AI spec book
Project specifications made answerable — ask what a section requires and get the cited clause back.
AI & automation
The busywork writes itself.
AOS doesn't bolt a chatbot onto the side. AI is wired through the record where the work already lives — it reads the invoice and codes it, turns a foreman's voice note into a structured daily log, levels the bids, prices the ROM, drafts the schedule update and grades it. Every call runs on a metered budget you can cap per feature, so you always know what AI is doing and what it costs.
Invoice & document parsing
Voice → structured forms
AI bid leveling & ROM pricing
Metered per-feature budgets
OCR
Invoices, COIs, and documents parsed and routed to the right record automatically
Voice
Field voice notes transcribed and extracted into structured forms with confidence scoring
Judge
AI-written summaries graded by a second model against the source data before they're trusted
Meter
Every AI call runs on a wallet with per-feature budgets, auto-refill, and full usage visibility
Document AI
Email it a PDF. It files itself.
An inbound document hits a classifier that figures out what it is — vendor invoice, certificate of insurance, lien waiver, submittal — and routes it to the right workflow. Invoices are parsed for vendor, PO, line items, and totals; COIs are read for carrier, limits, and expiration; certified-payroll docs are extracted into the report. The unstructured pile that used to need a human becomes structured data.
- Inbound classifier — incoming documents identified by type and routed to the matching workflow.
- Invoice parsing — vendor, PO, line items, totals, tax, and remit-to pulled from a PDF or image.
- COI extraction — carrier, coverage limits, and expiration read off a certificate of insurance.
- Certified-payroll extract — wage data lifted from submitted payroll docs into the WH-347 flow.
- Textract-backed OCR — document text and tables extracted with AWS Textract before the model reads them.
Inbound — classify & route
| Document | Classified | Routed to |
|---|---|---|
| atlas-inv-422.pdf | Invoice | AP 3-way match |
| steelco-coi.pdf | COI | Sub compliance |
| waiver-388.pdf | Lien waiver | Pay-app release |
| scan-unknown.pdf | Low conf. | Human review |
High confidence auto-routes
Low confidence → review
Voice capture
Talk through the day. Get a structured log.
A superintendent shouldn't type. They talk — crews on site, work put in place, delays, weather — and AOS transcribes it, extracts the structured fields a daily log needs, and scores its own confidence so low-certainty values get a second look instead of being silently wrong. The same voice-to-form pipeline fills inspections and other field forms.
- Audio intake — capture a voice note in the field, geotagged, and queued for processing.
- Transcription — speech to text via Deepgram or Whisper, built to handle a noisy jobsite.
- Structured extraction — the transcript mapped into daily-log and form fields, not just a wall of text.
- Confidence scoring — each extracted value scored; uncertain ones flagged for quick confirmation.
- Voice assistant — a session-based assistant that drafts and runs tools against the record on request.
Voice → daily log · field capture
1Foreman records 90s notegeotagged
2TranscribeWhisper
3Extract crews + workstructured
4Confidence check2 to confirm
5Daily log drafted1-tap submit
Estimating & pricing AI
Level the bids. Price the unknown.
Bid leveling AI reads sub proposals, normalizes them to a common scope, and surfaces the gaps and exclusions that make a low number not actually low. ROM pricing recommends an order-of-magnitude number from comparable historical work, and the markup recommender suggests margin from the factors that drove it on jobs like this one — each a recommendation you accept, not a black box that decides.
- AI bid leveling — proposals extracted, normalized, and reduced to an apples-to-apples comparison.
- Gap & exclusion detection — scope a sub didn't include surfaced before it becomes a change order.
- ROM pricing — order-of-magnitude pricing recommended from comparable historical projects.
- Markup recommender — margin suggested from the factors that shaped it on similar work, with comparables shown.
- Estimate analytics — search and expand historical estimate data to inform the next number.
Bid leveling AI — glazing package
| Sub | Base | Gap found |
|---|---|---|
| Steel City | $1.21M | Complete |
| Allegheny | $1.18M | No sealant |
| Three Rivers | $1.25M | Complete |
| Normalized low | $1.21M | Steel City |
Drafting & coding AI
The first draft, already written.
The weekly schedule update, the GL coding on an incoming transaction, the 1099 classification at year-end — the judgment calls that eat an afternoon get a first pass from AI, with a human approving rather than starting from blank. And where it matters, a second model grades the first one's work against the source numbers before anyone relies on it.
- AI schedule narrative — the weekly update drafted from the live rollups; detailed on the CPM page.
- AI-as-judge — a second model grades the narrative against the underlying schedule data before it's sent.
- Transaction auto-coder — GL account, job, and cost code suggested for incoming transactions.
- 1099 classification — vendors classified for 1099 reporting with a recommended box at year-end.
- Toolbox-talk generation — safety talks drafted from upcoming scope; see Safety & Quality.
GL auto-coder — incoming transactions
| Transaction | Suggested code | Conf. |
|---|---|---|
| Ready-mix delivery | 03 30 00 · Westside | 0.97 |
| Crane rental | 01 54 00 · Harbor | 0.91 |
| Fuel card | 01 50 00 · shared | 0.72 |
| Misc supply | Review | 0.55 |
Governed AI
AI with a meter and a ceiling.
Every AI call in AOS draws on a wallet. You set a budget per feature, watch usage in real time, and let it auto-refill or hard-stop at the cap — so estimating's experiments never blow up the whole month, and you always know what AI cost and what it did. The model provider sits behind one governed interface, not scattered keys across the app.
Metered wallet
Every call debits a wallet; usage is visible by feature, in real time, with a running balance.
Per-feature budgets
Cap each feature independently — estimating, documents, voice — so one can't drain the rest.
Auto-refill or stop
Top up automatically at a threshold, or hard-stop at the ceiling. You choose the behavior.
One provider interface
Model calls run through a single governed client with pricing and budget enforced in one place.
More of the AI layer
The work nobody else automated.
Quote parsing
Insurance renewal quotes parsed and compared so the renewal decision isn't a PDF pile.
Owner CRM AI
AI assists on the owner CRM — drafting, enrichment, and follow-up on the development pipeline.
Voice forms engine
A reusable voice-to-form engine with typed schemas powering inspections and field forms.
Estimate search
Search and expand historical estimate data to ground the next bid in what actually happened.
Smart intake
Inbound documents and emails triaged and routed, with low-confidence items sent to a human.
See the AI layer on your own documents.
30 minutes. Bring a stack of invoices and a voice note from the field. We'll parse them, route them, and show you the wallet that keeps it all on a budget.