Retrieval beats generation: the best AI use case in AECO is finding things

It is nine in the evening and the site office is down to one person. A project engineer, four documents open, trying to establish which specification clause governs a pour at six tomorrow morning. One clause came off the portal. One arrived by email in March. One was printed by somebody who has since left the job. The fourth is a PDF called spec_final_updated. She is not short of documents. She is short of any way to tell which one the site may build from.

A fortnight later the AI demo arrives. It writes a method statement, drafts an RFI response, produces a drawing. The slide photographs beautifully, and it answers a question nobody in that office at nine at night was asking.

The first useful job is smaller and duller than generation: find what the project already issued. These containers, this revision, this purpose, these recipients. Retrieval wins because generation quietly writes a second record, one the site was never given permission to build from. If a proposal cannot tell you what it will retrieve, it is already generating.

Generation makes a second record

An issued container is a dated act. A numbered transmittal, a revision the title block owns, a purpose or suitability, a list of companies who received it, and a predecessor marked as withdrawn. ISO 19650 is unromantic about the states — work in progress, shared, published — and each boundary is a decision somebody made on a date. A fluent paragraph sits in none of them.

What generation produces looks like information management from three metres away. A summary of "the structural issue". A rewritten clause. A sketch that resembles last week's general arrangement. The pixels are new; the issue is not. Nobody transmitted the paragraph, nobody marked the earlier revision superseded, and nobody can say who was entitled to rely on the sketch.

That is how the unofficial set in the trailer gets born without anybody behaving badly. The official set was hard to find, so someone asked a model what it said, and the answer came back confident and well organised. It was printed, or pasted into a chat, or built from. Months later the claim asks which identity was live on the pour date, and the generated file has nothing to say for itself. It was a second instruction that never went through a gate.

Generating a drawing is the same failure with geometry attached. Coordination is a comparison between issued models and issued sheets. A model that invents a riser, a grid line or a beam depth is proposing a design, and a proposal that cannot appear on a transmittal is no use to the gang installing at eight the next morning.

Generating text about the issued set is only slightly more respectable. A paragraph that cannot point at C03, the transmittal number and the date of issue is a rumour with better grammar. If it can point, you never needed the paragraph in the first place.

What retrieval has to mean

Retrieval here is not "search the folder". Last Thursday's ZIP, a document library, a mailbox full of plots — those are locations, and a location tells you only where files sit.

The issued set is bounded, and the boundaries are the whole value:

These containers — named sheets, models or documents, not whatever happened to be in the download.
This revision — the identity the title block declares, checked against the register rather than the letter in the filename.
This purpose — the suitability, or the action requested. S2 is not a licence to build.
These recipients — who received the issue as an act, not who had access to the folder.
This predecessor — what was withdrawn when the new identity arrived.

A tool that answers those five, and refuses when it cannot, is doing the job that pays for itself. A tool that summarises a folder is generating with extra steps. The pilot that dies usually dies at exactly this point: nothing pins the answer to the issued set, so every fluent sentence floats.

It is a reading-and-pointing job — register reconciliation at a wider radius. Read the title block. Find the current identity, the sheet nobody ever superseded, the transmittal that named the installer, the clause that was in the contract that was signed rather than the one that sounds as though it ought to have been.

The labour-exposure work hints at why this fits construction in particular. Construction's ordinary hours sit near the bottom of current AI coverage, while the hours that look like office work — find, check, retrieve, reconcile — sit nearer the top. Those are the only exposed hours the occupation has, and spending them on generation buys a second unofficial record with the one budget that might have bought a findable first one.

Five questions to put to the next proposal

Ask them in a room, write the answers down, and treat a shrug as an answer.

What does the live job already ask? Which revision is current for this pour. Who received C02. What was published for this package on a named date. If the proposed use is "write a method statement", it is generation, and you can stop there.

What is the corpus? Named issued containers, current revisions, stated purposes, named recipients. "We'll upload a folder" is a different project, and a longer one.

What must every answer point at? Container number, revision, issue record, date. An answer that cannot point is a sentence some software wrote. Treat it as unissued.

Who can stop a wrong retrieval? Models get it wrong. They return the superseded sheet with complete composure, or hand back a coordination issue as though it were for construction. Somebody with authority over the issued set has to be able to say this one does not go to the workface. If that person is "the model", or "everyone", you do not have one.

What happens to the leftovers? If the tool also drafts a summary or a suggested reply, that text is not an instruction. It does not get plotted, pasted into a transmittal, or pinned up in the cabin.

Run the five against one live package — the pour, the floor going in this week, the coordinated services zone — and time-box it to an afternoon. A pass is boring. The current sheet comes back, with the transmittal that carried it and the earlier identity marked withdrawn, and a person still decides whether the site may use it. A fail is fluent: three paragraphs about "the latest structural issue", no revision anywhere in them.

Why finding things is the use that lasts

An adjudicator, eighteen months later, has no interest in what a model thought. The questions are narrow and factual. Which sheet was live on the fifteenth. Who had it. What it was released for. What it replaced. Retrieval pinned to those facts can be checked by someone who dislikes you, which is the only test that counts. Generation cannot: you can produce the container, but not the proof that a generated paragraph was ever the instruction.

Construction has more artefacts than it can find, and no shortage of new text or new drawings. What is scarce is people who can put a hand on the identity that was live, at the workface and again six months later. ISO 19650, CDM, the Building Safety Act, NEC, FIDIC and JCT all assume that identity is knowable. None of them assumes a chatbot. They assume a contemporaneous record, kept by someone.

Where retrieval stops

Finding C03 does not publish C03. No amount of confidence in an answer changes who may release, withdraw or refuse a plot, and that should stay with a person who has a name and a role. Nor do five questions amount to a certification: the standards and the numbering protocols have owners, and none will accept a chat log as evidence that something was issued.

Three things worth doing

Only the last of them involves buying anything.

Fix the issue record first. Somebody has to reconcile the register against what was actually transmitted, title by title, and mark the retired identities as retired. On a live job that is weeks of a document controller's time, and the least popular line in an information-management budget. It is also unavoidable: every retrieval tool worth having reads that register, and none of them can invent it.

Then write the requirement into the appointment. Appointing parties and principal contractors set the information requirements, so they are the only ones who can make this a condition rather than a preference: any tool given access to project information answers with a pointer — container, revision, purpose, issue record, date — and stays quiet when it cannot. Put that in the tender documents, while it can still change a price. Added later, it becomes a variation and an argument.

Last, name the person allowed to refuse an answer and give them hours in the programme to do it. Usually that is the information manager or the lead document controller, and usually they are already at capacity, which is the part that costs money rather than goodwill.

Do those three and the engineer at nine in the evening is not choosing between four documents. She has one, with a date on it, a purpose, and a list of who else received it — and she can go home.