term matters more than people expect. An error that a reviewer will obviously catch is annoying. An error that no one internal can detect, because the only person who would know the truth is a community partner who will never read the document, is a different category of problem entirely.

Budget narratives score well. The facts live in the library, errors are checkable against numbers that exist, the reasoning is mechanical, and the time saved is substantial — this is often the single most tedious section to produce. Methodology sections and prior-award summaries score well for the same reasons: verifiable inputs, detectable errors, high drudgery.

Organizational capacity statements score moderately. The facts are mostly in the library — staff counts, tenure, credentials, infrastructure — but the framing requires judgment about what this specific funder values. We draft these and mark them for heavier review.

Community context scores badly, and it is the clearest case in the product. The facts are not in any document. Errors are invisible to a reviewer who was not in the room. And the damage — a program officer recognizing a mischaracterization of a community they have funded for fifteen years — is severe, immediate, and very hard to repair, because it does not read as a software error. It reads as an organization that does not know the people it claims to serve.

Lived-experience narrative scores worse. A model writing in the voice of a participant is producing testimony that did not occur. We will not do it, and the refusal is not configurable.

Letters of support are a related case. We help structure and format them; we will not generate the substance of a third party's endorsement of you. That is somebody else's claim to make.

So in those sections Maven offers structure and questions instead of prose. What changed in the last twelve months. Who told you. What exactly did they say. What did the organization do in response. Which of those things can you point to evidence for. The writer supplies the substance and the interface makes it fast to assemble — but the sentences originate with a person who was there.

There is a version of this argument that is purely about accuracy, and that version is incomplete. The deeper issue is provenance. A funder reading a community needs section is not only checking whether the claims are true. They are inferring whether this organization has real relationships in the community, and the texture of the writing is the evidence. Generated texture is a false signal about the thing the funder is actually trying to assess.

We also apply a lighter version of the rule to editing rather than drafting. Maven will tighten, cut, and flag inconsistency in a human-written community section, because those operations cannot invent facts. Rewriting the whole paragraph can, so the interface makes tightening easy and full rewrite deliberately awkward.

None of these lines are permanent. As the library gets richer — once an organization's own interview notes, site visit reports, and partner correspondence are indexed and citable — some of what is currently unsafe becomes safe, because the model would be assembling from real sources rather than generating from priors. The rule stays the same; the inputs change what it permits.

This costs us a demo moment. Competitors show a full application generated from a two-line prompt, and ours stops in the middle to ask who you talked to. We have lost evaluations over it.

It is still the right trade. The tool is being used on documents that carry an organization's reputation with the institutions that fund its existence. Being slightly less impressive in a demo is a very cheap price for never being the reason a fifteen-year funder relationship ends.