Generative AI on the Project

Generative AI is the branch of the field that creates content - text, images, code - rather than only analysing it, and the text-generating kind, built on large language models, is what has swept into offices everywhere, construction included. Its point of entry is the enormous volume of routine writing that construction administration demands. Consider what a manager or engineer actually drafts in a week: progress reports and daily narratives; requests for information and their responses; letters and emails to clients, consultants and subcontractors; meeting minutes; scopes of work and instructions; and technical documents such as method statements and risk assessments. Much of this is templated, repetitive and formulaic - and therefore exactly the kind of writing a language model is good at producing a competent first draft of, fast.

The ways it is used cluster into a few patterns. As a drafter, it turns a short prompt - the key facts, the audience, the purpose - into a structured first draft of a report, letter or RFI that a human then edits, which is far faster than starting from scratch. As a rewriter and improver, it takes rough notes or a clumsy draft and makes them clear, correctly toned and professional - valuable on multilingual projects where the writer may not be composing in their first language. As an assistant, answered in plain language, it can explain a concept, suggest how to phrase a difficult letter, or help structure a document. As a summariser, overlapping with the previous lesson, it condenses long inputs. What unites these is that generative AI attacks the blank-page problem and the drudgery of routine writing, freeing time and lowering the barrier to producing the documentation a project needs. This is genuinely useful and genuinely here - no pilot, no data platform, no integration required. But precisely because it is so easy and so fluent, it is easy to forget what it is doing under the hood - predicting plausible words, not retrieving verified facts - which is where the risks begin.

It is worth being positive and specific about the upside, because this is one of the few AI applications that delivers immediately and broadly. The blank-page problem - the friction of starting a document from nothing - is one of the biggest hidden drains on an admin-heavy role, and generative AI dissolves it: a competent first draft in seconds means the work becomes editing rather than composing, which for most people is far faster and less draining. Speed on routine text is transformative: an RFI, a standard letter, a progress narrative, a scope description that might take half an hour can be drafted in a minute and refined in five. Quality and consistency rise too, especially for those who find writing hard or are working in a second language: the model produces clear, correctly structured, professionally toned prose, so the standard of routine communication improves and a real barrier - the discomfort of writing - is lowered. And unlike most construction-AI applications, this one needs no special data foundation: it works on any project, today, for anyone who can describe what they need. For an industry where productivity has stagnated for decades and where administrative burden is a genuine drag, a tool that measurably speeds a large slice of daily work is not hype - it is a real gain worth adopting.

But naming the gain honestly means naming its shape precisely, because that shape defines the safe way to use it. Generative AI is powerful at form - producing fluent, well-structured, appropriately-toned text - and unreliable at truth - the facts, figures and technical content inside that form. It is an excellent drafter and a poor authority. This is why the productivity is real but conditional: the speed comes from letting the model handle the form, and the safety comes from a human always owning the truth. Used as a first-draft engine whose every factual and technical claim is verified by the accountable person, generative AI is a genuine productivity tool. Used as a source of answers to be trusted and sent, it is a liability generator that produces wrong reports, mis-stated letters and dangerous method statements at the same impressive speed. The gain and the risk are two sides of the same fluency.

The first risk is hallucination, and it is inherent, not a bug to be patched. A large language model generates text by predicting the most plausible next words, not by looking up verified facts, so it produces output that is fluent and confident regardless of whether it is true. It will invent a plausible-sounding figure, misquote or fabricate a standard or clause, state a technical detail that is subtly wrong, or answer a question it has no basis for - all in the same authoritative tone as its correct output, with no signal that it has crossed from fact into fiction. In casual use this is a nuisance; in construction it can be serious. A hallucinated number in a cost or progress report misleads decisions; a mis-stated obligation in a letter creates contractual exposure; a fabricated code reference in a technical document is worse than none. Most gravely, method statements and risk assessments are safety-critical documents that describe how dangerous work will be done safely - and a generative-AI draft that includes a plausible but wrong procedure, omits a real hazard, or invents a control measure is a direct danger to life. The rule is absolute: every factual, technical and especially safety-critical claim in a generated draft must be verified by a competent, accountable human against real sources before it is used. Automation bias - trusting the confident, polished output - is the exact failure mode to guard against.

The second risk is confidentiality. Generative AI works from what you give it, and what you give it may not stay private. Many public AI services can store, log or use inputs, so pasting a confidential contract term, a commercial rate, a client's sensitive information or dispute correspondence into one can amount to disclosing it outside your control - a breach of confidentiality obligations, client agreements and possibly legal privilege. On a project handling commercially and legally sensitive material daily, this is a real and easily-made mistake. The disciplines are practical: use tools your organisation has approved and whose data handling you understand (enterprise arrangements that do not train on your data, or self-hosted models); never paste client-confidential or privileged material into a general public chatbot; anonymise or generalise where you only need help with form; and follow your firm's and client's data policies. Together these two risks define the safe posture: generative AI drafts, a human verifies every fact and protects every confidence, and nothing binding or dangerous is ever trusted to the draft alone.

The safe and productive way to use generative AI on a project is a single discipline: draft fast, verify always. Treat every output as a first draft, never a finished document. Prompt it well by giving it the real facts to work with - the actual figures, the specific context, the audience and purpose - rather than asking it to supply facts it will only invent; the model is there to shape your information into good form, not to be the source of the information. Then have the accountable person check every factual, numerical, technical and contractual claim against real sources before the document is used or sent. Build this into the workflow so verification is automatic, not optional, and so the person who signs the document owns its content whoever drafted it. Protect confidentiality by using approved tools and keeping sensitive material out of public ones. Used this way, generative AI is a genuine accelerator of the routine writing that clogs a project, with the human firmly in the loop.

Some things, though, must never be delegated to the draft, and it is worth naming them. Anything safety-critical - a method statement, a risk assessment, a safety instruction - can be drafted by AI to save time, but must be authored and verified in substance by a competent person who is accountable for the safety of the work, because a plausible-but-wrong procedure can kill and the duty of care never transfers to the software. Anything binding - a contractual notice, a commitment, a legal or commercial decision expressed in correspondence - is a professional determination that AI can help phrase but must not decide; the substance stays with the accountable professionals and the governing law. And any factual record that others will rely on - a report, a certificate - is only as good as the human verification behind it. The Indian context adds both promise and caution: generative AI is especially helpful on multilingual Indian projects for producing clear professional English or bridging languages, and it works today without any data foundation, so uptake is broad; but the same hallucination and confidentiality risks apply, safety-critical and contractual documents remain governed by Indian law, codes and standards (NBC India, the applicable IS standards, and construction-safety and labour law) and belong to the accountable professionals, and the discipline - draft fast, verify always, protect confidences, never delegate the binding or the dangerous - holds exactly.