“Is our firm ready for AI agents?” is the wrong question. It covers the whole company, so nobody can answer it, and the meeting ends where it started. The better question is smaller: can an agent take this one decision? That one has an answer, and you get it by asking two things. Can you list every input the decision uses? Can you write down what a good result looks like?
Table of Contents
Try it on something real. One superintendent is assigned to two jobs in the same week. The inputs are two sets of dates. A good result is the overlap, flagged before anyone promises a client anything. That check is already running at firms that are still holding meetings about whether they are ready for AI.
Below, three common workforce planning decisions go through the test. They land in three different places.
Why a staffing decision is harder to hand off than a submittal
Most general contractors already use agents on documents. Something routes a submittal, flags an RFI, or reads a daily log and reports what changed. Bridgit’s explainer on what separates an assistant from an agent covers that ground. Document work has one property that makes it easy to hand over: when it goes wrong, somebody notices right away. Route a submittal to the wrong trade package and the person who opens it sees the mistake and sends it back.
A staffing decision does not give you that signal. Pick one of four qualified people for an April start and a wrong pick looks like a reasonable pick. Nobody sends it back. The planner may frown at the name, change it by hand, and move on. The job runs and the work gets done. The cost of sending the second-best person shows up months later, when the client does not call for the next job.
That is why the second question in the test carries most of the weight in workforce planning and almost none in the document layer. Writing down what a correctly routed submittal looks like takes one sentence. Writing down what a good project team looks like at your firm means deciding what you value: client relationships, sector experience, who gets stretched. Most firms have never made those decisions explicitly.
Exposure is what makes operations leaders cautious about handing this over. Gaurav Singal of ConstructConnect put the unspoken version of the objection plainly: “Am I going to be the person who trusted the AI and got burned on a bid?” On accountability, he added: “The smart ones know the accountability doesn’t transfer when the software is wrong.” Sorting your decisions before you delegate any of them limits that exposure. The cost is an afternoon on three recurring decisions. The alternative is losing a pursuit because nobody can explain how the team was picked.
The two-part test for AI workforce planning decisions
Hugh Seaton is Managing Director of AI Edge, an AI consultancy that works with contractors. On a Bridgit webinar on AI agents in construction he gave the rule his team uses before building anything: “if you could list all of the inputs, in other words it’s a finite number, and you could write down what good looks like, or give it an example of what good looks like, what finished looks like, that’s most of what you need to make an agent.”
You can answer both questions about one task in a meeting.
The first question is a completeness check. An agent that meets an input it has never seen does not stop and ask. It carries on with what it has and produces an answer anyway. So you need to be able to name every input, on a whiteboard, today.
The second question decides whether the agent can exist at all. An agent needs a way to check its own answer without handing the work back to you. That means somebody writes down the standard it checks against. Seaton listed what fails here: “things that require judgment, things that are open-ended, things that have lots and lots of different possible inputs are really hard to make an agent for.”
Writing in ZDNet in January, Ethan Mollick said the people who thrive will “know what good looks like, and can explain it clearly enough that even an AI can deliver it.”
Most people say yes to the second question and find out otherwise when they try to write the standard down. Seaton, who has spent thirty years around process, says he has never had a standard operating procedure he followed perfectly, and that most teams cannot accurately describe their own steps. Answering honestly takes work. A later piece in this series covers what to do when your process only exists in people’s heads.
None of this depends on what the vendor calls the thing. Seaton, on the label: “I don’t think it matters if it’s an agent or something you know, or a ham sandwich. What matters is, is it reliably doing what you want and adding value to your team?” The test matters. The label does not.
Three workforce planning decisions, run through the test
The test gives one of three verdicts. “Not yet” is the useful one, because it names the specific piece of homework that is missing.
A scheduling conflict an agent can flag today
One person, two assignments, overlapping dates. Every input is a date already in your system. A good result is unambiguous: the overlap, surfaced before it becomes a promise. Nobody disagrees about the right answer.
This passes so cleanly that it has already shipped, and it is also the one most likely to be dismissed as too small to bother with. It is not small. A conflict caught on Tuesday is an email. The same conflict caught when a superintendent does not turn up is a schedule problem with a client attached. Start here.
A team shortlist that needs your definition of good first
Assembling a project team splits the test in half. The inputs can be listed, though the list is long: build type, market sector, availability, cost, who has worked with this client, who has worked with each other. The first question passes.
The second question is where most firms stop, and the reason is not technical. “Good” here means a strong team for this job, and that standard belongs to your firm. Three calls decide it:
- Does a repeat client relationship outrank sector experience? The PM the client already trusts, against the one who has built four of these.
- Do developing people go next to senior ones on purpose? That pairing costs something on this job and builds the bench for the next one.
- What do you give up when the ideal person is booked? Every shortlist reaches this question, and most firms answer it differently each time.
This does not have to be a single rule, Seaton says. An agent can hold several definitions of good and check itself against all of them. Somebody still has to write them down first.
The work this verdict names pays off whether or not you ever deploy an agent. The agents in Bridgit AI include Team Builder, which proposes a team and waits for a planner to confirm it. What that proposal is worth depends on whether your firm has said what a strong team means. The software is available now. The definition it runs against is still yours to write.
A hiring decision that should stay with a person
Deciding which project manager to hire fails both questions, and it should. The inputs include who will work well with a difficult client, who is ready to be stretched, who is quietly looking elsewhere. None of that sits in a database, and some of it has never been written down anywhere. There is no evaluation you could hand a machine that would let it check its own answer.
A test that says yes to everything is not a test. The near-term value of agents in workforce planning sits where something proposes and a person confirms. Bridgit’s Advanced AI Workforce Planning guide sorts agent work into assistive, semi-autonomous and autonomous, and puts the near-term value for contractors in the first two. Anyone promising more is selling past what has shipped.
Where the decisions land
Each verdict tells you what to do next.
| Verdict | What it means | What to do now |
|---|---|---|
| Ready now | Every input can be listed and “good” is unambiguous | Deploy it and stop debating readiness at company level |
| Not yet | Inputs can be listed, but nobody has written down what good means here | Write the definition down. That is the work, and it is yours |
| Never | Inputs live in people’s heads and no evaluation could check the answer | Leave it with a person and stop feeling behind about it |
Bridgit’s explainer on agents ends by telling teams to decide where the line sits and write it down together. Run that exercise on one decision and you have a verdict by the end of the meeting.
The middle row asks for a definition of good. For a staffing decision, that means naming which experience actually matters: what someone has built, for which clients, in which sectors, and alongside whom. A firm that holds a real record of who has done what can write that definition in an afternoon. A firm whose experience data lives in a spreadsheet and three people’s memories cannot, and for that firm the agent question was never the real obstacle.
Run the test on one decision this week
Pick the staffing decision that comes around most often, put the two questions to it in a meeting, and write the verdict down. If it lands in the ready column, you have found your first agent. If it lands in not yet, you have found the afternoon of work that was actually in the way, and it was never a technology problem. Either way, keep watching once something is running, because an agent that was right in March can be worse by September while the logs still look normal.
If you want something concrete to run the test against, Bridgit’s five agents are described one at a time, each with what it proposes and what it leaves to the planner. That is a reasonable place to see where your decisions land.
