AI in the Loop: The Right Way to Use AI on Proposals

Why one-shot AI fails on proposals

A proposal is not a writing task. It is a compliance task, an evidence task and a persuasion task that happens to produce writing. One-shot generation — prompt in, draft out — skips all three. The output is fluent, generic, unsourced and unscored. It reads fine and loses anyway, because nothing checked it against the buyer’s requirements or your actual evidence.

• No requirement coverage — nothing verifies every “shall” is addressed
• No grounding — claims come from the model’s general knowledge, not your past performance
• No evaluation — nobody scored the draft the way the buyer will
• No accountability — you can’t trace a sentence back to a source

AI in the loop, defined

AI in the loop means AI operates inside your workflow, not instead of it. The software does the heavy reading, drafting and checking; you make the decisions that decide the bid — what to pursue, what to claim, what wins. Every output is inspectable, every claim is cited, and nothing ships without your approval.

The cycle: Understand → Draft → Evaluate → Refine

The improvement cycle that separates working AI from demo AI has four stages, run repeatedly until the draft scores:

• Understand — shred the solicitation into every requirement, evaluation factor and submission rule before a word is written
• Draft — generate each section grounded in your knowledge graph of past performance, people and evidence, with citations
• Evaluate — score the draft against the buyer’s own criteria, the way their evaluators will
• Refine — fix the weakest-scoring sections first, then re-evaluate; repeat until the score holds

One-shot AI stops after Draft. The Evaluate and Refine stages are where compliance gaps, unsupported claims and weak win themes actually get caught — before the buyer catches them for you.

You stay in command

AI in the loop is not autopilot. The division of labor is explicit: software handles volume and verification; you handle judgment.

• You decide go/no-go, win themes, pricing posture and discriminators
• You approve every claim before it ships — citations make review fast
• You direct the Refine passes: which sections, which criteria, how aggressive
• The system flags what it can’t support instead of inventing it

What it looks like in practice

On a real pursuit: the system shreds a 200-page RFP into a compliance matrix in minutes; drafts sections from your cited evidence; a built-in red team scores the draft against Section M or the buyer’s weighted rubric; you spend your week on strategy and the ten weakest answers instead of on formatting and hunting for boilerplate.