Why Oversee AI
Every other construction platform bolted a chatbot onto old software. Oversee built the AI into the data model.
That's the actual difference, and it shows up in four concrete ways: every answer is grounded in your real project data with a visible source list, every write happens only after a human taps confirm, every prediction is judged against your own project's history instead of a generic industry rule, and every record is checked against your whole organization the moment it's created — not just the project you happen to be looking at.
01 · Grounded Answers
Not a black-box chatbot. A visible list of what it actually looked up.
Every answer in Oversee's AI Assistant comes from a real function call against your project's live Firestore data — never a guess dressed up in confident language. And you don't have to take that on faith: the Assistant's "Grounded In" panel lists, after every reply, exactly which tools it called to answer your question. If an answer didn't need to look anything up, the panel says so — it's general guidance, not something drawn from your project's data.
That distinction — traceable versus general — is shown on every single answer, not buried in a settings page.
Grounded In
Representative example — not a live screenshot.
02 · Agentic Actions
It can actually do the work. Nothing touches your project without a real tap.
Ask Oversee to create an RFI, log a safety observation, or update a task, and it will — but never in one step. Every write-capable action follows the same propose → confirm → execute path, and every step is checked against your real project role, the same permission system the rest of the app already enforces.
Propose
The model calls a tool — create_rfi, update_task_status, and the rest. Nothing is written yet. Oversee checks the request against your real project role, using the same permission matrix the rest of the app already gates on, then stages the action and writes a plain-English preview.
Confirm
You see the exact preview in the chat, with a real Confirm / Cancel control. No write happens without an explicit tap — not an auto-approve setting, not a timeout, not a default.
Execute
Confirming re-checks your role — in case anything changed between propose and confirm — then performs the real write. It lands in the same activity feed and fires the same outbound webhooks as a change a person made by hand.
A confirmed action lands in the same activity feed and fires the same outbound webhooks as a change your team made by hand — an AI-created RFI is never a second-class record.
03 · Predictive — and It Shows Its Work
Told why, not just what — and judged against your project, not an industry average.
Per-trade pace risk, before the due date passes
Oversee compares how far a task is through its own scheduled window against how much of it is actually marked done — and flags it while there's still time to act, not after the due date has already come and gone. The bar it's judged against isn't one flat rule for every trade: it's this project's own history. A trade that's historically run 22% over its own schedule here gets weighed against that number; a trade that's always been on time gets held to a tighter one.
Change-order justifications, grounded — or left blank
Ask Oversee to draft a change order's justification and it writes the cost narrative from that record's own title, description, and linked field data. A schedule-impact narrative only gets added when there's real delay evidence tied to it through the shared cost code — no evidence, no section. It leaves the paragraph out rather than inventing a delay that never happened.
04 · Cross-Project Memory
It remembers what happened on a job six months ago, on the other side of your company.
Every RFI, safety observation, change event, submittal, and change order created in Oversee is checked against everything already on record for your organization — every project, not just the one you're in — the moment it's created. A real match doesn't wait for someone to think to go search for it: it surfaces as a proactive alert.
This isn't keyword matching. In testing, two RFIs describing the same underlying HVAC condensate issue — one written as "condensate line backing up into ceiling," the other as "water stain appearing below mechanical unit" — scored a 0.783 similarity match under Oversee's real embedding model. Worded nothing alike. Flagged anyway.
Possible Match Found
RFI — "Water stain appearing below mechanical unit"
Harborview Tower · this project
RFI — "Condensate line backing up into ceiling"
Riverside Commons · a different project, six months earlier
Representative example — not a live screenshot.
See it against your own project data
Ask us for a demo and try grounded answers, agentic actions, and cross-project memory on real projects — yours, if you want to bring them.