How to make a general contractor AI-native: a playbook

Picture a Tuesday on a federal job. The superintendent's daily log says the owner's consultant held the level 3 slab pour over a reinforcing question, and the crews moved elsewhere. The hold lasts a week. The log is accurate and signed. Nobody turns that line into a change event. A month later the PM hears about it in the OAC meeting and writes the notice.

By then part of the cost is gone for good. Under the federal changes clause, "no adjustment ... shall be made for any costs incurred more than 20 days before the Contractor gives written notice"[11]. Private contracts have their own clocks. Nobody did anything wrong in the field. The paper did not move.

The story above is a scenario, not a reported case. The money behind it is real. Tutor Perini, a US-listed building and civil contractor with $5.5 billion of 2025 revenue, ended the year carrying $324.7 million of claims and $402.1 million of unapproved change orders as contract assets, together about 13% of its revenue: money spent on work whose price nobody had agreed[4]. Its 10-K defines the second as work "performed before the parties have agreed on the corresponding change in the contract price". For a GC, notice, backup and pricing sit on the balance sheet.

An agent that reads every daily log on every job would have opened that change event on the Tuesday and put the deadline in front of the PM. That is worth more than a faster RFI, though the staged path still starts with documents, because a change event is only as good as the drawings and logs behind it.

Who this is for. US commercial GCs, CMs at risk and design-builders with roughly $50 million to $2 billion of annual volume, who buy most of their work from subs and run jobs in Procore, Autodesk Construction Cloud or Oracle Aconex. Specialty subs and homebuilders are not the subject.

The field stays under the superintendent and the competent person. Agents work in the office around them and read what the field records. A useful test: on any job, can the firm show which drawing revision a draft came from, who approved it and when, and that no contract clock ran out while a log entry sat unread?

A GC will not replace Procore, its accounting system or its schedule to become AI-native. It will add agents around them: the ones its project system ships, the ones from startups, and ones it builds. Each comes with its own permissions and its own idea of what a person must approve, and nobody holds one set of rules or one record across them.

OrchKernel is that layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened. Procore's own agents already mirror Procore permissions and leave the final call to people[44]. OrchKernel's job is the part no single vendor covers: one set of rules, one approval queue and one log across Procore, email, accounting and any other agent the firm uses.

What it is not. OrchKernel is not a project management, accounting, drawing-log or scheduling system; those stay the systems of record. It does not keep the drawing log current, verify a certificate with a broker or decide what is controlled information, and it has no role in physical safety. Details are in the OrchKernel blueprint.

Building GCs earn a thin spread on large revenue, and disputes swing the result. Operating margins of US-listed contractors for fiscal 2025[5,6,7,8,9]:

Self-performers and specialists earn more than the building GC model, whose spread a few bad jobs can take. Tutor Perini's 10-K ties its swings to specific projects, including a $25.9 million hit in 2024 on a Florida government building "primarily due to increased costs associated with external subcontractors and resolution of certain delay change orders"[4]. Private-GC benchmarks are member-only at CFMA, so we quote none.

Input prices for new nonresidential construction rose 8.9% in the year to August 2026[37], while ABC members held 8.5 months of backlog[38]. Work priced a year ago is being bought out now, so escalation clauses, buyout timing and change pricing carry more of the margin.

Of 8.364 million US construction jobs in September 2026, 5.28 million were at specialty trade contractors[1], the subs a GC buys from. 87% of firms in the AGC and NCCER survey have craft openings[36], and Deloitte puts the need at 499,000 new workers in 2026[41]. Agents in the office do not lay block. What they can take is the paper that keeps superintendents and project engineers at a desk while they coordinate subs who are short of workers.

The problem is old. Goolsbee and Syverson find US construction productivity "falling or, at best, stagnant over multiple decades"[40].

The most quoted RFI figures are from 2013 and the time survey is from 2018, sponsored by a software vendor. The baseline you record in Stage 0 will be better evidence for your firm than anything above.

At each step of a job, some of the work is drafting and some is a decision. The diagram splits the two; the numbers point to the control points in the blueprint.

Examples, not recommendations. Performance figures are the vendors' own.

Nearly every category works on documents, where a mistake can be caught before anyone sends it. The governance each vendor describes stops at its own product: three tools mean three sets of permissions and three logs.

Construction's AI use roughly doubled in ten months and is still low. In the Census Bureau's survey, 15.4% of construction businesses used AI in early September 2026, up from 7.4% in November 2025; 17.7% expect to within six months[2]. Manufacturing sits at 19.5% and professional services at 43.9%.

