Neural · Construction
The margin is decided before the crew gets on the roof.
A commercial contractor already has the data to see a job going sideways. It's in the estimate, the contract, the daily reports, the cost ledger and somebody's inbox. It is never in one place at one time. An agent layer that reads across all of it can put the exception in front of the person who can still do something about it.
Read this first. Neural is the agent architecture we run inside our own company and are adapting for construction. It has not yet been deployed inside another contractor. You would be the first. Every capability below is labelled for exactly where it stands, and the labels separate what runs at CKS from what is in client beta from what is only designed.
Contents
- 01Where margin leavesThree places it goes, in the order they cost you
- 02One job, one morningThe whole loop, start to finish, on a real condition
- 03The workforceFive workflows and exactly where each one stands
- 04What we've actually doneStated precisely, including what we haven't
- 05DeploymentThe questions your IT group will ask first
- 06How it startsOne workflow, sixty days, read-only
[01]Where it goes
Three places margin leaves.
The gap between the bid and the job
The number that won the work and the number that closed it are different, and the reason is spread across a dozen documents nobody reads together. By the time the variance is visible in the monthly, the crew has been off the roof for weeks.
Change that never gets papered
Tear-off finds wet deck. The foreman calls it in, the work proceeds because the crew is standing there, and the change order gets written days later. Sometimes it doesn't get written at all. Every hour between the discovery and the paperwork is unbilled risk.
Risk that only surfaces when it's expensive
The exclusion nobody caught, the responsibility that landed nowhere, the allowance that won't hold. It's all in the scope documents at bid time. Nobody has the hours to read every one of them the way they'd need to be read.
[02]One job, one morning
What it looks like when it works.
- 06:40
A foreman's daily report notes saturated deck on the north bay, two sections beyond what the estimate assumed.
- 06:41
The agent layer reads the report against the contract, pulls the notice provision that governs it, assembles the T&M backup and the photos, and drafts the pricing off your unit rates.
- 07:15
It's the second line of the PM's brief, above the noise. She reads the draft, corrects the quantities, decides it goes.
- 07:20
The request is documented and transmitted while the crew is still on the roof, instead of being reconstructed from memory next week.
- ·
Every step of that is in the log: what the agent read, what it proposed, what she changed, when she approved it.
The agent never sends anything to an owner. A person does, every time.
[03]The workforce
Five workflows, and where each one actually stands.
Every one has the same shape: a trigger, agent work, a human checkpoint, an artifact. The status under each name is the honest part. One of these runs our own company today. One is in beta on our construction platform build. Three are specified against it and not yet built.
Scope risk read
Trigger
A scope document lands: invitation, spec section, owner exhibit.
Agent actions
Reads it against the estimate. Flags exclusions, unassigned responsibilities, allowances that look thin, terms that differ from your standard.
Human checkpoint
Estimator confirms or dismisses each flag before it reaches the bid.
Output
A risk memo attached to the opportunity, with the clause quoted.
Field-to-change-order
Trigger
A foreman reports a condition that differs from the contract documents.
Agent actions
Pulls the notice and authorization provisions that actually govern it, assembles the T&M backup and photos, and drafts the pricing off your unit rates and markups.
Human checkpoint
PM reviews, corrects the quantities, and decides whether it goes out. Nothing reaches an owner without a person.
Output
A documented change request with its backup attached, while the crew is still on the roof.
Margin watch
Trigger
Nightly, across every open job.
Agent actions
Compares committed cost, field-reported quantities and open change exposure against the estimate by phase. Deliberately not accounting alone, which lags.
Human checkpoint
Project executive triages the list each morning and tells it what was noise.
Output
A ranked exception list. Five jobs, not eighty.
Executive brief
Trigger
Every morning, before anyone is at a desk.
Agent actions
Reads across the systems that hold your operation and writes what changed, what slipped, and what needs a decision today.
Human checkpoint
You reply in the thread; the work routes from there.
Output
One brief, in the order you'd have asked for it.
Closeout assembly
Trigger
A job reaches substantial completion.
Agent actions
Gathers the warranty registration, inspection records, submittals and lien releases the package requires, and names what's missing.
Human checkpoint
PM signs off before the package leaves.
Output
A closeout package and a list of what's still outstanding.
[04]Standing
What we've actually done.
Stated precisely, because in this category precision is the differentiator.
We run our own company on it
Neural Ops plans the work, writes and reviews the code, runs the infrastructure and files the exceptions at CKS every day, including the site you're reading. We'll demo the live console rather than a slide.
We're building a construction platform now
We are actively building and beta-testing software with a commercial contractor. No client-specific scope is disclosed here, and none will be. That same discretion is what you would get. We'll talk about our own architecture in as much depth as you want.
We've consolidated a contractor's stack before
A custom management platform for a general contractor over $300M in revenue, replacing a set of disconnected third-party tools and removing more than $150K a year in licensing. Client-reported, from the original engagement.
We ship at enterprise scale
Mobile and kiosk ordering deployed across 100+ major stadiums and arenas. Connected hardware and inventory control across 500+ industrial sites. Both delivered before the agent platform existed.
The pilot below is scoped on that basis: first deployment, priced and measured accordingly.
[05]Deployment
The questions IT will ask.
Answered here rather than in the third meeting.
- Where it runs
- Your infrastructure or a private tenant. Not a shared multi-tenant product.
- What it reads
- The systems you already run: accounting, project management, email, files. Read access first, always.
- Who approves
- Actions you designate as consequential stop at a named person before they happen. You set that boundary; the platform enforces it rather than relying on discipline.
- What's recorded
- Each agent action is logged with its inputs, the sources it cited, what it did, and who approved it. The things that hold up in a review.
- Who owns it
- Your data and configuration are yours, and we'll put the exit terms in the contract rather than asking you to trust a web page.
[06]How it starts
One workflow. Sixty days.
Not a platform rollout. We pick the single workflow where your margin is leaking hardest, stand it up against your real systems, and you judge it on whether it caught things your people did not have time to.
Week 1
The workflow and the success metric are agreed in the first session, not discovered over a month. Read-only access to the systems it needs; your security review runs in parallel.
Weeks 2–7
It runs alongside your people on live jobs. Every output reviewed by your PM, nothing acted on automatically, false positives tuned weekly.
Week 8
Measured against the metric we set in week 1. Keep it, change it, or walk away. Read-only means there is no platform to unwind.
- What it costs
- Fixed fee, quoted after the working session, so it is priced against your actual systems rather than a guess. No per-seat licence during the pilot.
- How it's judged
- One number, agreed before we start. Usually exceptions caught that your team didn't have hours to catch, measured against a baseline period.
- What we need
- Read-only access to two or three systems, one PM who reviews the output, and an executive who reads the weekly.
- What you keep
- The configuration and everything it produced, whether or not you continue. Data handling and retention are in the agreement before access is granted.
A working session is ninety minutes with your operators, not a pitch. We map where your margin actually leaks and tell you which parts of it are worth pointing an agent at, including the parts that are not.