Who Embeds on the Jobsite as Your Site AI Engineer?

Start With the Coordination Meeting Nobody Can Act On

Thursday coordination. The mechanical contractor says the main duct run clashes with a beam at grid line 12, the BIM coordinator pulls up the model, everyone agrees to look at it. Three weeks later the same clash arrives as a change order. The data existed the whole time: the model, the RFI log, the schedule, the submittal register. Nobody in the room owned turning it into something the field could act on by Friday.

That absence is what a site AI engineer fills. Suffolk Construction has been hiring the title in Boston and New Haven and, per Construction Dive, embedding AI engineers directly on jobsites under its Jobsite of the Future program, with CEO John Fish and CTO Jit Kee Chin framing it as a permanent operating change rather than a pilot ¹. The scope in those listings is deliberately broad: drawing conflicts, coordination gaps, supply chain risk, pay application flow.

The category is new enough that the title is not stable. Postings surface on ZipRecruiter, Glassdoor, Lensa and LinkedIn under a handful of names ¹, and most general contractors still keep this capability inside corporate IT or a VDC group. Hiring one now means writing the job yourself, which is an advantage rather than a burden: the scope gets defined around the delivery problems on your projects instead of around somebody else's template.

What Separates a Real Site AI Engineer From a Dashboard Builder?

The tell is what a candidate does with an ambiguous field problem. A dashboard builder asks what data is available and proposes a view. A site AI engineer asks who makes the decision, when they make it, and what they do today instead. Then they build the smallest thing that changes that Friday decision, and they go stand next to the superintendent to see whether it gets used.

Ask for a walkthrough of something they built that people kept using after they stopped maintaining it. Real answers are unflattering in specific ways: the first version was wrong, a foreman told them the schedule activity codes were fiction, they rebuilt around what the crews actually tracked on paper. Performed answers stay at the level of stack and accuracy percentage. The number matters less than whether the person can say why anyone believed the number.

Ask how they used AI to get good at this work, and listen for friction rather than fluency. The strong answers name a moment where the model was confident and wrong: an extraction pass over 400 submittals that quietly mislabeled revision letters, a schedule summary that produced a float figure nothing in the file supported, a code lookup citing a section the adopted local amendment had changed. People who work that way have built the habit of checking a confident claim against a primary document, which is the same habit that keeps a bad clash report from reaching a subcontractor. It is the trait doing most of the work in an AI estimate validation specialist too, where the entire job is deciding when a generated number deserves belief.

One more trait is harder to test and worth as much as the rest: tolerance for a trailer. This person spends real days in boots, in noise, being interrupted, and being the least visibly busy body in a room of people under schedule pressure. A candidate who describes the field as a data source rather than as a workplace tends to be back at headquarters within two quarters.

Which Backgrounds Actually Produce This Person?

Two obvious pipelines and one that gets overlooked. The obvious ones are VDC and BIM coordinators who taught themselves to write real code, and software engineers with some construction in their history: a parent in the trades, a summer on a crew, a few years at a construction technology vendor. The third is the project engineer who automated their own paperwork out of irritation and got caught doing it.

The VDC path produces someone who already speaks Revit, Navisworks, coordination sequencing and the politics of a clash meeting. The weakness is engineering discipline: no version control, no tests, tools that die when the author rolls off the project. That gap closes in a quarter with a code review partner and a repository. The domain knowledge underneath it does not close in a quarter, which is why this path is usually the cheaper hire even when the resume looks lighter.

The software path is the mirror image. The engineering is there and the field is not, so the risk is a beautifully built thing nobody asked for. Nine months of a genuine rotation fixes it, meaning a desk in the trailer and a superintendent willing to say plainly when the tool is wrong. Screen this candidate for curiosity about the work rather than about the data, and ask what they think a submittal is before you tell them.

The third path is the one most contractors miss, partly because the person is not applying for an engineering job. A project engineer who built a spreadsheet that reconciles pay applications, or a scheduler who wrote a script to diff two schedule exports, has already proved the two hardest parts: they can find the friction, and they finish something without being assigned it. Look inside the company first. The digital project controls analyst on your largest job is often one conversation away from this role, and already has the credibility with the field that an outside hire spends a year earning.

Where Do You Find a Field-Resident AI Engineer in 2026?

Not on a general AI job board, where the posting competes with every model company in the country and loses on both money and prestige. Look where construction technology people already gather: VDC and BIM user communities, the Autodesk and Procore ecosystems, construction technology meetups in the cities where you actually build, and the university programs that pair construction management with computing.

Those programs are an unusually good source because the hybrid already exists as a degree. Stanford's Center for Integrated Facility Engineering community and Carnegie Mellon's advanced infrastructure systems group both produce people who chose this intersection on purpose, which is a stronger signal than a general machine learning graduate who is willing to try construction. Recruit them a year early and let them spend a summer in a trailer before either side commits.

Watch the visible hirers too. Suffolk's own postings work as a recruiting map ¹: the cities named tell you where the competitive market for this person currently sits, and the people hired into those seats become the second generation of candidates for everyone else in about two years. Ask your own subcontractors as well, since the BIM lead at a large mechanical contractor has usually solved a version of your problem already.

Poaching from construction technology vendors works, with one caveat worth an interview question. A solutions engineer at a construction software partner has seen a hundred jobsites and shipped against real workflows, but has rarely owned a delivery they could not walk away from at the end of an engagement. Ask what happened after go-live on the account they are proudest of, and whether they were still there six months later.

Close Them on the Jobsite, and Pay Against the Right Band

Close on autonomy and proximity rather than on the technology. This candidate has offers from companies with bigger models and better laptops. What those companies cannot offer is a live building with a schedule, a superintendent who will tell them the truth at 7am, and permission to ship something on Tuesday that changes what happens on Thursday. Name the specific project in the pitch.

On money, the honest structural answer for a title this new is the band question rather than a number. Pay against the software engineering band in the market where the project sits, not against the VDC or project engineering band, and be ready to defend that internally when someone points out the person works out of a trailer. The premium behind that advice is real but general rather than construction-specific: PwC's 2026 AI Jobs Barometer reports an average wage premium of 62 percent for roles requiring AI skills, measured across roughly a billion job advertisements ², as of that report's publication. Treat any point salary figure for this exact title, including one a candidate quotes at you, as an anecdote until you have three of them.

Remote is not on offer here, and pretending otherwise wastes a month of everyone's time. The Suffolk version of the role is field-resident by design ¹, and the value sits in the overheard conversation. What is negotiable is the shape of the week: four days on site and one at home for deep work is defensible, as is rotating one person across two nearby projects. What breaks the role is assigning them to five jobs in three states, at which point they become a corporate consultant in a hard hat and the coordination meeting goes back to being unactionable.

Write the offer with a review date on the scope. The category is still forming, and the job posted this quarter will not be the job in eighteen months. Say that to the candidate directly. The people worth hiring into a role this new are the ones who hear an unfinished definition as room to work rather than as a risk.

See a sample report >