"Human in the loop" appears in nearly every construction AI vendor's materials, including ours. The phrase has become table stakes precisely because buyers are rightly wary of software that writes to their systems of record unsupervised. But the phrase describes at least three fundamentally different designs, and the differences determine what happens on the day the AI is wrong. Before trusting any tool with your project data, it's worth knowing which design you're actually buying.
Three designs that share one name
The optional review. In the most common design, the AI drafts something—a record, a summary, a classification—and presents it inside a portal for a person to look over. Nothing technically stops review, and nothing technically requires it. Under deadline pressure, "review the AI's 40 suggestions" becomes "scroll and click accept," because the interface makes wholesale acceptance the path of least resistance. The loop exists in the demo and evaporates in month three.
The after-the-fact audit. In the second design, the AI acts autonomously and a person can inspect what it did afterward—an activity log, a dashboard, a weekly digest. This is better than nothing, and for low-stakes work it can be an appropriate trade. But auditing is detection, not prevention: by the time a person sees the error, the wrong sub has been notified or the wrong status recorded, and someone is now cleaning up in front of a counterparty.
The hard approval gate. In the third design, the AI's output cannot reach the system of record until a named person approves it. The AI does the reading, extraction, and drafting; the person reviews, corrects, and signs off; only then does anything commit. The gate is structural—not a checkbox the vendor added, but the shape of the workflow itself.
Why the difference matters in construction
In some industries an AI error is an inconvenience. In construction, committed data triggers consequences with contractual weight. An observation record notifies a subcontractor and starts a clock. A COI marked compliant lets a sub mobilize with a coverage gap your risk team believes is closed. A lien log entry with a mistyped deadline misinforms decisions about statutory response dates. These aren't display errors; they propagate outward into schedules, payments, and legal exposure, and unwinding them costs far more than preventing them.
That asymmetry is why we regard the hard gate as the only appropriate design for this industry's back office, and why every Bex module is built around one—the full argument is in how Bex keeps humans in the loop.
What a real approval gate looks like in practice
A gate that people bypass isn't a gate, so the practical question is how to make review genuinely light. When Bex Punch processes a consultant's observation report, it emails back an approval request with every item extracted and classified—project, assignee, trade, location, photos. The reviewer reads it in their inbox like any other email. If something's off, they reply in plain English: "reassign item 4 to Vasquez, set all types to Deficiency." Bex applies the corrections and sends a fresh request. Only when the reviewer replies with an approval does anything reach Procore or ProjectSight.
Correction-by-conversation matters more than it might appear. A gate that only offers accept-or-reject forces a reviewer who spots one wrong field to either swallow the error or reject the whole batch and start over; either way, the gate gets resented and then circumvented. A gate where fixing an error costs one sentence gets used.
Equally important is what happens when the AI itself is unsure. Bex escalates ambiguous cases—an incomprehensible reply from a sub, an insurance endorsement with unusual wording—to your team with its analysis attached, rather than guessing confidently. An AI that never says "I'm not sure" should worry you more than one that sometimes does.
Doesn't the gate slow everything down?
The reasonable worry about mandatory approval is throughput: if a person must sign off on everything, haven't we just moved the bottleneck? In practice, no—because the gate reviews work that is already done. Reading a 30-item consultant report and keying it into Procore takes a project engineer a couple of hours; reviewing Bex's extraction of the same report takes minutes, because the reading, classifying, and photo-matching are finished before the reviewer arrives. The 33-item report on our proof page went from a two-week backlog to subs notified within the hour, approval gate included. The gate costs minutes and buys certainty about what entered the record; the hours were never in the approving.
Where the approval lives
One design detail with outsized consequences: the medium of the approval. When sign-off happens through a click inside a vendor's portal, the evidence of who approved what lives in that vendor's database, on that vendor's retention schedule. Bex's approvals happen over email, which means every request, correction, and sign-off sits permanently in your own Exchange or Google Workspace archive—searchable by your people, governed by your retention policy, and available years later when someone asks why item 12 went to the drywall sub. The audit trail is a side effect of the workflow rather than a feature you configure, part of the posture described in secure, observable, trustworthy.
Questions that expose the design
When a vendor says "human in the loop," these questions will surface which of the three designs they mean:
- What exactly happens if no human acts on the AI's output—does the data commit anyway, later, or never?
- Can the reviewer correct the output, or only accept and reject it?
- Where is the approval recorded, and who controls that record?
- What does the AI do when it isn't confident—guess, or escalate?
- Show me the reviewer's actual experience on a 40-item batch, not a 3-item demo.
Vendors with a real gate answer these quickly, because the answers are the product. If you'd like to see Bex's answers demonstrated on one of your own documents—an observation report, a stack of COIs—send it over and watch what the approval loop looks like from the reviewer's chair. The contact link is below.