Most of the effort on AI in capital delivery goes into producing documents. The better return is in checking the ones that already arrive, the slow work that usually gets skipped.
Tools now draft progress reports, summarise meetings, fill in bid responses and turn a stack of documents into a tidy page in minutes. That has its uses, but it addresses the easier half of the problem, because a shortage of documents was never what held a programme back.
The better use runs the other way. An owner receives a steady flow of material from its supply chain, monthly reports, revised programmes, cost forecasts and, at handover, registers of everything that has been built. Most of it is read for its headlines and filed. The checks that would test it are easy to describe and tedious to do. The forecast should agree with the programme it claims to rest on, and every asset in a handover register should have a location and someone responsible for it.
Those checks get skipped for ordinary reasons. The work rarely finds anything on a given day, and whoever does it gets no credit when it comes back clean. The discrepancy it would have caught turns up months later instead, as a forecast that jumps without explanation or an asset nobody can find when it fails. By then the cost has landed somewhere else.
This is work a model suits well. It does not get bored, so a reconciliation that a person would do carefully in the first month and skim by the sixth gets the same attention every month. It can read two documents against each other and report where they part company, and running that on everything that arrives costs little. What comes back is a short list of places where the documents disagree, which takes far less reading than a generated report and tells the owner more.
Someone still has to read what comes back. Some flags will be wrong, because a model can misread a table as easily as a tired reviewer can. Telling a timing difference from the first sign of a real problem still needs a person who knows the job. What changes is where that person's attention goes, which is now to the few places where the documents have stopped agreeing.
For an owner, this kind of checking is also a cheap place to start. It works on the documents already arriving each month, and the main addition it needs is a written statement of what should agree with what. That statement is worth having for its own sake, because the knowledge behind it usually sits in a few people's heads.
Most owners already hold more documents than anyone will read. The useful question to put to a model is which of them disagree.