Two cautions. The survey counts firms, not jobs or dollars, so small contractors dominate it. And "any AI in any function" includes a chatbot opened once.

RICS's 2025 survey of 2,200+ construction professionals found 45% with no AI in their organization, 34% piloting and under 1% using it across the organization. Top barriers were skills (46%), integration with existing systems (37%) and data quality (30%), ahead of cost[39]. Its respondents are mostly in the UK, but integration and data quality are the problems of a GC whose documents live in five systems and change by revision.

The platform is large and still loses money. Procore had $1.32 billion of revenue in 2025 at a 79.5% gross margin and lost $124.3 million from operations[10]. It is the system of record most AI tools for US GCs sit inside or connect to.

The technology-native builder failed. Katerra took in more than $2 billion to be manufacturer, architect and contractor at once, peaked at 7,500 staff and went bankrupt in June 2021[54]. The funded AI companies in this section sell tools to GCs and subs; none builds. The realistic path is a GC that keeps its model and changes how its office works.

We found no evidence of an AI-native general contractor operating at scale.

The order follows three facts. Documents and revisions are where current AI works best, and errors are reversible until a person sends. Contract clocks and payments are where a GC's money is lost. And field safety authority cannot move to software. Start on one job, then spread across jobs. Different jobs can be at different stages.

Stage 0 plus one Stage 1 or Stage 2 flow, on one job. The baseline you record in the first month tells you by the end of the quarter whether it worked. We run this with design partners.

Our recommendations, inferred from the work rather than from a survey.

What must not change: the field's safety authority, the design team's authority over design, and the PM's authority over money and notice.

An agent can prepare each of these. A named person makes the call, however good the drafts get.

Across contracts and states, the same requirements keep coming back: give written notice on contract clocks; pay subs and protect lien rights on statutory deadlines; sign payroll statements truthfully on prevailing-wage work; report serious injuries fast and keep a competent person in charge of hazards; leave design decisions with licensed professionals; prove where materials came from on federally funded work; and protect controlled facility data. Each one ends in a person's decision or signature, which is why each one maps to a control below.

For scale on safety: 1,032 people in construction and extraction occupations died at work in the US in 2024, 370 of them in falls, slips and trips[3].

Several of the best-known collapses and fires of the past 45 years turned on document control and who had authority: a connection change approved by phone, cracks observed while work went on, product claims taken on trust. Each points to a control that agents can either support or undermine.

These are scenarios, not reported cases. Each is what an agent with too much access, or too much trust, could do on a job, and each maps to a numbered control point.

OrchKernel sits between AI agents and the systems a GC runs on, as drawn in the missing layer. Agents ask it before they act; it checks the rules, holds what needs a person, and records what happened.

Where a GC needs a control whatever tools it uses, who owns it, how OrchKernel enforces it, and what stays in another system.

OrchKernel governs what agents do in the systems a GC connects. It does not make a model provider FedRAMP-equivalent or decide what is controlled information, and whether a self-hosted OrchKernel with a cleared model fits a firm's CMMC boundary needs a CMMC practitioner's view. Some of these controls, such as approval on RFIs and permissions that mirror the project system, also exist inside Procore's own agents[44]; OrchKernel holds them in one place across the other systems and agents too.

OrchKernel is source-available under the Business Source License and can run on your own servers, so you can read the code that enforces these controls.

Record your baseline in Stage 0, then track the same numbers at each stage. Public benchmarks for GC operations barely exist, and we say so row by row rather than quote rules of thumb.

• How much of a GC's office time goes to RFIs, submittals, changes and billing today. No current independent time study exists; design partners' baselines will be the first good data.
• Private GC margins. CFMA's benchmarker is member-only, so we quote only public contractors.
• Whether AI drafting changes change-order recovery or dispute outcomes. Nobody has published it.
• The exact wording of AIA A201-2017 sections 1.5 and 15.1.2, which we have not read, and whether firm-wide AI search over past drawings fits the license.
• Lien and prompt-payment timelines beyond California and Texas, the Washington, Oregon and Nevada heat rules, and the 2024 change to Illinois' biometric law.
• How Autodesk, Oracle and Trimble's AI features compare. Their pages were blocked when we checked, so vendor claims here come mostly from Procore and startups.
• Whether a self-hosted control layer with a cleared model fits a given firm's CMMC assessment boundary.

Last reviewed October 2026. Sources were read in October 2026 unless the entry says otherwise; dates are publication or data dates. Vendor figures are directional, not an industry benchmark.