AI in Construction

·

8 min read

The Change You Didn't Catch: Three Generations of AI in Construction

Copilots answer questions and workflows automate steps. The next gain comes from AI that takes a defined job off the team's plate and hands back only what needs a person.

Yaniv Holzer

CEO at Conndor

The five-minute problem

Imagine a highway widening project. A revised drainage plan arrives on a Tuesday. It changes a culvert the crew backfilled last week.

Depending on the contract and the resulting direction, the change may trigger a written-notice deadline. The project manager now needs to connect the revision to work already in the ground, find the supporting records, check the contract requirements, and decide what needs to happen next — while managing dozens of other issues.

Nobody misses that notice because they don't care. They miss it because nobody had five uninterrupted minutes to connect all the pieces.

That is the problem AI has to solve in construction: less work to manage, not just faster answers. And the people who most need that help usually have the least time to set up, check, and maintain another tool.

AI use is common. Operational adoption is still rare.

Individual use of AI is becoming common. Embedding it into how an organization actually operates is a different challenge.

The RICS 2025 AI in Construction Report, based on more than 2,200 professionals worldwide, found that 45% of respondents reported no AI implementation in their organizations, while another 34% were still in early pilot phases. Less than 1% reported having operational AI systems scaled across projects.

The barriers it identified are all forms of scarce capacity: a lack of skilled personnel (46%), challenges integrating with existing systems (37%), and data quality and availability (30%).

Meanwhile, Mastt's 2026 State of AI in Construction Project Management survey, which gathered responses from 108 construction professionals globally, found that 72.2% used AI at least weekly, while only 9.3% said they were using AI agents.

The surveys measure different things and are not directly comparable. But together they highlight an important distinction: people are using AI at work, while relatively few organizations have embedded it into how work gets done.

Three generations of AI in construction

AI in construction is developing through three broad generations: copilots that answer questions, workflows that automate a sequence of steps, and AI specialists that take on a defined job and bring people in for decisions.

Most teams will use all three at once. What changes between them is how much of the work stays with the person.

Think of three ways to work with a new project engineer. You can ask them questions. You can hand them a checklist. Or you can make them responsible for a job and ask them to bring you only what needs your call.

Generation 1: Copilots answer questions.

You upload a specification or an RFI and ask what it says. The AI provides a summary or an answer. It is fast and useful, but the person still decides what to ask, checks the answer, and works out what happens next. When tomorrow's document arrives, someone has to start again.

Generation 2: Workflows automate steps.

When an RFI response arrives, a workflow can identify the project, update the log, and notify the right person. But someone has to design it, maintain it, and handle every case it wasn't built for. Construction projects produce those cases every week.

Generation 3: Specialists take on a defined job.

A specialist has an objective, access to relevant project records, and clear limits on what it may do. It monitors new information, connects it to what is already known, prepares the work, and escalates anything that needs judgment or approval.

The person stays accountable for decisions. The specialist carries more of the work that leads to them.

One plan revision, three generations

Consider the same revision on a highway widening project. The scenario is illustrative.

A revised plan upsizes a culvert at a particular station from 36-inch to 48-inch pipe. Crews have already installed and backfilled the original. The change may affect excavation and backfill quantities, the traffic control phase, a subcontractor's scope, and the contractor's potential entitlement to additional compensation.

Here is how each generation would handle it.


Generation 1: Copilot

Generation 2: Workflow

Generation 3: Specialist

What starts it

A PM uploads both sheets and asks what changed

A new revision arrives in the document system

The specialist is already monitoring relevant project records

What it produces

A list of differences between the sheets

Differences routed to the PM and logged in a tracker

A dated impact record: what changed, where work is affected, supporting RFIs, daily logs and photos, missing information, and the contract clause to verify

What the PM still does

Identifies the backfilled culvert, finds records, checks the contract, and prepares the necessary response

Performs much of the same work and maintains the routing rules

Validates the finding, confirms applicable deadlines, and decides whether to pursue it

Where the risk sits

On the PM's memory and calendar

On the PM's memory and calendar

In a reviewable record surfaced while the notice window is still open

The plan comparison is the easy part. The value is in the work between noticing a change and knowing what to do about it: connecting it to work already in the ground, to the records that support the impact, and to the contract requirements that may start the clock.

