Per-seat desks that draft the admin work (vendor onboarding, COI tracking, work order coordination, bid assembly) and a signal engine that reads permits, public filings, and bid boards for funded work.
General contractors, specialty contractors, and facilities maintenance contractors from about 20 to 500 people. In production today at a Chicago-area commercial and facilities contractor.
Takeoffs, scheduling optimization, and jobsite cameras. Good vendors exist for those. See what AI is actually best for construction.
A two-week operations assessment that ranks the seats by hours lost, then one build measured against a number agreed up front.
Where AI pays first in a contractor's office
Every seat in a construction office drags an admin shadow: the hours spent collecting, re-keying, chasing, and formatting before the skilled part of the job can happen. The shadow is where AI belongs first, because the work is repetitive, it lives in email and PDFs, and the person in the seat already knows what a correct result looks like. The build order is arithmetic: hours in the shadow, times what that hour costs.
| Seat | The admin shadow | What a desk does |
|---|---|---|
| Vendor coordinator | Chasing W-9s, COIs, licenses, and prequal forms across dozens of email threads; tracking expirations by hand | Assembles the onboarding packet from the emails and PDFs, flags gaps, drafts the request for whatever is missing, watches every certificate's expiry |
| Work order coordinator | Reading each request, finding who is available, writing the dispatch note, updating the customer | Matches the order to available crews, drafts the dispatch note and the customer update, queues both for a yes |
| Estimator | Pulling specs, past bids, and sub quotes into a bid package; re-typing the same scope language | Assembles the package from your own history, drafts the scope sections, leaves the numbers to the estimator |
| Controller | Month-end reconciliation, lien waivers, pay app assembly | Collects, matches, and drafts; a person signs |
| Business development | Watching bid boards, permits, and news for work worth chasing | An engine, not a desk. See below. |
A desk is where AI agents report to work. Most of a desk is ordinary software: connections into the systems you already use, one play per task, a memory that compounds with every job. Agents handle only the parts that need judgment, like deciding whether a certificate actually covers the scope. That split is what makes a desk reliable enough to run every day.
The approval gate is the whole point
Ask an AI assistant whether a general contractor should build or buy AI and it will tell you to buy the platform, hire a specialist for the integration, and be careful of anyone promising autonomous agents. It will ask for documented approvals, an audit log, and a human override. That is the right answer, and it is how every Keystone desk is built.
- Output is always a draft. A desk cannot send, submit, or sign. The gate is enforced by the architecture, not promised by a policy.
- Every draft cites its sources. The onboarding packet says which four emails and two PDFs it was assembled from. The person approving can check in seconds.
- Every decision is logged. Who approved what, when, and what changed. That is your audit trail for the insurance carrier and the owner.
- The seat stays. The coordinator keeps the chair and loses the re-keying. There is no replacement story here.
More on this in human in the loop AI agents.
Finding work before it is public
Most contractors hear about work when it reaches a bid board, which is the same moment every competitor hears about it. The funded work is visible earlier: in permit filings, municipal agendas and capital budgets, state procurement postings, federal inspector-general reports, and storm tracks that will produce repair demand in a known set of buildings.
A signal engine reads those sources continuously across your footprint, scores each item against criteria you set (building type, size, owner type, distance), and delivers a ranked list on a cadence. It never acts on the world. It writes to your record, and a person decides what to chase. For the Chicago-area contractor running one today, the footprint is Northern Illinois municipal and state records plus a federal facilities lane and a storm lane.
The engine is Keystone-operated: no one on your team is asked to supervise it. It is your engine, running on your criteria. Keystone runs it.
What to buy instead of build
Custom work is the wrong answer for jobs where a category vendor has already spent years on the problem. The assessment says so, in writing.
- Takeoffs and estimating math: Togal.AI, Bluebeam, or your existing estimating platform. General models like ChatGPT cannot reliably do quantity takeoffs from drawings, and you should not let them try.
- Project management and documents: Procore, Autodesk Construction Cloud. A desk connects to these read-only; it does not replace them.
- Schedule optimization: ALICE and similar. Specialist math, buy it.
- Jobsite capture: OpenSpace, CompanyCam. Cameras and 360 capture are a product category, not a custom build.
The custom layer is the office work between those tools, the part no vendor covers because it is specific to how your company runs. Full comparison in best AI for construction.
How it starts
- Assess (two weeks). Where the hours go, seat by seat. What your current software already does. How work reaches you today, by source. It ends in a ranked build list you can act on with or without Keystone.
- Build (weeks, not quarters). The top item on the list, shipped as production software on your real data, measured against a number agreed before the build starts.
- Expand. Monthly counts: hours returned, drafts approved, leads qualified. When a build pays for itself, the next one goes in beside it. When it does not, you can walk. Each client gets a separate instance and a written offboarding process.