AI in Construction Claims: What Can Actually Be Automated, and What Still Requires Expert Judgment?

Artificial intelligence is beginning to enter construction claims and disputes. But the useful question is not whether AI can "do claims".

It is which parts of claims work are fundamentally exercises in processing, organizing, tracing and calculating, and which require interpretation, methodological choice and professional judgment.

That distinction matters.

A construction claim or expert analysis may involve thousands of documents, multiple programme updates, cost ledgers, instructions, RFIs, meeting minutes, daily reports and competing factual narratives. Much of the work required to assemble that material is systematic. It is also expensive and time-consuming.

But identifying a document containing a delay event is not the same as deciding what caused critical delay. Running a calculation is not the same as choosing the appropriate methodology. Drafting an argument from a set of records is not the same as forming an independent expert opinion.

For experienced practitioners evaluating AI for construction claims, this is the line that matters: analytical labor can be accelerated without treating professional judgment as automatable. What is far less obvious, and far more consequential, is how hard it is to accelerate that labor to a standard the work can actually bear.

That difficulty is consistent with the way established delay guidance treats analysis itself. AACE International describes Recommended Practice 29R-03 as a technical reference rather than a prescriptive standard and expressly states that it should be used alongside professional judgment. The Society of Construction Law's Delay and Disruption Protocol similarly places importance on transparency of information and methodology.

Why construction claims are unusually document-intensive

Claims analysis sits at the intersection of multiple project records that were usually created for different operational purposes.

The relevant evidence may be dispersed across contract documents and amendments; baseline and updated programmes; programme narratives; correspondence; notices and instructions; RFIs and responses; meeting minutes; daily and site reports; progress information; change and variation records; resource records; cost reports, timesheets and invoices; and witness and expert materials.

The SCL Delay and Disruption Protocol itself distinguishes programme, progress, resource, cost, correspondence and administrative records, and contract and tender documents as important categories for managing progress and substantiating delay or disruption claims.

The difficulty is therefore rarely simply finding a relevant document.

It is reconstructing what happened, when it happened, what the contemporaneous programme showed at the time, what other events were occurring, how the parties responded, what costs were subsequently incurred, and whether each proposition being advanced can be traced back to reliable evidence.

That reconstruction can consume substantial professional time before the harder analytical questions even begin.

The analytical labor is real, and harder than it looks

Before any expert question can be answered, an enormous amount of preparatory work has to happen. Records created for different purposes have to be assembled and structured. A chronology has to be built and tied back to source. Assertions have to be connected to the evidence that supports them. Schedule and cost data have to be processed. First-pass workpapers have to be produced from an unstructured repository rather than a blank page.

This is analytical labor, and it is where the near-term role for software sits. But "software can accelerate this" and "software can do this to a standard that survives scrutiny" are very different claims, and the distance between them is the entire problem.

A summary that cannot be traced back to the record it came from creates a second review problem rather than solving the first. A chronology that reads well but misattributes an event is worse than no chronology. A cost reconciliation that is arithmetically clean but silently mishandles entitlement produces a confident wrong answer.

Getting this work to a defensible standard is not a matter of pointing a general-purpose model at a data room. It takes depth in the specific contract standard, an understanding of how entitlement and quantum actually behave, and verification engineered into the workflow rather than assumed. That is craft, not commodity, and it is the difference between output that impresses in a demonstration and output a professional is willing to put their name to.

Where judgment cannot be delegated

Everything above surrounds the judgment; none of it replaces it. Methodology selection, causation and criticality, concurrency, the weight to attach to conflicting records, contractual interpretation, whether a cost is genuinely recoverable, and the final opinion itself remain the professional's, not the software's. These are not gaps to be closed by a better model; they are the exercise of expertise. Any workflow that quietly converts correlation into causation, or a calculated number into a recoverable one, is doing harm dressed as help. The right output surfaces the relevant language, records and issues for review. It does not present a position as a finding.

The hallucination problem is really a traceability problem

Most discussion of generative AI fixates on hallucination. For construction claims, the sharper standard is traceability.

A paragraph can be entirely correct and still be useless if nobody can establish where it came from. Conversely, an imperfect extraction is easy to correct if the reviewer can open the underlying source and inspect its context.

So the categories that matter in serious claims work must never collapse into a single confident block of prose: what was extracted directly from a record, what was calculated from defined inputs, what was inferred by the system, and what was concluded by the responsible professional after review. Holding those apart is what makes an output defensible rather than merely persuasive.

This is the principle Hecato is built around: source-linked analytical work product that preserves the path from any answer back to the underlying project record, with the expert in control of methodology, interpretation and the final opinion.

The distinction that should shape adoption

The near-term opportunity in construction claims is not an autonomous expert witness. It is AI software that compresses the vast evidential and analytical preparation standing between fragmented project records and the point at which expert judgment can properly be exercised.

That compression only counts if it is done to a standard the work can bear. The analytical labor is automatable; automating it defensibly is hard, and that difficulty is the whole point, not an inconvenience to be papered over.

Methodology, causation, concurrency, evidential weight, contractual interpretation and the final opinion stay with the professional.

Automate the analytical labor. Preserve the judgment.

The work is the same whichever side you sit on

This work is neutral by construction. The same record has to be read the same way whether you are acting for a claimant, a respondent, a contractor, an employer or a tribunal-appointed expert. Nothing in the analysis should assume the claim is valid, and nothing should assume the objective is to maximize a figure.

What a professional needs from AI software is the ability to interrogate the project record, test propositions against it, trace every position back to the document it came from, identify evidential support, gaps and contradictions, and analyze the relationship between events, delay and cost. That work substantiates a position where the evidence supports it and challenges it where it does not.

That is the approach behind Hecato: a platform that helps claims and disputes professionals move from fragmented project records to source-linked analytical work product across FIDIC, NEC, JCT, AIA and ConsensusDocs, while the practitioner keeps control of methodology, interpretation and final opinion.