Why Generic AI Writes Dangerous Proposals

I’m going to tell you a true story a prospect shared with us: When their marketing coordinator pulled up her firm's newest AI-drafted proposal, one line described a $12M wastewater project as "successfully delivered" for a repeat client. Turns out, that project was never successfully delivered. In fact, it wasn’t even awarded. The firm bid on that job eighteen months ago and lost. Nobody caught it until a teammate who worked the original pursuit happened to skim the file before it went out.

That's the danger with generic AI in proposal writing, and it's a strange one to explain: the thing that makes it good is the same thing that makes it dangerous. It writes fluently. It writes confidently. It writes a better first draft than most people can manage on deadline. 

Unfortunately, none of that has anything to do with whether what it's writing is true, and the model has no way to know the difference. So a lie about your track record comes out sounding as convincing as the truth. Writing well and knowing what actually happened are two different skills, and a generic model only has one of them.

This blind spot shows up in two places, and neither one is harmless. 

The wastewater line is one version of it. Somewhere in that firm's proposal archive, the same project shows up twice: once in the original submission, hedged as "our proposed approach," and again a year later in a project sheet that got reused for the next pursuit, caveat stripped out, because nobody flagged that the deal never closed. Feed both documents to a generic model and it has no way to know which paragraph came before a loss and which came after a win. It just sees a wastewater plant described twice, with no instinct for which version to believe.

The second version shows up between firms instead of between projects. Proposal content gets reused constantly in this industry: a project sheet built for a 2022 pursuit where three firms teamed up gets pulled again years later for a solo bid, because it's still the best writeup of that project sitting in the drive. What doesn't survive the copy-paste is the context. A model reading that sheet later has no way to recover any of it. It just sees "we delivered this" and repeats it.

Neither mistake stays harmless once it's sitting inside a submitted proposal. Overstate delivered experience to a private client and you're issuing an awkward correction. Do it on federal work and the legal risk is real. The government doesn't have to prove you meant to mislead anyone, only that you didn't check. (Fox Rothschild, FCA 101: Falsity) On an SF-330, that same gap shows up in Section F, where past performance gets scored line by line, and what federal evaluators are actually scoring leaves no room for an incorrect byline.

The fix is not a better prompt 

Telling a model to "only describe delivered work" doesn't help. Especially if it was never taught what counts as delivered, or trained to notice a logo that isn't yours. The fix has to happen before a draft ever gets generated, on the documents themselves, sorting every one of them by what it actually is and who it actually belongs to.

But of course, that’s a mammoth task. One you can use AI for, but just not generic AI. 

Sorting these documents right takes someone who's done the job, not just someone who can read.

Picture handing the same stack of proposals, resumes, and project sheets to two different people to sort. Person A spent a decade as an AEC marketing coordinator. Person B is sharp and organized, but has never seen an SF-330 in her life. Both can read every word on the page. But only one of them knows, without being told, that a letter of interest isn't a signed contract, that a joint-venture partner's logo in the corner of a project sheet means that project doesn't belong in your capabilities section, or that a resume listing someone as "PE candidate" in 2019 needs a second look before anyone calls them a licensed engineer today.

Person B could learn all of that eventually, but only if someone wrote down every rule first and even then, they'd miss the ones nobody thought to write down. That's the difference between a generic AI pipeline and the AEC-specific one we built. 

Kantiv's sorting logic didn't start from a blank slate. It came from a team that had done the job of running pursuit calendars, assembling SF-330s, and living the same seller-doer chaos this piece is about, for years. The checks exist in Kantiv because someone on our team already knew, from experience, exactly what an experienced coordinator checks by instinct on every pursuit. Then we automated it.

Encoding that instinct is what took two and a half years

Reading a PDF and pulling out names, dates, and dollar figures is the easy 80% of a document pipeline. Plenty of generic AI tools can do that well. Whether a project is a bid you lost or work you delivered, or whether a resume belongs to your staff or a subconsultant's: those aren't generic document-processing problems, and a general-purpose extraction tool wouldn't know to check for them. 

Encoding that judgment into something repeatable meant building checks for exactly those blind spots. We run many checks on every document before that information is allowed anywhere near a draft.

Four (of the many) checks, that close the trust gap

• Type. Every document gets classified on the way in as an RFP, proposal, resume, and case study before extraction even starts.
• Status. A project only earns "your experience" status once it's shown as delivered work inside a submitted proposal or case study, not the moment it's mentioned somewhere.
• Ownership. Teaming partners get separated by logo and name detection on the source page, cross-checked against your actual CRM or ERP records (Dynamics, Deltek, Vantagepoint) instead of guessed from a scan.
• Source. Every claim in a draft links back to the exact document it came from, so verifying it is a click-through, not a research project.

So now, instead of rereading every paragraph, hunting for a lie dressed up as fact, your team clicks through one tag and confirms it in seconds. What used to mean proofreading the whole proposal now means glancing at a handful of receipts.

Whatever AI touches your next pursuit needs to know your track record

If you're using or exploring AI tools for your pursuits, a question worth asking is: how often can it actually tell a bid from a win, and how easy does it make for you to verify claims?

Because, generic or not, AI output is not 100% infallible. Nobody should use AI without a human in the loop. The difference is how much manual time, effort and energy the tool saves.

You can use our ROI tool to find exactly what that looks like when you use Kantiv. Or set up a chat with our team if you have any more questions, or want to run your firm’s data through our tool for a trust check.