AI and Data in Forensic Delay Analysis: Promise, Pitfalls, and Professional Judgement

Consider a familiar scenario. A major infrastructure contract has run eighteen months late. The tribunal has before it 47 baseline revisions, some 900 weekly progress updates, three years of site diaries, tens of thousands of emails, a document management system holding a quarter of a million records, and two expert reports that reach opposing conclusions on causation. The delay analysts on both sides have, in effect, sampled this ocean of data. Each has selected windows, chosen an as-built dataset, and formed a view. The tribunal must decide which analysis better reflects what actually happened. It is precisely this problem of volume, selectivity, and the reproducibility of expert reasoning that artificial intelligence and modern data techniques are now being asked to address.

The question for senior construction, commercial, and legal professionals is not whether AI will feature in forensic delay work. It already does, quietly, in document review platforms and scheduling tools. The question is where it genuinely improves the reliability of analysis, where it introduces new risks, and how its outputs stand up when tested in dispute resolution.

WHAT THE DATA PROBLEM ACTUALLY IS

Forensic delay analysis has always been constrained by the quality and accessibility of records, not by the sophistication of method. The SCL Delay and Disruption Protocol has consistently emphasised contemporaneous records as the foundation of any credible analysis. The recurring difficulty is that records are voluminous, inconsistent, and dispersed across incompatible systems. Progress updates may be unreliable, as-built dates may be reconstructed after the fact, and the narrative in correspondence rarely maps cleanly onto activities in a programme.

This is where data techniques earn their place. The core value of AI is not in reaching conclusions; it is in organising, connecting, and surfacing information at a scale no team can manage manually. When 250,000 documents can be reviewed in days rather than months, the analyst spends less time locating evidence and more time interpreting it.

WHERE AI IS ALREADY ADDING VALUE

Several applications are now mature enough to be used with confidence, provided they are supervised.

Document review and technology-assisted review (TAR) have been accepted in litigation for over a decade. Predictive coding and clustering allow large disclosure sets to be prioritised, de-duplicated, and searched conceptually rather than by keyword alone. For delay work, this accelerates the identification of the contemporaneous record that underpins any defensible as-built.

As-built reconstruction is being improved by natural language processing that extracts dates, events, and references from site diaries, daily reports, and correspondence, then cross-references them against programme activities. Rather than replacing the analyst's judgement on what an activity's true start and finish were, these tools assemble the candidate evidence for that judgement and flag inconsistencies between sources.

Schedule interrogation is another practical area. Programmes can be parsed automatically to detect open ends, negative lags, excessive constraints, high float values, out-of-sequence progress, and logic changes between revisions. A tool that maps how the critical path migrated across 47 baseline revisions does in minutes what once took an analyst weeks, and it does so consistently.

Disruption and productivity analysis, long the poor relation of delay work, benefits particularly. Measured mile analysis depends on identifying comparable impacted and unimpacted periods within noisy production data. Pattern recognition across labour allocation, output records, and cost data can help locate genuinely comparable baselines and quantify the productivity differential more objectively.

THE LIMITS THAT MATTER

None of this displaces the analyst's core function, and it is important to be precise about why.

First, causation is a matter of reasoned judgement, not correlation. An algorithm can establish that an activity finished late and that a variation was issued in the same window. It cannot, of itself, determine that the variation caused the delay to completion, still less resolve concurrency or apportion responsibility. Those questions require an understanding of the contract, the facts, and the counterfactual, and they remain squarely the province of the expert.

Second, the choice of method is not a data problem. Whether an as-planned versus as-built, time impact analysis, windows, or collapsed as-built approach is appropriate depends on the contract, the quality of records, the timing of the assessment, and the nature of the dispute: the considerations set out in the SCL Protocol. AI does not select the method; it executes and evidences the method the expert has chosen.

Third, and most seriously, generative AI introduces the risk of fabrication. Large language models can produce fluent, confident text that is factually wrong, including invented citations. There have been well-publicised instances in litigation of practitioners submitting AI-generated authorities that did not exist. In a forensic context, where the expert's duty is to the tribunal and every proposition must be traceable to a source, an unverifiable output is worse than useless.

Fourth, there is the black-box problem. An expert must be able to explain, defend, and reproduce their analysis under cross-examination. A model whose reasoning cannot be articulated, or whose results cannot be replicated from the same inputs, undermines the very transparency that gives an analysis weight. Reproducibility is not a technical nicety; it is a condition of admissibility and persuasiveness.

DATA QUALITY AND GOVERNANCE

The value of any AI-assisted analysis is bounded by the quality of the data fed into it. Garbage in remains garbage out, now at greater speed. Where progress updates were manipulated, where as-built dates were reverse-engineered, or where records are simply absent, no algorithm can manufacture reliable evidence. If anything, the apparent authority of a data-driven output makes uncritical acceptance more dangerous.

This places a premium on data governance during the project, not merely during the dispute. Contractors and owners who maintain structured, well-organised, contemporaneous records, properly dated, version-controlled, and linked to the programme, put themselves in a far stronger evidential position. The same investment that supports good project controls also produces the clean datasets on which credible forensic analysis, AI-assisted or otherwise, depends.

There are also confidentiality and privilege considerations. Feeding privileged or commercially sensitive documents into third-party AI services can waive privilege or breach confidentiality obligations. Any deployment must sit within a controlled environment with clear data-handling protocols, and legal advisers should be involved in that decision from the outset.

HOW TRIBUNALS ARE LIKELY TO RESPOND

Tribunals value transparency, contemporaneity, and reasoned judgement. An analysis that uses AI to organise records and interrogate programmes, while the expert retains and explains every substantive judgement, is entirely consistent with those values. An analysis that outsources reasoning to an opaque tool is not. The distinction will increasingly be probed in cross-examination: which parts of the opinion are the expert's, and which the machine's; whether the process is reproducible; and whether the underlying records support the output.

The prudent position is to treat AI as an instrument of efficiency and rigour that must be disclosed, supervised, and verifiable, never as a substitute for the expert's independent duty.

A PRACTICAL POSTURE

For firms and their clients, three principles hold. Use AI to handle scale, such as disclosure, extraction and schedule interrogation, where it demonstrably outperforms manual effort. Keep judgement human, particularly on method selection, causation, concurrency, and float. And verify everything, because an assertion that cannot be traced to a contemporaneous record has no place in a forensic report.

The technology is changing quickly; the professional obligations are not. The analyst who combines strong data discipline with sound method and defensible judgement will produce better work, faster. The one who mistakes fluency for reliability will produce something that fails at the first serious test.

REFERENCES

1. Society of Construction Law, Delay and Disruption Protocol, 2nd edition (2017).
2. AACE International, Recommended Practice No. 29R-03, Forensic Schedule Analysis.
3. AACE International, Recommended Practice No. 25R-03, Estimating Lost Labor Productivity in Construction Claims.
4. Keith Pickavance, Delay and Disruption in Construction Contracts, 4th edition.
5. FIDIC Conditions of Contract for Construction (Red Book), 2017.
6. Nicholas Dennys and Robert Clay (eds), Hudson's Building and Engineering Contracts.