What actually mattered in our live contract-review build, and why the agent was the least important thing in the room.
Last month I told you a story about a principal at a well-run Arizona subcontractor who signed a seven-figure contract from the cab of his truck and relearned a $180,000 lesson that was sitting in the back pages. The headline of that piece was simple: nobody reads the whole contract, and the fine print is counting on it.
On June 11, in front of a room full of Arizona builders, we did something about it.
We took a real subcontractor’s actual GC contracts and, live and unscripted, watched AI read one cover to cover. It flagged the clauses worth a second look, summarized what was unusual, numbered every issue so a human could react to it, and then drafted the response back to the GC in the project manager’s own voice. Start to finish, the working tool came together in about forty minutes.
It was a great demo. And the demo was the least important thing that happened.
Because the point was never the agent. My co-host Rowan Steel-Hall of Smartbuild said it best about thirty seconds in: we are here to teach you how to fish, not to hand you a fish. The contract reader was the fish. The handful of transferable skills underneath it are the part you can point at any messy workflow in your business. Those are the takeaways. Here they are.
1. "Which AI should I use?" is the wrong question.
Rowan opened with the clearest framing of the AI landscape I have heard. The tools everyone argues about, ChatGPT, Claude, Gemini, Grok, are not really products. They are engines. Almost nobody in construction is building their own AI from scratch. The contract tool, the takeoff plugin, the scheduling assistant, the thing bolted into your project management platform: pull the cover off and you will find one of a handful of the same engines humming underneath. It is not that different from Formula 1, where only a few companies actually build engines and everyone else builds a car around one.
Which means the question most people lead with, which model is best, is the wrong question. They leapfrog each other constantly. A brand-new one literally dropped the night before our session. For the work most of you are doing, chasing the “best” model is like buying a superbike to ride to the corner store. They are all wildly more capable than the job in front of you. Pick the one that fits your business and your budget, and get good at it. That is the whole strategy.
2. Level one to level two is the entire game.
Rowan uses a simple ladder. Level one is what almost everyone does: open a fresh window, type a question, get an answer, close it, start over tomorrow. By his read, close to ninety percent of people live there. Level five is custom automation and code that a couple of the curious folks on your team will chase on their own time.
The leap that matters for the rest of us is level one to level two, and it has a boring name: persistent context. Instead of re-explaining yourself every single time, you build a reusable tool once, a Project in Claude, a Custom GPT in ChatGPT, a Gem in Gemini, and load it up with your context so it already knows the job before you say a word. That is exactly what the contract agent is. You do not need to chase level five. You need to stop starting from scratch.
3. Teach it by showing, not by telling.
This is the one I want you to underline. When Rowan built the agent, he did not sit down and write a rulebook for how Kristi at SHADE Industries reviews a contract. He did almost the opposite. He fed it a handful of her real contracts and the real emails she had sent back to GCs, and he asked the AI to study them and write its own guide to how she works.
It learned her voice. The bullet style, the “is this okay?” at the end of every flag, the way she points to a page and a section. Interviewing her for every rule would have taken days and still missed half of what makes her writing hers. Five good examples got there in minutes.
The lesson generalizes to anything. If you want AI to produce work that sounds like your company, do not start by describing your company. Show it five examples of the thing done right and ask it to reverse-engineer the pattern. Provide the content first, then refine with instructions.
Examples beat adjectives every time.
4. Treat it like a sharp intern, and make it show its work.
A couple of quiet process habits did a lot of heavy lifting in that build.
First, Rowan did not ask the AI to solve the problem. He asked it to write a plan to solve the problem, then built from the plan. His framing: treat it like a really smart intern, not a seasoned pro who has worked your trade for twenty years. Brilliant, fast, eager, and in need of a clear assignment and a review.
Second, he refused to let it sprint to the finish. The natural temptation is to have it read the contract and spit out the email in one shot. Instead the agent works in phases: analyze first, number every issue (critical items, red flags, and blue clarifications), then stop. The human reads the numbered list, keeps this one, kills that one, and only then asks for the email. The numbering is the small trick that makes it work, because “dismiss C2 and include R1” is precise in a way that “the second thing, no, the other second thing” never is. Kristi can even rip through the whole list in voice dictation.
