AI in proposals
Almost everyone on a proposal team is already using AI. Almost no one is using it systematically, safely, or in a way their organization has actually thought through.
This is an honest map of an open, fast-moving space, for the whole bid team: capture and BD, proposal managers, technical writers, MEL specialists, and reviewers. What early adopters are doing and what works so far, not settled best practice, because there is none yet. The MEL section gets the deepest treatment, because that is where the hard part lives.
- of aid workers surveyed have tried AI
- ~93%of aid workers surveyed have tried AI
- use it daily or weekly
- ~70%use it daily or weekly
- of their organizations report wide, integrated use
- ~8%of their organizations report wide, integrated use
- have a formal AI policy
- ~22%have a formal AI policy
Source: Humanitarian Leadership Academy and Data Friendly Space survey of 2,539 aid workers in 144 countries, 2025. Only about 64% report any training, and about 3.5% describe themselves as expert-level. In the proposal world, about 89% of roughly 750 proposal and sales professionals had experimented with AI (APMP-backed research by Responsive), which is breadth of trial, not depth of practice.
The shadow-AI paradox, named up front
The typical team has individuals quietly using ChatGPT or Claude on their own, with no policy, no training, and no sanctioned tool. The point is not to scold anyone for it. It is that the practical questions, which tier is safe, what to disclose, what AI quietly gets wrong, mostly go unanswered in that informal mode, and those unanswered questions are where bids get hurt.
Source: synthesis across Shipley, Lohfeld, GovEagle, APMP, and M&E-community practice. The manual baseline for analyzing a 200-page solicitation is commonly cited at around 5 to 7 days. The throughline: AI is good at handling text you already have and bad at inventing substance you do not.
AI is on both sides of the table
Funders and agencies are starting to use AI to screen the proposals they receive. If an automated compliance screen can eliminate you before a human reads your approach, then structure, compliance, and machine-readability matter more, not less. The early, mechanical uses (shredding the solicitation, building the matrix) harden exactly the layer a machine reviewer checks first.
Trust splits by stage
Trusted to AI: shredding the RFP into a compliance matrix, outlining and storyboarding, first drafts, retrieving past performance, mechanical review against evaluator scores.
Kept human: win themes, discriminators, red-team and gold-team judgment, final accountability, and the context-specific causal logic that holds a theory of change together.
Five habits separate the people who get value from AI from the ones who get burned.
- 1Ground prompts in source text.Ask AI to quote the exact requirement or the exact sentence it relied on, not to summarize from memory.
- 2Retrieve, do not invent.Point AI at your own win library, past performance, and bios. Retrieval over your real material is the antidote to hallucinated experience.
- 3Say "do not fabricate," explicitly.Tell the model in plain language not to invent statistics, citations, credentials, or past performance, and to write "no support found" when it has none.
- 4Review against the published criteria.Score drafts the way the evaluator will, using the actual Section M or equivalent, not against vibes.
- 5A human owns strategy and sign-off.Win themes, discriminators, causal logic, and accountability stay with people. AI assists; it does not decide.
Where AI fits in your bid
Most teams either over-use AI (let it write the narrative wholesale and ship generic prose) or under-use it (never let it touch the mechanical work where it is genuinely strong). For each stage: what AI actually helps with, the pattern early adopters trust, and the one caution that matters most. Drafting the MEL section is the anchor, because it is where AI helps most and can cost you the most.
1Capture and pre-position
Helps with: Synthesizing your own past notes, intel, and prior designs about a client or geography into a readable capture picture; drafting a project summary from your design notes for a given funder.
Trusted pattern: Feed it your own material and ask for a structured summary or a fit assessment, grounded in what you actually know.
The one caution: Capture strategy, the read on what the client really wants and how you will be different, is human work. AI can organize intel; it cannot tell you your discriminators.
2Solicitation analysis
Helps with: Summarizing a long solicitation, extracting requirements, and producing a first-pass compliance matrix in minutes rather than days.
Trusted pattern: Shred-to-matrix, grounded in source. Ask it to capture only "shall" and "must" statements, quote the exact sentence, and cite the section and page.
