Don't Be the 95%: What to Ask an AI Proposal Vendor Before You Sign

Every AI proposal tool looks incredible in a 30-minute demo. The draft appears fast. The formatting is clean. Someone on your team says "wait, it already knows who worked on the airport project?" and you can feel the room warming up to it.

But the thing to remember is that the demo is not the product. The product is what happens in month three, when pursuit specialists and engineers-turned-seller-doers are all trying to use the same tool, your SharePoint libraries haven't synced the way you were promised, and the AI just called a bid you lost a "case study." In fact, MIT's NANDA initiative found that 95% of generative AI pilots inside companies deliver no measurable return.

So how do you make sure that your firm falls in the other 5%? By asking the right questions.

We sat in on a call recently with a global engineering firm evaluating AI tools, and the questions their proposal and BD leaders asked were sharper than anything in a typical RFP. Below is the checklist worth stealing, built from that conversation.

Does It Know a Promise From a Project?

This is the one that should worry you most, because it's invisible until it isn't.

Somewhere in your content library sits a proposal for work you didn't win. It described, in confident future tense, everything your firm would do. If an AI tool trained on your past submissions can't tell the difference between "we will deliver this" and "we delivered this," it will eventually write that lost bid into a new proposal as a completed case study. Your client catches it. Your credibility takes the hit, not the AI's.

This isn’t something a smarter prompt will fix–It needs better architecture. A well-built system tags a document's status the moment it's ingested and before a single word gets reused, whether it is a proposal, project profile, resume or case study. If a firm's project only counts as "experience" once it's actually shown as delivered work, on paper, that's the tell you're dealing with a tool that did the unglamorous plumbing work up front instead of hoping the language model would sort it out later.

Ask the vendor directly: how does your system distinguish a submitted proposal from a completed project? If the answer is some version of "the AI is pretty good at figuring that out," that's not an answer. 

Can It Tell Your People From the Other Firm's People?

Joint ventures and teaming partnerships are normal in this industry. They're also where AI content generation falls apart, because a resume or project profile with a partner firm's logo on it can get scooped up and presented as your own experience if nobody built a way to catch it.

The detection here should be almost boring in how simple it sounds: if a partner's logo or name appears on a page, the system flags that content as external. And if you've got a system of record, Microsoft Dynamics, Sharepoint, Deltek, or whatever you're running, that data should always outrank whatever gets extracted from an old PDF. Ask where the tiebreaker sits. If there isn't one, you're the tiebreaker, manually, every time. 

Is It Built For Your Sellers & Doers?

Most AEC firms don't have one proposal team. They have two, even if nobody's called it that.

There's a smaller, dedicated group who lives inside pursuits full-time who are neck deep in deadlines, storyboards, compliance checks, all day. Then there's a much larger group of engineers and architects who write proposals occasionally, usually for the smaller pursuits, usually while also doing their actual job. One group needs depth and control. The other needs something closer to a fast, low-friction assist that doesn't ask them to learn a new system for a $100,000 submission.

A tool built only for the first group will frustrate the second into ignoring it. A tool built only for the second group won't hold up under the complexity the first group deals with daily. Ask the vendor how their product changes for a casual, occasional user versus a full-time pursuit lead. If the answer is "it's the same experience for everyone," that's a company that hasn't sat with your actual org chart yet.

Does It Meet You Where Your Data Already Lives?

Every AEC firm we talk to has already sunk real time into organizing resumes, project profiles, and past proposals, usually in SharePoint and sometimes layered with a CRM like Dynamics. That's the work. That's the asset.

A vendor asking you to re-upload all of it into their platform, by hand, is asking you to redo work you already finished. It's also the single way to ensure a pilot dies. You’ll see a slow fade as the "temporary" manual upload step never gets automated, and six months later nobody's touched the tool.

What you want to hear is that the platform syncs with your existing libraries as a source, applies its own structure on top, and leaves your original files exactly where they are, so your other tools pointed at that same SharePoint hub still work fine. Structured data you already paid to build should make the AI's job easier, not get thrown out and rebuilt from scratch.

What's the Real Onboarding Timeline?

If a vendor tells you two weeks to full adoption, run.

That promise usually means "two weeks until the software is technically installed," which is a very different thing from "two weeks until 30 full-time proposal staff and 700 occasional users actually trust and use it correctly." Enterprises that have been burned by this before, and a lot of AEC firms have, know the difference by now. The outside data backs them up: research on enterprise AI rollouts puts the real average timeline at 6 to 12 months. 63% of AI projects run past their original schedule anyway, usually because nobody budgeted time for the unglamorous parts, like getting the data and the people ready.

A realistic rollout runs closer to two to three months and looks less like a launch and more like a project plan: a small group making early configuration decisions (what data sources connect, what tags get created, what the style guide should enforce), then phased training across cohorts so the system is actually right before more people are using it. Slower at the start. Faster forever after.

Will It Catch Your Mistakes Before the Client Does?

Public sector reviewers, in particular, do not forgive a proposal that misses a page limit or skips a mandatory clause. A compliance check that runs your draft against the actual RFP requirements before submission, therefore, becomes the last line of defense between your team and an embarrassing scoring cut.

Ask whether the tool checks your response against the RFP directly, flags missing or mismatched clauses, and lets you run it against your own internal checklists too whether that’s legal, IT, sustainability, whatever your firm tracks. If compliance checking is an afterthought bolted onto the product, it'll behave like one.

The Actual Question You're Answering

The real question was never "can this AI write a proposal." Every tool on the market can write a proposal. The question that separates a good demo from a good decision is whether the system will actually work the way your firm works, in sync with your data, your two very different types of users, your teaming partners, your compliance requirements. And whether it will get the facts right when nobody's double-checking it.

That's a longer conversation than 30 minutes. It's also the one worth having before you sign anything. If you want to run your own firm's setup through this list, or you've got a question none of the above covers book a demo, or send in your questions!