Ask ten people in construction what AI is doing for the industry and you will get ten different answers. Some describe autonomous plant and robotic bricklaying. Some describe a chatbot that summarizes a specification. Some describe a pilot that quietly ended after four months.

The gap between how AI is marketed to the industry and what it is actually completing on live projects is wide, and it makes planning difficult. Budget holders are being asked to approve spend against claims they have no way of testing.

A narrower question is more useful: on a real project, in a real week, what tasks is AI completing to a standard a professional would put their name to?

This guide answers that in two parts. First, what AI is doing for contractors delivering the work. Second, what it is doing for the claims professionals who deal with the consequences when delivery goes wrong.

The headlines are about machines. The production use is paperwork.

That is not a disappointing finding. Paperwork is where construction projects are won and lost. Arcadis has tracked the causes of global construction disputes for well over a decade, and the leading causes have stayed remarkably stable: failure to properly administer the contract, and poorly drafted, incomplete or unsubstantiated claims [1]. Both are document problems. Both are expensive. Neither is solved by a robot.

The adoption data points the same way. A Dodge Construction Network survey of 235 US general and trade contractors, run in late 2025, found 40 percent already allocating a dedicated AI budget and 51 percent actively evaluating AI-related changes across their teams [2]. The most common expectation, held by 85 percent, was simply spending less time on repetitive tasks. Three quarters expected AI to help them learn from historical project data.

The leading concern in the same survey is worth sitting with: 57 percent named lack of reliability or accuracy in AI output [2]. That concern is well founded, and it shapes everything below.

It is more useful to think in capabilities than in products, because products change and capabilities do not. Five capabilities cover almost all defensible construction use today.

Extraction. Pulling defined fields out of unstructured documents. Clause numbers and notice periods out of a contract. Liquidated damages rates and caps out of a schedule of particulars. Revision numbers and dates off a title block. Quantities out of a bill.

Classification. Sorting a large undifferentiated pile into categories. Which of 4,000 emails relate to the substation delay. Which RFIs are design queries and which are site queries. Which drawings are superseded.

Summarization and chronology. Condensing long documents, and assembling dated events into sequence. This is already in serious use in disputes: one industry panel demonstrated working across roughly 1,000 project documents to identify relevant material, summaries extension of time claims and list delay events [3].

Drafting to a fixed structure. Producing a first draft in a defined format: a scope of works, a procurement package, a daily report, a departures register. The value is not eloquence. It is that the first draft arrives already in the shape you wanted.

Consistency checking. Comparing one document against another and flagging where they disagree. Specification against drawing. Estimate against bid documents. Subcontract against head contract.

The common thread is that all five are high-volume, well-defined, and verifiable. Every output can be checked against a source document in seconds. That last property is what makes them safe.

The limitations are as important as the capabilities, and they are consistent across the tasks where teams get burned.

Causation. AI can establish what documents say. It cannot reliably establish why something happened or who carries the risk. Forensic practitioners are direct about the boundary: AI can parse schedules, extract dates, draft chronologies from timestamped events and compute float, but distinguishing genuine delay from float management, resolving ambiguous baseline candidates, selecting an analytical methodology and apportioning responsibility in novel circumstances all remain expert judgment [4].

Ground truth. AI knows the record. It does not know the site. If the as-built record is thin, wrong or contradicted by what actually happened, AI will confidently process the record it was given.

Defensible quantification. AI will produce a total whether or not the underlying data supports one. It does not stop and say the sample is too small or the rates are stale.

Commercial judgment. Whether to press a claim, concede a point, or price a risk is not a document-processing task.

Consistency with itself. The same input can produce different output on different runs. In drafting, that is a nuisance. In forensic work, it is disqualifying, because an analysis that cannot be reproduced cannot be defended [4].

The workflows that work share a pattern: AI does the first pass over volume, and a professional does the judgment and owns the output.

