Blog › AI Document Review

Why AI Must Never Sign: Professional Sign-Off in AI-Assisted Review

ParitySense team · September 2026 · 6 min read

Professional sign-off is the act by which a qualified person — a building surveyor, certifier, engineer, or a contract administrator issuing a formal position — puts their name, and everything attached to their name, behind a conclusion. AI can accelerate every step that leads to that moment. It must never perform the moment itself. This is not a compliance nicety we've bolted onto a product; it is a structural fact about how accountability works in Australian construction, and any tool built as if it were optional is a tool built for someone else's industry. (What follows is practical framing from people who build review software, not legal advice — your obligations come from your registration scheme, your contract and your insurer, and those are the authorities to consult.)

Accountability doesn't compress

Consider who signs things on an Australian project. The building surveyor or certifier whose approval sits within a statutory registration scheme, with a board that can discipline them personally. The engineer whose design certification carries their registration in the growing number of states that require one. The superintendent certifying payment or assessing a claim under the contract. The CA whose formal correspondence creates positions their employer will live with in adjudication under security of payment legislation.

Every one of these acts shares a property: it binds a person — their registration, their professional indemnity insurance, their standing. That is deliberate. The system works because someone identifiable can be asked, later and under pressure, "why did you conclude this?" — and must have an answer. Software cannot hold a registration, owe a professional duty, appear before a board, or carry PI cover. A tool that "signs" hasn't absorbed the accountability; it has merely obscured which human is actually carrying it — usually the one who clicked fastest.

What "review aid" honestly means

So the ceiling for AI in this domain is fixed, and the honest product question is what fits underneath it. Quite a lot does:

• Reading and cross-checking the entire set — the mechanical layer where machines genuinely outperform sampled human review (the capability case, and the failure modes, are in the honest guide).
• Surfacing candidate findings with evidence attached, so the professional adjudicates claims rather than hunts for them — provided every claim is citation-backed, which is the point of evidence gating.
• Drafting instruments — RFIs, comparison schedules, report sections — for a person to review, amend and issue.

And the line it must not cross: the tool never decides a finding is confirmed, never sends correspondence, never issues a report. In ParitySense that boundary is enforced in the architecture — the system has no code path that confirms a finding or dispatches an RFI without a named human acting, and the workflow is built around that adjudication step. We'd make the same recommendation about any competitor: if the boundary is a policy, it erodes under deadline pressure; if it's architecture, it holds.

The audit trail is what makes assistance defensible

Here is the part that decides whether AI assistance strengthens or weakens your position when a conclusion is later challenged. The question that arrives — from a board, an insurer, an adjudicator, opposing counsel — is always some form of "demonstrate how this decision was reached." A defensible answer separates machine work from human judgement, with records for both:

1. What ran: which checks, on which documents at which revisions, when.
2. What was found: each candidate finding with its citation to the evidence — and, just as importantly, what the system withheld as unverified.
3. What the human did: confirmed, dismissed with reason, corrected, escalated — per finding, with actor and timestamp.
4. What was issued: which report or correspondence went out, containing which confirmed findings, sent by whom.

Notice what this record does to the AI-assistance question: it dissolves it. "Did a machine help?" becomes uninteresting when you can show that every finding in the signed report was individually verified against a cited page by a named professional. Conversely, a process that can't produce this record is fragile whether or not AI was involved — the transcript of a chat session is not an audit trail, because a surveyor must be able to demonstrate the reasoning without replaying a conversation. This is the same logic your profession already applies to the RFI register: numbered, nothing deleted, every action attributable. AI-assisted review just extends the register discipline to the review itself.

The position worth holding

Sceptics in your office are right that a machine's opinion is worth nothing at sign-off. Enthusiasts are right that unassisted humans cannot exhaustively cross-check a modern document set. Both are answered by the same arrangement: the machine reads everything and proves what it claims; the professional judges everything that matters and signs; the record shows exactly where one ended and the other began. Tools should be evaluated on how well they implement that arrangement — and rejected, whatever their demo looks like, if they blur it.

FAQ