The one caution: The matrix is a draft, not the truth. A human still has to do the compliance read, especially for eligibility nuance the model glosses over.
3Go / no-go
Helps with: A fast, structured fit assessment against funder priorities; surfacing gaps in your capability or past performance early.
Trusted pattern: Ask it to assess fit against the stated priorities and list where you are weak, using your real history as input.
The one caution: Go/no-go is a judgment call with money and reputation attached. Use the analysis as an input to the decision, not as the decision.
4Planning (outline and storyboard)
Helps with: Turning the compliance matrix into a compliant outline, drafting a storyboard, and keeping the structure aligned to the evaluation criteria.
Trusted pattern: Generate the outline from the requirements you extracted, so structure traces back to the solicitation.
The one caution: Check that the outline maps to how the proposal will actually be scored, not just to the table of contents. Structure that mirrors Section M is what survives a machine screen.
5Drafting (technical)
Helps with: Fast first drafts, reformatting, tone consistency, and translation, which is a real equalizer for non-native-English and local organizations.
Trusted pattern: Treat AI as an overeager graduate assistant. Have it draft from your own inputs and approved boilerplate, then a human rewrites for substance and voice.
The one caution: "AI smell." Generic prose, RFP-mirroring, and vague, non-measurable outcomes are exactly what reviewers are trained to spot. A first draft is a starting point, never the submission.
6Drafting (MEL)
Helps with: Stress-testing a theory of change, logframe, or indicator set you have already written. The defensible use here is AI-as-critic, not AI-as-author.
Trusted pattern: Paste your existing ToC, results framework, or indicators and ask AI to flag weak causal links, hidden assumptions, unmeasurable indicators, and misalignment with the framework. You keep authorship; it pressure-tests.
The one caution: AI is least reliable on MEL. Ask it to write the framework from scratch and it will fabricate indicators, data sources, and citations, flatten the theory of change into a template, and cite mismatched donor frameworks with confident formatting.
The MEL section in depth7Past performance and personnel
Helps with: Retrieving relevant past-performance examples and bio material from your own library and drafting in your real voice with real experience.
Trusted pattern: Retrieval over your own past projects. Ask it to find examples matching a scope and value threshold and to extract the summary, the point of contact, and the rating from your records.
The one caution: This is the highest-fabrication-risk area when the model is not grounded. Ungrounded, it will invent plausible past performance and citations. Ground it in your real files and verify every specific.
8Review (color teams)
Helps with: A mechanical pre-review pass: scoring a section against the published criteria and flagging where points might be lost before a human color team meets.
Trusted pattern: Evaluator simulation. Ask it to score against the actual Section M (or equivalent) criteria and explain the likely point losses.
The one caution: The AI score is a prompt for human attention, not a verdict. It does not replace a subject-matter expert or customer insight, and a high AI score is not a win.
9Production
Helps with: Consistency checks, formatting cleanup, acronym and cross-reference checks, and catching mechanical compliance slips (page limits, font rules, required attachments).
Trusted pattern: Narrow, checklist-style passes over the near-final document.
The one caution: Do not feed a near-final, sensitive document into a consumer tool to "tidy it up." Production is exactly when the full bid exists in one file, so the data-security tier matters most.
10Disclosure and submission
Helps with: Drafting an AI-use disclosure statement, and building an internal record of which sections used AI, which model and version, and which prompts.
Trusted pattern: When the funder is silent (most are), document it anyway: sections, model/version, prompts. That record protects you.
The one caution: Disclosure rules vary widely and are changing. Check the specific funder; do not assume "everyone allows it" or "everyone bans it."
11Post-award and debrief
Helps with: Summarizing debrief notes, comparing your bid against feedback, and extracting reusable, win-library-ready content from a submitted proposal.
Trusted pattern: Use it to mine your own submitted material and debrief into clean, taggable reusable content for next time.
The one caution: Lessons about why you won or lost are judgment, not summary. Let AI organize the notes; you draw the conclusions.
By role
For MEL, treat AI as a critic, not an author.