The machine drafts. The human decides. That order is not a nicety. It is the whole safety model.
5. It does not get tired on a Friday afternoon.
Here is the part that ties back to my actual job. For most of construction history, the only defense against these mistakes was a careful human with enough time and enough coffee. Time is the one thing nobody signing a contract between job sites actually has. That is changing, and it is worth knowing what to watch for.
On the scope-and-number side, estimating and takeoff platforms have gotten genuinely good at catching the misses in bucket two. Tools like Bluebeam now use AI-assisted symbol detection to scan a drawing set and count every fixture in seconds, work that used to eat hours, while keeping the estimator as the validator rather than the clicker. Bid-leveling features line up subcontractor quotes side by side to expose the scope gaps that create the Winner’s Curse, and platforms like Procore keep the estimate, the contract, and the field tied to one source of truth.
On the contract side, an entire category has grown up to read agreements the way a tired human cannot. Platforms like Document Crunch are trained on construction contracts, not generic legal language, so they flag the pay-if-paid, indemnification, and lien provisions that general-purpose tools miss, and they translate the legalese into plain-English summaries in minutes. The category is consolidating fast: Trimble just announced it is acquiring Document Crunch to make contract-risk intelligence a built-in layer of its construction platform. When a software giant buys the contract-review startup, that is the market telling you this problem is worth solving with software.
I am not here to sell you any of these, and this is not a product bake-off. The right tool depends entirely on how your team works. The point is simpler than a shopping list: as this software gets better, and as AI gets layered on top of it, the gap between a contract that got a real review and one that got a windshield skim is going to get very wide, very fast. Be on the lookout for what these tools do well, and start asking how they would fit your process.
The part I am obligated to say honestly
Dimensions does not sell you the contract agent. We do not sell the AI. What we kept circling back to on the call is the stuff underneath it, because the agent is only ever as good as the foundation it stands on.
Two foundation lessons from the build are worth repeating. One: do not build the one tool to rule them all. The instinct is to stuff every brand document and every workflow into one giant project. Do that and quality quietly falls off a cliff, because you have buried its attention. Build scoped tools instead, a contract reviewer, an RFI assistant, a submittal helper, each one lean and specific. Two: your data is the lever. Name things consistently, organize your files, and use the paid tier so your contracts are not parked on a server training someone else’s model. Clean infrastructure first, then clean data, then the AI on top. Skip a step and the whole stack wobbles. That part, the network, the security, the data hygiene, the access controls, is the part we actually own.
What this looks like on Monday
You do not need a budget or a committee to start. You need to copy the move.
Pick the one document-heavy task your team does over and over and quietly dreads. Contract review, RFI responses, submittal logs, whatever yours is. Gather four or five examples of it done well, the input and the finished output both. Build one scoped tool on the paid version of whatever engine you already pay for, feed it those examples, and ask it to learn the pattern. Then run it beside your manual process, not instead of it, until it earns your trust. That is the entire playbook. The contract was just our example of it.
One More Thing
When we asked the room whether they wanted this to be a regular thing, ther majority raised a hand for a regular session. Message received.
So we are running it again in mid August, and we are picking another workflow that eats your team’s week to build live, end to end, the same way. I have a few candidates circling: the RFI black hole, the submittal grind, the estimate that goes out the door without a second set of eyes. I would genuinely like to know which one hurts the most at your shop. Reply and tell me. The workflow that makes the most of you wince is probably the one we should tackle.
Same format. Real construction problem. No slides full of theory, just the work. And yes, everyone who comes walks out with the tool we build, the way our June attendees walked out with the contract agent and a setup guide.
The agent was never the point. The skill to build the next one is. That part travels.
For over 20 years, Dimensions has been the trusted IT partner for Arizona’s architecture, engineering, and construction industry. We help AEC firms communicate better, collaborate smarter, and actually use the technology they’ve invested in. Because in construction, the tools only work if your team does.
IT Built For Builders.
