AI review

Before you can price a bid, someone has to shuffle the PDFs: find your scope pages, pull the relevant specs, read the addenda, and spot the traps. AI review does that first pass for you. It does not do your job. It gets the pile sorted so you can start yours.

Running a review

Open a bid after uploading its documents and select Run document review. The app records the request, adds it to the background work queue, and returns control to you immediately. You can keep working in the bid or leave the page while review runs.

The bid shows the durable review status:

Queued means the request is waiting for a worker.
Running means document triage is in progress.
Succeeded means the findings are ready to review.
Failed or Canceled includes a safe explanation and lets you queue new work when appropriate.

The status refreshes automatically while the bid page is open. The attempt number is visible too. Unexpected worker failures can retry automatically, up to three total attempts, without exposing document contents in the error message.

Duplicate submissions do not create duplicate active work. If review is already queued or running, another click reuses that request. If the full review inputs already finished successfully, the app reuses the completed result. Uploading another document, changing the review profile, or deploying changed review logic creates a new input. An unfinished attempt may repeat work after a worker failure, but durable findings stay deduplicated. If documents change after a request was queued but before it starts, that stale request is canceled and the app asks you to queue a fresh review.

The trust rule

This is the rule the whole feature is built on, so it goes first:

You should not have to trust an AI summary. You should be able to check it. Every document-derived finding carries at least one citation: the document, the page number, and the quoted text that triggered it. The app rejects findings that arrive without evidence. Click the citation and the document viewer opens on that page, showing you the finding that sent you there.

What review finds

Run review on a bid’s documents and it produces findings grouped by type:

Scope pages: candidate pages that touch your trade.
Spec sections: specification sections relevant to your scope.
Addendum changes: what an addendum changed, compared against what you had.
Bid form requirements: submission requirements pulled from the bid forms.
Dates and deadlines: bid dates, RFI cutoffs, and similar.
Alternates and allowances.
Exclusions and qualifications.
Red-flag terms: language worth a second read before you sign up for it.
Missing documents: categories that look absent from the package.

Each finding carries a severity (info, needs attention, critical) and an origin: some findings come from deterministic checks, some from AI. The origin is visible, so you always know which kind you are looking at.

Suggestions, not verdicts

Findings are reviewable suggestions, not final truth. Two things make that concrete:

Confidence is visible. A finding is either confirmed or suggested. Suggested findings are the ones that need a human eye before anyone relies on them.
You decide. Mark each finding reviewed, dismissed, or important. Your decision is the durable record, and it follows the finding into exports. Dismissed findings are kept but set aside, so downstream estimating does not price rejected items.

What AI review will not do

It will not write quantities, prices, or proposal numbers. Ever.
It will not measure anything off the drawings. See the overview for what the product is not.
It will not present an unsourced claim about your documents as fact.

Treat review as a sharp-eyed first pass from someone who read everything but signs nothing. The estimator still owns the bid.