Proposal Materials

AI in Proposals

51 starting points for using AI on a bid. Adapt each one to your solicitation; none of them is boilerplate.

Prompts

  • The four-part prompt pattern: The default skeleton for any serious proposal prompt. The "No support found" line is what stops AI from inventing requirements, indicators, or facts. Keep it in.
  • POP (Persona, Objective, Parameters): A lighter, faster structure for everyday drafting and analysis when the full four-part pattern is more than you need. Same idea, fewer fields.
  • Shred an RFP into a compliance matrix: At solicitation analysis, to turn a long document into a structured checklist in minutes. Swap "Sections L and M" for the relevant sections of your solicitation. Always run a human compliance check over the result; the matrix is a first pass, not the final word.
  • The shall/must anti-hallucination guardrail: When you need a clean, defensible list of binding requirements and you cannot afford an invented or softened one. Forcing the exact quote and source page is what keeps the AI from paraphrasing requirements into existence.
  • Donor and program research: Early, to turn design notes into a funder-aligned summary, and to pressure-test an objective against a funder's stated priorities before you build the bid around it.
  • Retrieve past performance from your own library: When you need real, citable past performance fast. The key is that AI retrieves from your own real records rather than generating from memory, which is the difference between true past performance and fabricated past performance. The "do not invent" clause is non-negotiable here.
  • Audit a theory of change: Paste a theory of change you already own and use AI as a skeptical reader to find the weak joints before an evaluator does. Treat every flag as a question you answer, not a verdict.
  • Audit a results framework or logframe: To test that your levels line up end to end. Trust the mechanical findings (a missing output, an indicator at the wrong level); reconcile the level vocabulary yourself against the funder's definitions.
  • Audit an indicator set: To tighten indicator wording and feasibility. Keep the precision improvements; verify every suggested indicator and data source against a real library or the funder's handbook, because this is where AI invents the most.
  • Draft a results framework from a finished theory of change: Only after the theory of change is finished and human-owned. AI is a capable drafter when it is translating logic you already stand behind, not inventing it.
  • The anti-fabrication line (add to any drafting prompt): Append it to any prompt that asks AI to draft content with facts in it. It is the single most useful guardrail in this whole library, and it costs you one line.
  • Simulate the evaluator's score: As a pink-team or self-review pass before a real review, to find where a section underperforms against the published criteria. Adapt "Section M" to your solicitation's evaluation section. Treat the score as a prompt for human judgment, not as the truth; it is no substitute for a subject-matter expert and customer insight.

Texts

  • General AI-use disclosure (proposal or report): The default disclosure line when a funder asks you to declare AI use but does not specify a format. The "fully responsible" sentence mirrors the language funders themselves use and is the part reviewers care about most.
  • Disclosure with a carve-out for routine assistance: When you want to declare substantive AI use while making clear you are not listing every spell-check. This matches how most funder disclosure rules actually draw the line, on substance rather than grammar.
  • Footnote / marginal-notation style (for funders that require it inline): When your funder requires AI-generated text to be acknowledged inline, by footnote or marginal note, rather than in a single statement. Attach it to the specific text that AI helped draft, not to the whole document. Check the exact format your funder asks for.
  • Minimal "drafting only" disclosure: When AI's role was genuinely light, drafting and tidying prose rather than shaping the approach, and you want a short, honest line that does not overstate the tool's contribution.
  • Disclosure plus data-handling assurance (one block): When you want the disclosure and the data-protection assurance in a single statement, for a funder or partner sensitive to how a not-yet-public bid was handled. The middle sentence is what reassures a reviewer that your solicitation and program data stayed protected.
  • Reviewer-confidentiality note (when you are the one reviewing): When your organization reviews others' documents, for example as a sub-granting or pass-through partner, and needs to state that confidential material is protected. This mirrors the reviewer-side confidentiality rules that funders apply to their own panels.
  • Responsible AI use in the MEL approach (core paragraph): The default paragraph for describing AI in a MEL methodology or plan. Keep the human-validation, grounding, and triangulation sentences; those are the credibility load-bearers.
  • AI to scale qualitative and synthesis work: When the program's MEL load is high-volume qualitative or document-heavy and you want to show AI extends capacity without ceding judgment. The closing inclusion-check sentence answers the bias concern before a reviewer raises it.
  • Human-in-the-loop commitment (short): A compact, drop-in line when you need to signal responsible use in a sentence or two, for example in an executive summary or a capacity statement.
  • Sensitive-data handling with AI tools: A methodology or data-management section that handles personal or sensitive data. Pick the bracketed option that matches your real practice. The consent-and-reuse sentence is the one funders increasingly look for.
  • No client or community data used to train external systems: When a funder or partner is sensitive to data sovereignty or third-party training, which is increasingly common in federal and research contexts. State only the protection you can actually deliver.
  • The pairing rule, in one line: The template behind every framing line. Fill the first bracket with a specific task and the second with the check that keeps it honest. Never claim the benefit without the guardrail attached.
  • Qualitative coding (the highest-value, highest-risk claim): A MEL narrative or win theme built on qualitative analysis at scale. This is the most-cited AI-in-M&E use and also where it most often goes wrong, so the codebook-and-reliability guardrail is not optional.
  • Data cleaning and harmonization: When you want to claim an efficiency saving on data preparation. Frame the saving against the specific task, not against "M&E costs" in general.
  • Evidence synthesis across many documents: A synthesis-heavy or evidence-review program, especially in [country/region] settings where important evidence is oral or in local languages. The inclusion check is what keeps the claim defensible.
  • Reporting from structured evidence: When reporting volume is high and you want to claim a drafting efficiency without ceding the judgment a reviewer cares about. The human-owns-framing half is the credibility anchor.
  • Cost framed as a shift, not a cut: A budget narrative or cost story. Framing the saving as redeployed effort rather than headcount you can cut is what reviewers believe; naming the hidden costs up front makes the rest of the claim survive scrutiny.
  • Where AI saves, and where it does not (the maturity signal): When you want to show a considered, bounded approach rather than blanket adoption. The "where we do not" half is what signals maturity to a reviewer.

