Methodology · The AI + human split that wins

AI handles the research.
A human writes the narrative.
That's the split that wins.

Reviewers can spot generic AI prose. The pattern is documented: passive voice, hedge words, absence of program-specific detail. NIH now refuses applications "substantially developed by AI." NSF requires disclosure. The agencies are not converging — they're diverging.

The winning approach in 2026 is clear: AI does the heavy lifting on structure, research, rubric compliance, and data synthesis. A human writes the narrative that carries the mission. That's how I work. Every proposal I ship runs through a 6-stage pipeline with this exact split.

The split

What AI does. What the human does.

The line is clear. Crossing it in either direction hurts the proposal — AI writing narrative sounds generic; human-only research is slow and incomplete.

⚡ AI handles

Funder research — cross-referencing 990 data, past awards, stated priorities, and giving patterns against your org's profile
Rubric mapping — extracting every scoring criterion from the RFP and mapping it to a coverage matrix before any narrative is drafted
Evidence synthesis — digesting your prior reports, impact data, financials, and program docs into structured evidence blocks the narrative can cite
Compliance checking — flagging eligibility gaps, missing documentation, word-count violations, and formatting errors against the funder's spec
Budget line traceability — linking every budget item back to a named program requirement so the justification is auditable
Adversarial review — a separate AI pass simulates a skeptical program officer, surfaces objection points, and flags rubric gaps before submission

Why this works: These tasks are data-matching at scale. A human doing rubric compliance by hand takes 8–12 hours per application and still misses items. AI does it in minutes and catches the missed eligibility documentation that would have rejected the application on technical grounds.

✍ Human writes

The narrative — project description, needs statement, significance section, specific aims. The parts reviewers form their judgment on.
The story — why THIS organization, doing THIS work, for THESE people, at THIS moment. AI can't write institutional voice. A human carries the mission.
The judgment calls — what to emphasize, what to cut, where to take a position vs. hedge. Reviewers reward conviction; AI defaults to balanced hedging.
The relationship framing — how this proposal connects to the funder's stated priorities in language that feels like partnership, not compliance.
The revision response — when a funder asks for changes, the response requires human reading of subtext, not just literal editing.

Why this matters: Stanford Medicine's published guidance says it directly: "AI is a brainstorming and editing partner, not a substitute for the investigator's voice or scientific judgment." The agencies are watching. The narrative is where you earn the grant — and it has to sound like your organization, not like a chatbot.

The 6-stage pipeline

Every proposal runs through six audited stages.

Two human gates. Four AI-augmented production stages. Nothing ships until the final stage clears.

Human gate

Stage 01

Intake

30-minute scoping call. We agree on the test criteria and the funder profile before the timer starts.

Stage 02

Evidence synthesis

AI digests your prior reports, impact data, and program docs into a structured evidence base — the foundation every narrative section will cite.

Stage 03

Rubric mapping

Every scoring criterion extracted from the RFP, mapped to a coverage matrix. Every required element named and assigned before drafting.

Stage 04

Narrative draft

Human-authored narrative — needs statement, project description, logic model — built against the rubric map with cited evidence blocks.

Stage 05

Adversarial review

A skeptical-reviewer AI pass surfaces objection points, rubric gaps, and unstated assumptions before a real reviewer sees them.

Human gate

Stage 06

Ship-ready package

Final consolidation in the funder's required format — Word, PDF, portal text blocks. Rubric-match audit + modular reuse pack included.

The 2026 compliance landscape

Every agency has different AI rules. I track them so you don't have to.

The disclosure and originality rules are not converging. Each agency has its own stance. My pipeline is built to comply with all of them — because the same proposal structure works regardless of the funder's AI policy.

NIH

Strictest. Refuses AI-heavy applications.

NOT-OD-25-132: refuses applications "substantially developed by AI." Caps each PI at six submissions per calendar year. Reviewers are trained to spot generic AI prose. Cost disallowances for non-compliance.

My pipeline: AI handles research + compliance. Human writes every narrative section. Full compliance.

NSF

Disclosure required. No caps — yet.

Current PAPPG permits AI use but requires disclosure of "extent and manner" of use. No originality declaration. A new PAPPG (26-1) is deferred while OMB updates the Uniform Guidance — disclosure language could tighten.

My pipeline: full disclosure included in every NSF submission. AI stages documented in the methodology note.

DOE

IT governance approach.

AI-assisted content must meet federal accessibility and plain-language requirements. Accessing AI tools on DOE computers needs a business justification. Less prescriptive than NIH but compliance-oriented.

My pipeline: all outputs meet plain-language and accessibility standards by default.

Foundations / private

No formal policy — but reviewers notice.

Private foundations haven't published AI disclosure rules, but program officers read proposals carefully. Generic AI prose triggers the same "this wasn't written by someone who knows the work" reaction as a poorly researched human draft.

My pipeline: voice-matched narrative. Evidence from YOUR data. No generic fills.

AI handles the research. A human writes the narrative. That's the split that wins.