A contractor with offices in three cities asks its workforce planning system how many superintendents it employs. The answer comes back low. One office calls the role site superintendent. The other two call it project superintendent. Both spellings are in the record, neither is wrong, and the roster comes up short when a pursuit needs staffing. Nobody notices, because nothing is broken.
Table of Contents
About a quarter of contractors say . The standard advice is to clean the data first and automate second. That advice has held for a decade and it is good advice. Nobody had the budget or the hours, so nobody followed it. The change now is that cleaning the workforce record is a job you can hand to AI, and it should be the first one.
A firm of two thousand people cannot put somebody through every workforce record looking for stale assignments and expired certifications. The job has no owner, no budget line, and no end date. It is also exactly the kind of reading a model does quickly and cheaply.
What makes construction workforce data different
Construction has spent the past few years pointing AI at its messy data, and almost all of that effort went to the field. Badge events, job tickets, dictated notes, photos of equipment nameplates: all of it arrives from a sensor or a device, in volume, with a timestamp. The workforce record has no sensor. People assemble it, from what other people did on jobs, and it goes stale between the moments anyone needs it.
The questions an operations team asks of that record are experience questions. Who has run a hospital job? Who holds a current certification? Who has worked with this architect before? Generic HR systems do not carry those fields, and the tools that track the jobsite never touch them.
The standard data governance playbook assumes a company with a data team. It calls for a governance council, named data stewards, a warehouse table for quality results, automated checks on a schedule, and a ticketing queue with resolution targets. No construction operations department is getting budget for that.
Hugh Seaton, who works with construction firms on AI adoption, said this on a :
“So one of the things that you’ll hear people say is you can’t do AI if you don’t have your data in order. And the problem with that statement is the people who are listening to it are probably spending an hour a day on ChatGPT or whatever. So they’re already using AI and their data is where it is.”
Where errors enter the workforce record
The first errors arrive at setup, and none of them look like mistakes. Whoever configured the system in each office used the vocabulary that office already spoke. A firm that has grown by acquisition inherits one vocabulary per acquisition.
More errors accumulate once the work is running, as skipped updates. An assignment ends and nobody closes it, so the person shows as committed for another six weeks. A certification lapses on a date nobody is watching. Somebody moves between divisions and the old title follows them. Each of these takes thirty seconds to fix at the time. A year later it is an archaeology project.
Nothing in a contractor’s week amounts to an inspection of the workforce record. There is no submittal review for a roster. Nobody’s Monday includes reading it for things that stopped being true. A spreadsheet will hold a stale assignment forever and never mention it.
The error becomes expensive when somebody acts on it, and the most expensive version is a wage classification. Prevailing wage rules under Davis-Bacon classify a worker by the duties actually performed, not by the title on the payroll. Workers’ compensation reaches the same place from a different direction: where an employee’s payroll is split across classifications, the split has to be documented, and without those records an insurer can book all of that payroll at the highest-rated code. Neither regime accepts “that is what we call the role” as an answer.
Four ways to use AI on your construction workforce data
Â鶹´«Ã½’s lists four ways to point AI at the record itself: assisted audits, standardization, record enrichment, and ongoing health monitoring. Each move below is one of those, applied to a contractor’s roster.
Run the audit before you plan the cleanup
Start by reading the records. All of them. The reason the cleanup never begins is that nobody knows how big it is. A model can read every profile in the system and report what is wrong: expired certifications, assignments nobody closed, project histories that stop in 2023, offices that spell the same role three ways. Each of those is a plain-language question about the workforce, which is what Ask Â鶹´«Ã½ is built to take.
What comes back is a list with names and record numbers on it. Somebody on your staff still has to work through it, but a list can be assigned, split across a team, and finished. “Our data is a mess” cannot. The audit is the cheap half, and it turns the expensive half into a number you can take into a budget meeting.
Standardize role names across your offices
Sitzler eliminated 90 redundant role titles across roughly 250 staff. They did it by hand, before connecting their systems. That is the version of this move most contractors know: one analyst, one pass, a few hundred people. It was worth doing at 250. Nobody does it by hand across a national contractor’s whole roster, which is where the reading has to go to a model.
Matching records that describe the same thing under different names is close to a solved problem in other industries, and a roster of a few thousand people is a small instance of it. Two things make it the move to run first: the wage exposure above, and the fact that you can check the work.
More than half of contractors name as their chief concern about AI, ahead of data security. Pointing a model at the system of record is the highest-stakes version of that worry, and role normalization answers it directly. Write down the titles your company considers correct. That list is the standard. The model proposes a mapping from what is in the record onto the standard, and you read the mapping before a single field is written.
A model will match site superintendent to project superintendent without difficulty. What it cannot see is that a superintendent who runs a crew in one region and a superintendent who manages at job level in another are two different jobs with one name. Collapse those together and the data comes out cleaner and less true.
Fill gaps by cross-referencing the systems you already run
Most of the missing fields already exist somewhere in the company. The certification is in the HR system. The project history is in the PM system. The person who could confirm either is on a job site. The systems do not talk to each other. Scott Kramer, President and CEO at Spartan Contracting, described the cost of that: “Our software stack is mostly siloed information, which also leads to double, triple, sometimes quadruple data input.”
Cross-referencing what those systems already hold replaces that duplicate entry. The gaps that remain are not equally important, so rank them. A missing certification on somebody about to be proposed for regulated work is a different problem from a missing project from 2019. Working the ranked list keeps enrichment from turning into a completeness exercise nobody finishes.
Monitor completeness so the cleanup holds
A cleanup with no monitor behind it is a cleanup you will run again in two years. Pick the fields you would stake a pursuit on: certification status and expiry date, assignment end date, role title against the standard list, and the last two projects worked. Set a completeness threshold on those fields and put someone in charge of watching it, so a drop gets caught early, before it spreads across the roster. Â鶹´«Ã½’s own guidance on building an internal resume database makes the same point from the other end. It recommends that firms set accuracy standards for workforce records and hold to them.
Monitoring is a standing practice that the one-time cleanup pays for. Once the backlog is gone, staying current costs a few corrections a week.
What the model settles and what still needs a person
The four moves do not carry the same amount of automation.
| What the model settles | What still needs a person |
|---|---|
| Matching the same role spelled differently across offices | Deciding that a regional superintendent is a genuinely different job |
| Filling blank fields from a system that already holds the answer | Chasing a certification missing from every system you own |
| Proposing the mapping from your titles onto your standard list | Reading the mapping before it gets written to the record |
| Flagging which profiles are incomplete and by how much | Deciding which gaps are worth anyone’s Thursday |
The model handles the cases it is sure about and hands you the rest. Audits and monitoring work the other way: they generate work for people. Surfacing an expired certification is one act. Chasing it down is another, and only the first one is automatic.
Run the audit first
The audit is the cheapest of the four moves and the other three depend on it. It turns “our data is a mess” into a list with names and record numbers on it, and a list can be assigned to somebody and finished. Everything the guide says about agents rests on that record being trustworthy: an agent working from incomplete or inaccurate data does not produce better outcomes, it produces poor outcomes faster.
If your workforce record already lives in Â鶹´«Ã½, the audit starts as questions. Ask Â鶹´«Ã½ takes them in plain language, and the answers are the list. Give it to somebody, and give that person a Thursday.