Bid review. AI produces a first-pass summary of scope, key dates, unusual obligations and onerous terms across a large bid package. The bid team decides what matters and what it prices.

Contract review. AI extracts key commercial clauses and compares them against the company's accepted positions, producing a draft departures register. Commercial staff set the risk ratings and decide what gets negotiated.

Procurement packaging. AI drafts scopes of work and pricing schedules from the specification and drawings. The package engineer checks scope gaps and interfaces, which is where the money is.

Document control. AI classifies incoming correspondence, flags revision changes, and maintains registers. Document controllers handle exceptions rather than data entry.

Estimate checking. AI cross-checks an estimate against the bid documents for missing items, double counts and arithmetic that does not reconcile. The estimator judges whether each flag is real.

Progress reporting. AI drafts daily and weekly reports from site notes, inspection records and photographs. The site team corrects the record before it becomes the record.

Claims work is the most natural fit for current AI capability, because claims work is document-intensive by definition, and it carries the highest penalty for getting it wrong.

Document review at scale. Narrowing tens of thousands of documents down to the set that touches a specific delay event. This is the single largest time saving available, and it is verifiable, because every document either mentions the event or it does not.

Chronology assembly. Building a dated event sequence from correspondence, site records, minutes and instructions, with each entry linked back to its source.

Delay event extraction. Listing the events pleaded in an opposing claim and mapping each to the documents relied on.

Contract mechanics checking. Testing a claim against the contract's own machinery. Was notice given, in the required form, within the required period, to the named recipient? This is mechanical, high-value, and frequently decisive.

First-draft narrative. Producing a structured draft of an extension of time or variation narrative that the claims professional then rewrites, evidences and stands behind.

Rebuttal preparation. Identifying where an opposing position is unsupported by the documents disclosed.

Claims work carries an evidential standard that ordinary project work does not, and that changes the rules.

Two AI failure modes matter here. The first is fabrication, meaning an invented document reference, date or schedule detail that does not exist. The second is irreproducibility, meaning different output from identical input. Either one, discovered in cross-examination, damages more than the point in dispute. It damages the credibility of everything else in the report.

The legal profession has already run this experiment publicly. As of May 2026 a widely cited academic database tracked roughly 1,490 decisions worldwide, over 1,000 of them in the United States, in which AI-fabricated material reached a court [5]. Sanctions escalated from a $5,000 fine in the first well-known case in 2023 to $15,000 per attorney, plus opposing counsel fees and double costs, in Whiting v. City of Athens before the US Court of Appeals for the Sixth Circuit in March 2026 [5,6]. That court set out a standard which travels well beyond law: no filing should contain any citation, whether provided by generative AI or any other source, that the professional has not personally read and verified [5].

Two further patterns are worth borrowing. Courts reserved their harshest treatment not for the error itself but for practitioners who defended or denied it [5]. And experienced dispute practitioners describe current AI as roughly an 80 percent tool, meaning a good start that a human must review independently, with a clear record of where AI was used and what quality control was applied [3].

Five rules cover most of the exposure.

Start retrospectively, not live. Take a project that has closed and a task where you already know the right answer, such as a contract review you completed or a claim you have already run, and put AI through it. You get a measurable error rate instead of an impression.

Then take one workflow, not six. The most common failure mode in AI adoption across industries is not the technology. It is deploying a tool without defining the task it is meant to perform, which is why MIT researchers found that 95 percent of enterprise generative AI pilots delivered no measurable return [7]. A single well-defined workflow, run for a month, with time saved and errors caught both written down, will tell you more than a platform trial.

AI in construction today is a document processing capability, not a decision-making one. That sounds modest until you consider that document processing is where contract administration fails and where claims are lost.

The contractors and claims professionals getting real value from it are not the ones with the most tools. They are the ones who picked repeatable, document-heavy tasks, defined precisely how they wanted those tasks done, and kept a named human accountable for every output that leaves the building.

References