A daily log can be complete and still miss the story. The specialist's job is to connect it to the rest.

The specialist does not decide entitlement, commit the company, or send a notice without authorization. Those responsibilities stay with the people accountable for them.

What this looks like in practice

On an infrastructure project, the owner rejected a contractor's claim for relocating trees, saying the work was already priced in the bill of quantities.

The project manager accepted that assessment.

But an AI specialist surfaced an earlier email describing an ancient olive tree on the site. That detail had not been connected to the commercial review.

The project team then revisited the contract. The special specification priced the relocation work separately, at approximately six times the standard rate, and took precedence over the bill of quantities.

The finding helped substantiate a potential entitlement valued at approximately US$32,000.

The important part wasn't that AI found an email. It connected a piece of information that had been overlooked to the project and contract context that changed its commercial significance.

This is the difference between finding information and doing the specialist work needed to make that information useful.

Other information-heavy professions are showing a similar direction.

At The Permanente Medical Group, ambient AI scribes draft clinical notes from patient visits, and physicians review and sign them. Across more than 2.5 million patient encounters, the American Medical Association reported an estimated 15,791 hours of documentation time saved.

The AI took on a defined job. The physician kept the judgment and accountability.

That is the model worth considering for construction too.

What makes this hard

Faster individuals do not automatically make a better business.

In McKinsey's 2026 State of AI survey, 80% of 1,719 respondents said AI had improved their individual productivity, but only 37% reported that AI had contributed positively to their organization's EBIT.

The gains show up when the work itself is redesigned, not simply when people get another tool.

For an AI specialist to make a meaningful difference on a construction project, three things matter:

Data. Daily logs vary by crew and region, and important correspondence can live in individual inboxes. A specialist that cannot access a record cannot connect it. It should identify what it could not find rather than guess.

Integration. Project records sit across project management platforms, ERPs, owner portals, shared drives, and email. If connecting them requires months of IT work, the specialist has become another project to manage.

Accountability. Every finding needs a source trail a PM or commercial manager can check quickly. Anything that commits the company or requires a commercial decision stays with an authorized person.

A smarter model does not make these problems disappear. Solving them is part of building a system that can be trusted with real work.

Five questions to ask any AI vendor

Calling a product an AI agent proves nothing. Before starting a pilot, ask:

1. What exactly is the job, and what does done look like?

"Help with documentation" is too vague. "Monitor plan revisions and prepare a traceable impact record" is a defined job that can be measured.

2. Which of our systems does it read, and how long will connecting them take?

Ask for a clear implementation timeline, the dependencies involved, and what your IT team will need to provide.

3. Can every finding be traced to a source record?

If a PM cannot check the evidence quickly, the team will end up redoing the work.

4. What can it do on its own, and what needs approval?

Get the boundaries in writing. Understand what it can read, prepare, update, communicate, and execute — and where human approval is mandatory.

5. Who maintains the system as the project changes?

Understand what the vendor maintains, what your team needs to manage, and who handles integration failures, changing requirements, and exceptions.

Then measure the pilot against a baseline from a recent comparable project.

Useful measures include:

  • Days from new information arriving to an identified potential impact.

  • Team hours spent collecting supporting records.

  • Percentage of findings the team validates as relevant.

  • Notice windows surfaced with sufficient time for review and action.

  • Number of manual steps still required to complete the job.

The point is to measure whether the team is doing less repetitive work and getting better operational outcomes — not simply how many tasks the AI claims to automate.

Give the five minutes back

Copilots will keep answering questions, and workflows will keep automating steps.

The next gain comes from AI that takes a defined job off the team's plate and hands back only what needs a person.

The question for construction leaders is no longer just where they can use AI. It is which jobs should no longer require a person to coordinate every step.

The goal isn't to save five minutes on one task and create another system to manage.

It's to give those five minutes back — and then more.

A plan revision is only the beginning.

See how Conndor's specialist get the job done.

See Conndor in action →

ABOUT THE AUTHOR

Yaniv Holzer

CEO at Conndor