Examples

  • A MEL section, before and after: Replace generic MEL adjectives with named measures, disaggregation, and a baseline; state the method specifically enough that a reviewer could picture the fieldwork; name your AI use plainly and attach the human check to it every time; say where sensitive data does and does not go; include a short honest disclosure matched to your funder. The point is the pattern, not the paragraph.
  • A compliance matrix, sample layout: It forces every requirement to carry its own source reference, keeps "what we will say" out of the extraction step (a human fills that column), and turns a 200-page solicitation into a checkable list. The risk is over-trusting the matrix: AI misses eligibility nuance and can paraphrase loosely, so a person verifies the extraction against the source before anyone relies on it.
  • RFP shred to a compliance matrix: Turning a long solicitation into a structured first-pass matrix in minutes instead of days. Pair it with a human compliance check; the matrix is a starting point, not a sign-off.
  • Anti-hallucination guardrail on the same task: Tightening the shred so the tool quotes exact source text rather than paraphrasing the funder back to you. "Quote the exact sentence" is the move that catches loose extraction.
  • MEL-audit prompts, AI as critic not author: The most defensible MEL use today, AI as a critic of a draft you already wrote, not as the author of the MEL section. You paste your own theory of change, logframe, or indicators; the tool flags weak links and gaps; you decide what to change. It does not invent indicators or frameworks for you, which is where AI most often goes wrong on MEL.
  • Anti-fabrication instruction: A one-line guardrail to append to any drafting prompt. Fabricated statistics, citations, and past performance are a documented failure mode; this line tells the tool to leave blanks rather than invent. It does not replace your own verification, but it reduces the surface for invented content.

FAQ

  • Will a funder reject my proposal just because I used AI?
  • What counts as "too much" AI?
  • Do I have to disclose AI use, and where?
  • Can I paste a funder's RFP, or a partner's or competitor's confidential information, into ChatGPT?
  • Will an AI evaluate my bid before a human sees it?
  • Can reviewers tell it's AI? How?
  • Who owns the AI-produced text? Can a partner reuse it?
  • If detectors are unreliable, why do reviewers still flag AI?
  • Will using AI make my proposal worse?
  • What's a safe division of labor between me and AI?
  • What should we put in an internal AI policy if our funder is silent?
  • Does using AI disadvantage local or non-native-English organizations?

Checklists

  • Pre-submission AI QA: Run this on anything AI touched before it goes into the final package. The point is to catch the failures AI is known for: confident wrong claims, generic prose, and fabricated specifics.
  • Data-security tier check: Run this before you paste anything into an AI tool. The single biggest blind spot in proposal AI use is sensitive bid information pasted into a tool that may train on it.
  • Disclosure documentation: Keep this record even if your funder has no published AI rule. It is the sane default when a funder is silent, and it is what you will need if a disclosure requirement turns up late.