You are the methodologist. AI is a fast, tireless reviewer who has never been to your project country and will state a wrong indicator definition with total confidence. Paste in your own theory of change, results framework, or indicator set and ask it to find what is weak, unmeasurable, or misaligned. Keep authorship and every final call with the human who owns the logic. This is also the one part of the landscape the proposal-AI tools mostly skip.
Critique your theory of change
AI reads the logic literally, with none of the optimism you built over months of design, so it finds the weak joints. Treat every flag as a question, not a verdict.
Open the audit promptCritique your results framework
It holds the whole structure at once: an outcome with no output, an indicator at the wrong level, a verification source that does not verify. Reconcile the level vocabulary yourself.
Open the audit promptCritique your indicators
Strong at pressure-testing whether an indicator is measurable; also where it most readily invents things that do not exist. Verify every reference.
Open the audit promptOne generation task does work, in strict order: once your theory of change is finished and human-owned, AI can draft the first-pass results framework from it. The prompt for that.
What AI reliably gets wrong on MEL
- Fabricated or generic indicators, data sources, and citations. One study found roughly two-thirds of AI-generated citations were fabricated or wrong. Verify every indicator, source, and reference against a real library or the funder's handbook.
- Bland, template-flat theories of change. Left to author, AI writes a causal story with no edges. Reviewers discount it because it could describe any program anywhere.
- Weak causal reasoning. AI can format a results chain perfectly while the logic underneath does not hold.
- Shallow or outdated donor-framework alignment. It may cite the wrong framework, a mismatched version, or a donor's level definitions applied incorrectly. Confirm all three yourself.
- Review burden that erases the time savings. Generate a MEL section wholesale and you have moved the time, not saved it. Critiquing a draft you own is fast; cleaning up one AI invented is slow.
Materials library
Prompts, copyable texts, worked examples, FAQ answers, and checklists, all tagged by proposal stage and role. Everything is a starting point to adapt, not a button that writes the proposal for you.
Filter by stage and role in the full library. The deepest set is MEL: the theory of change, results framework, and indicator audit prompts.
Open the materials libraryUse it with your eyes open
None of this is a reason to avoid AI on a bid. It is the specific ways AI quietly hurts proposals, so you can use it deliberately. Most are versions of the same root cause: AI is strong on text you already have and weak at inventing substance you do not.
Data security: know your tier
The most common blind spot and the easiest to fix. What matters is whether the tool may use your inputs to train its models.
| Consumer / free tier | Enterprise / Teams / API tier | |
|---|---|---|
| May use your inputs to train | Often yes (check the current terms) | Generally no |
| Admin controls and data handling | Limited | Added controls, retention settings |
| Safe for | Public information, generic drafting, your own non-sensitive notes | Solicitations, partner data, competitive intel, anything confidential |
The rule: a solicitation, a partner's or competitor's confidential information, or your own unannounced competitive intel does not go into a free or personal consumer tool. If your organization has not sanctioned an enterprise or API tier, treat anything sensitive as off-limits to AI until it has. The tier checklist
What stays human
The parts AI assists but never owns. Also, not coincidentally, the parts that actually win.
- Win themes. The handful of ideas the whole bid is organized around. AI can brainstorm against your history; you decide.
- Discriminators. Why you, and not the other bidder. Judgment about your real strengths and the client.
- Strategy. The capture read, the go/no-go call, the positioning. AI is an input, never the decision.
- Context-specific causal logic. The theory of change and the indicators you can actually measure in this place, with these partners. AI critiques it; it cannot author it credibly.
- Final sign-off. Accountability for everything that ships, including the duty to verify every fact AI touched, stays with a person.
AI smell
Generic mission statements, a tone that does not match your voice, vague outcomes, RFP language paraphrased back, and prose that is flawless but factually impossible. The fix is not to hide the AI better; it is real substance and real voice from a human.
Fabrication
Ungrounded, AI invents past performance, citations, DOIs, statistics, and credentials in flawless prose. Ground everything in your own files, tell it not to fabricate in plain words, and verify every specific against a real source.
Equity and detection
AI lowers the English-language barrier for local organizations, and detectors over-flag exactly those writers. Do not lean on detectors, and do not assume a clean result protects you. The protection is substance and voice that are genuinely yours.