Here's an uncomfortable truth about AI proposal tools, ours included: a reviewer that is confidently wrong is worse than no reviewer at all. If a compliance check tells you a mandatory requirement is missing from your submission when it's actually sitting on page 34, your team burns an evening rewriting content that already exists. And if it tells you something passes when it doesn't, you find out from the selection committee.
When we dug into the cases where our own Submission Review missed the mark, two patterns kept showing up. The first is language equivalence: the RFP asks for “10 years of experience,” the firm wrote “over a decade,” and a literal-minded check calls that a miss. The second is silence about uncertainty. Most AI tools return a verdict — pass or fail — with no signal for how sure the model actually is. Every answer looks equally confident, so you can't tell which ones deserve a second look.
We think the bar for AI in proposal work should match the bar for a junior staff member doing the same job: show your work, flag what you're unsure about, and never present a guess as a fact. That standard shaped everything in this release.
Confidence scores on every verdict. Each requirement in Submission Review now carries the AI's self-rated confidence, displayed as a chip next to the verdict. High means explicit, direct evidence. Low means look at this one yourself. And there's a hard rule underneath: if the AI wants to mark a requirement as failed but its confidence is low, it must downgrade to “partial” and explain what would resolve the ambiguity — instead of quietly guessing against you.
Click a verdict to see the source. Every requirement now has a View Source button that opens your own submission PDF in a side panel, jumps to the relevant page, and highlights the passage the AI is citing. Agree or disagree with the verdict in one click — no more downloading your own proposal and hunting with Ctrl-F.
Equivalence rules. The reviewer now explicitly bridges paraphrase: a decade equals ten years, OBC 2024 equals the Ontario Building Code 2024 edition, and a capability stated in any submission document counts regardless of which document the requirement came from.
Honest failure states. If part of a review can't be completed — a processing error, a timeout — those requirements are now labelled “not evaluated” with a one-click re-evaluate button, instead of being disguised as fails. And if you change your submission documents after running a review, a banner tells you the results may be stale.
The AEC industry is right to be skeptical of AI that can't be audited. Proposals carry real money and real reputations, and “the tool said so” is not a defence a principal can take to a selection committee. Verifiability isn't a nice-to-have feature — it's the price of admission for AI in this industry. This release is us paying it.
If you want to see what a verifiable AI review looks like on one of your own submissions, book a 15-minute demo.