Your Team Is Already Pasting the Proposal into AI. Here's What That Risks.

A capture manager told us recently she had a feeling some of her colleagues were feeding their proposal into an AI chatbot and having it write sections. She's almost certainly right — surveys keep finding that a large share of knowledge workers use AI tools their employer never approved, and proposal teams under deadline are exactly the population that reaches for one. The problem isn't that AI touched the proposal. The problem is which AI, on what terms, doing which job. Most teams have never asked any of those three questions.

Risk one: where the text goes

When someone pastes a technical volume into a free consumer chatbot, the words leave your company's control. Depending on the tool and the account tier, that content may be retained on the provider's servers, reviewed by humans for quality, or used to train future models. The paste takes three seconds. The obligations it can violate took your contracts team years to negotiate:

• Competition sensitivity. Your win themes, pricing strategy, and staffing approach are the most valuable secrets your company has during a pursuit. A tool that retains or trains on them is a disclosure — even if no human competitor ever sees them.
• CUI and safeguarding clauses. Federal work often involves controlled unclassified information, and contractors accept specific safeguarding obligations for it. A consumer chatbot is not a compliant place for that material, full stop.
• The solicitation's own terms. Many RFPs restrict disclosure of solicitation materials. Pasting the government's draft SOW into a public tool can breach the terms of the very competition you're trying to win.

The fix is not "never use AI." The fix is using tools with enterprise data terms: zero data retention, no training on your content, tenant isolation, and provable deletion. Before anyone on your team pastes anything anywhere, they should be able to answer four questions: Is this content retained? Is it trained on? Who else could ever retrieve it? Can we prove it was destroyed? If nobody knows, the answer is the free tier of a consumer app — and that's the wrong answer.

Risk two: what the AI writes

Suppose the data terms were fine. There's a second, louder risk: the words themselves. A proposal is not a writing sample — it's a set of promises your company becomes accountable for. AI-drafted proposal text fails in ways that are specifically expensive:

• It invents things you now own. Language models are confident fabricators. An AI-drafted past-performance section can cite work you never did; an AI-drafted technical approach can commit to certifications you don't hold or timelines you can't hit. If it survives review, it's not a typo — it's a commitment in your offer.
• It sounds like everyone else. Evaluators read stacks of proposals, and increasingly they read stacks of proposals written by the same handful of models. Generic AI prose — polished, symmetrical, empty — is becoming recognizable, and some agencies have begun asking offerors to disclose AI-generated content. Sameness is a scoring problem: evaluators reward specificity.
• It multiplies inconsistencies. When three authors each use AI on their own volume, you get three fluent documents that disagree — on headcounts, on dates, on tool names. The merge happens at 2am; the discrepancies survive; the evaluation board finds them.
• Your team stops knowing its own proposal. This is the sleeper risk, and it detonates at orals. If nobody wrote the approach, nobody can defend it out loud. The panel asks a follow-up about "your phased migration strategy," and the presenter is reading it for the first time on the screen behind them.

The job AI is actually built for in a pursuit

Here's the reframe that resolves the whole debate: AI is a mediocre author and a superhuman checker. The failure mode above — fabrication — comes from asking a model to generate commitments. But verification inverts the problem. Checking text against source documents is the thing models do reliably, because every output can carry its evidence: the quote, the page, the timestamp.

That's the role worth adopting, and it covers the whole pursuit:

• Before submission — score the written volumes against the RFP's own evaluation factors, and cross-check every number, headcount, and date between volumes. Humans wrote it; AI verifies it agrees with itself and answers what's scored.
• Before orals or interviews — score rehearsals against the evaluation criteria, catch spoken claims that contradict the written submission, and drill the panel's hardest questions. Humans present; AI holds the 200 pages in memory.
• Never — writing the approach, generating the past performance, or making the judgment calls. The humans who will stand behind the promises should be the ones who make them.

This is the line we build on at Take the Room: the tool never writes or rewrites a word of your proposal. It checks what your team wrote and said against what the solicitation demands — under zero-data-retention AI processing, per-organization encryption, and hard deletion with a receipt. AI as the red team, not the ghostwriter.

A one-paragraph policy your team can adopt tomorrow

"We write our own proposals. AI tools may be used to check, score, and rehearse our work — never to author content that becomes part of an offer. Any tool that touches pursuit material must guarantee zero data retention and no training on our content, and the capture manager approves every tool before first use. Where a solicitation asks about AI use, we disclose accurately."

Four sentences. It won't slow anyone down, it keeps the genuine productivity gains, and it means the next time someone has "a feeling" about where the proposal has been pasted, the answer is written down.