An independent magazine about AI at work in construction. Every story starts on a real job site, with the people who used the technology and the results they will answer for.
Colin Stoner Said No to Procore in 2011. Novo's AI Now Runs on What He Built | ConstructionMagazine.ai
Colin Stoner Said No to Procore in 2011. Novo's AI Now Runs on What He Built
The innovation chief of a 400-person California GC spent August explaining Sentinel, the system Novo built instead of buying, and the AI on it: insurance certificates 99 percent automated, RFIs drafted from 14 years of Novo's writing.
In 2011 Novo Construction looked at Procore and decided not to buy it. Colin Stoner, who had joined the Menlo Park general contractor seven years earlier, led the company into building its own system of record instead, a platform Novo calls Sentinel. He has described the decision, on Bricks & Bytes in August, as half-witted at the time. Fifteen years on it is the reason a 400-person contractor has fourteen years of its own RFIs, insurance certificates, and schedules in a shape a model can be trained on, and Stoner, now Novo's chief innovation officer, spent August explaining what he does with them.
This profile is built from three places Stoner spoke in the last five weeks: a Construction Dive Q&A published 5 August on how Novo compares drawing sets with AI, a Bricks & Bytes episode released 30 July with clips posted from 24 to 26 August, and the episode's published summary. He has not spoken to this magazine. Where a number is his, it is marked as his.
Twenty-two years, one email address
Stoner joined Novo in 2004 and has used the same email address since, a detail the podcast uses to explain why vendors can always find him. Novo is a 400-person general contractor headquartered in Menlo Park with work that now reaches Austin. Its chief innovation officer is also, on Construction Dive's byline, its chief information officer, and at a firm that writes its own software the two jobs are the same desk. The 2011 decision was his, and the podcast's chapter list frames it the way he does: ConTech platforms come and go, horizontal technology stays, and a contractor that builds on the second is not exposed to the first.
The summary's version of his cost case is that building Sentinel was cheaper than licensing, by a margin he describes in two words and does not itemize. The podcast also lists a chapter on getting site teams to actually use what the office builds, which is the failure mode of most in-house tools. Neither the figure nor the method is public. What is public is that the platform is still in daily use fifteen years later, which is longer than most of the products Novo declined to buy in 2011 have existed.
Novo's own system of record, fifteen years of data beneath it, and the three AI uses that sit on top: insurance certificates, RFIs, and look-ahead schedules. AI-assisted illustration.
The Bricks & Bytes summary lists three AI uses in production at Novo. Certificates of insurance, which once arrived by the envelope every quarter and buried an administrator, are now processed by Box AI with, in the summary's words, 99 percent automated and no human review. RFIs are drafted from a forwarded email chain by a model trained on 14 years of Novo's own RFI writing. Three-week look-ahead schedules are generated on request from a tool called Conduit. Stoner's cost comparison against licensing equivalent software off the shelf, as the summary quotes him, is two words: massive savings.
The RFI case is the one he explains most directly in the clips. Nobody goes to college to write RFIs, the podcast's post says, and the point of training on Novo's own history is that a project engineer never starts from a blank page. The draft carries the house style because the house wrote the training set. That is the argument for owning the data that Stoner makes throughout: the structured record Novo kept for fifteen years became an advantage the day models could read it, and a contractor whose record sits inside a vendor's platform does not have the same option.
Drawing diffs on two California jobs
The one use that is not Novo's own code is the one Construction Dive covered. Novo uses BuildCheck, an AI drawing-review product, and in August Stoner said its newer Diffs feature had helped on two California projects. Diffs compares a drawing set to the previous version and flags what moved. Before it, a project manager overlaid sheets in Bluebeam and looked, which over hundreds of pages, in Stoner's words, makes you go cross-eyed. Now the PM runs a diff check, reviews the flags, dismisses the ones that do not matter, and when a wall has moved goes to the framing and drywall subcontractor to price the change.
Stoner is precise about where the value lands. It is not in preconstruction and not in the field. It is the window before construction starts, when a moved wall or an added window changes the scope Novo is contracted for and the price can still be changed. He did not give hours saved or name the two projects, and Construction Dive did not press for them. What he gave instead was the rule he applies to every tool, his own or a vendor's: trust but verify.
He was asked what the team learned about AI from the drawing tool, and the answer was about trust rather than speed. Before Diffs, the learning curve was steep because a person had to find every change by eye. Now the tool flags changes the same way it flags issues on an initial drawing check, the PM keeps a log of what was flagged and what was dismissed, and clicking a flag highlights the change on every sheet it touches. Stoner says that has made the team more trusting of AI in general, and in the same breath says he verifies everything that goes to them, whether he built it or a vendor did.
How a drawing revision gets priced at Novo (Stoner's description, Construction Dive, 5 August)
1A revised drawing set arrives
→
2The project manager runs a diff check against the previous set
→
3The tool flags every change on the affected sheets
→
4The PM reviews each flag and dismisses the ones that do not affect scope
→
5For a real change, the PM takes it to the affected subcontractor for a price
→
6The change is priced before construction starts
The rules he gives other contractors
Stoner's advice on building is a warning first. The Bricks & Bytes summary describes a test he uses to tell chief executives whether to attempt their own system at all, and the clips give the short form: build only if the whole team is bought in, and without that do not start. He expects more mid-size contractors to try it anyway, because the cost has dropped. His estimate in a clip posted 26 August is a team of three to five people and a runway of twelve to eighteen months.
On accountability he is blunt. AI will never be the fall guy, as the podcast's 24 August post summarizes him. Someone owns every workflow, and when a tool built in-house breaks, the failure belongs to the person who owns it, not to the model. On readiness his answer is a sequence rather than a purchase: understand the workflow first, know why each piece of data matters, then build. His Monday-morning advice to a chief executive is two questions long, and the summary says most firms cannot answer either. The questions themselves are behind the podcast's paywall.
He also draws one line. AI-powered robocalls to subcontractors, where a vendor spoofs a phone number and lets a model collect bid information from a sub who thinks they are talking to a person, cross it. He says that workflow should not be automated, and that relationships get more valuable, not less, as the rest of the process does.
Headcount, pitches, and the number he will not give
Stoner says AI will not shrink Novo's headcount. His team, in the podcast's paraphrase, is too busy to lose anyone, and his target is the same number of people carrying more revenue and more projects. Revenue per employee is the measure he expects to move. He has not published the current figure or the target.
The other number he gave is from the buyer's chair. He receives close to 100 cold pitches a week, at an email address he has used for 22 years, and the pitch he objects to is the one that arrives two minutes after a LinkedIn connection. What gets a meeting, by the summary's account, is a product that fits into Sentinel rather than one that wants to replace it. Novo still buys software, and the rule for what it buys is whether it integrates with the record Novo already owns.
What to ask before copying Novo
Ask how many years of structured records the company actually holds, and where they sit. Novo's RFI model works because fourteen years of RFIs were in one database in one format. A contractor whose RFIs live in three platforms and an email archive has a data project before it has an AI project, and Stoner's readiness sequence, workflow first and data second, is a description of that project. Ask who owns each automated workflow by name, because his rule is that the owner and not the model answers for a failure. And ask whether the workflow being automated is one a subcontractor would object to if they knew, because his one hard line is a workflow where the sub does not know.
What is missing from all of it is the same thing missing from most contractor AI accounts this year: a measured before and after on a job with a name on it. The 99 percent figure on insurance certificates is the closest, and it is a process metric, not a project one. Stoner does not name the two California Diffs jobs or put hours on them. What he offers instead is fifteen years of a decision most of his peers did not make, and a set of rules that do not depend on which model is best this month.