A buyer's diligence framework for construction AI vendors

The demonstration will be a chat window. Someone drags last week's drawings into a panel, asks it about a beam, and a fluent paragraph comes back in a second or two. The words issued set get used, probably twice, in a tone suggesting the product knows what they mean on your job. Nobody asks the dull question, partly because the paragraph reads well and partly because the room is full of people who would rather not be the one slowing the meeting down.

What you have watched is a model answering over a folder. Whether the thing behind it is a construction information tool or a language model pointed at a heap of files is still an open question, and the demonstration is not designed to settle it.

So take questions in with you. The ones below are deliberately ones any supplier could sit, on a live package or in a sandbox you control; a list only one product can pass is a brochure with question marks in it. They pair with the thirty-one questions for a common data environment, which ask whether you are buying a CDE or a permissioned cupboard. These ask whether the model sitting on top of it can still be governed once the licence is live. Ask for each answer in the product rather than on a slide, and write down what you actually saw.

Before you sign, not after

Most construction AI pilots die beside the job rather than in it, and the pattern is monotonous: nothing pinned to the issued set, no owner once the champion moves to another project, no production workflow anybody is paid to run. Diligence done afterwards is an audit of a decision already made. Done beforehand, it comes down to four holdings a document controller can check without a consultant and without ranking anyone — who owns the training data, what the answers are pinned to, whether the record comes back out, and who carries the can when the model is wrong.

How to run it

Print the list. Do not email it over as a questionnaire for marketing to complete, because what comes back will be twenty paragraphs that are true in the way horoscopes are true. Sit it in a demonstration you control, on one live package that looks like yours.

For each question, accept one of four answers:

Shown — you watched the product do it, on data that looks like yours.
Configured — they changed a setting and then showed it. Note who will own that setting on day one.
Services — it exists if you buy a professional-services package. Write down the price and the lead time; that is the answer.
Not in product — which includes “on the roadmap” and “our partner does that.”

Resist the urge to score suppliers against each other. Score each answer against the job you actually run: one issued set, one RFI, one site engineer who will not wait a fortnight for a seat. A yes that cannot be shown is a no, however warmly it is delivered.

Who owns the pile once it is uploaded

The ordinary corpus is ugly and valuable. Title blocks. Superseded revisions nobody ever withdrew. Transmittals, RFIs, inspection photographs, a federated model that froze at practical completion, and the unofficial set in the trailer that someone scanned because the official cupboard could not be trusted on a Tuesday. A brochure calls all of that training data. On your job it is project information with duty-holders, recipients and a defects period attached, and uploading it does not change what it is.

If you cannot say who owns that pile after it has been ingested, you are not really buying a tool. You are granting a licence over the job.

Who owns the files, embeddings, prompts and derived weights once they have been uploaded — the appointing party, the appointed party who issued them, or the vendor?
Can the vendor use project information to train, fine-tune or evaluate a model that will later be shown to other customers, and is that use written into the order form or only into a click-wrap?
Can a named person on the project delete an upload and have the deletion reach the index, the prompt log and any fine-tune — or does “delete” mean hidden from the interface?
When a company leaves the job, do their past prompts remain attributable, and does their corpus leave without taking the issued set with it?
Is personal data in photographs, names on title blocks and site diaries covered by a data-protection impact assessment you can produce, or is that “handled by the platform”?

A right that lasts only as long as the subscription is not ownership. It is a lease of your own record.

What the answers are pinned to

An issued set is a bounded, authorised collection: these revisions, this purpose, this list of recipients, this moment. The newest files in a folder are not that, whatever the folder happens to be called. A model that cannot name the collection it read is retrieving over a picnic, and retrieval over a picnic is how you train a confident liar.

The CDE questions already ask whether the environment can produce a transmittal and withdraw a superseded sheet. These ask whether the model is reading that act, or a ZIP of last Thursday's export that somebody ran off for the pilot.

What is the pin — named containers, a revision, a suitability or purpose, a list of recipients — or is it a folder the champion uploaded for the demonstration?
When a revision is superseded, does the model refuse the withdrawn sheet, or can it still answer from a copy nobody took out of the index?
For one live answer, can you see the exact issued containers and revisions the model used, with the same identities your document controller would put on a transmittal?
Can a site engineer ask for the current issued set on one live package and receive that set, rather than a summary of whatever happened to be in the download?
If the pin cannot be shown, will the product still cheerfully let the answer be copied into an RFI, a mark-up or a site instruction?

If question six fails, stop. The rest of the tour is a chat window. Without a pin every answer is ungrounded, and ungrounded answers are fluent about beam depths that were superseded on Monday. The model will reconcile a register against sheets that were never transmitted. It will summarise a method statement that is not the one on the permit board. Meanwhile the site carries on building to the issued set, because that is what issued means.

Getting the record back out

The appointing party will want the record after the licence ends. A claim, a facilities team inheriting the building, a later enquiry, a regulator asking what the model said on a named date — none of those arrive conveniently timed to your renewal. Export is not a courtesy, and it is worth about as much as you can prove on the day you need it. The CDE list already asks whether containers, metadata and the audit trail leave in open formats. Ask the same of prompts, answers, citations and the pin.

Can you export every prompt, answer, citation and issued-set pin in open formats you can open without the product?
Is that export complete — withdrawn answers, human overrides, the identity of whoever closed the thread, the containers the model claimed to have read — or only the chat the interface still displays?
Who is entitled to run the export: the appointing party, the information manager, or only the vendor's services team? And what does it cost the second time?
How many days after the contract ends can you still take a complete export, and does that right live in the order form or in the demonstration script?
From the export alone, can you rebuild what the model said about one named package on one named date, and which issued revision it was pinned to?

Ask to watch an export happen, then open the files on a machine with none of the vendor's software installed. If you cannot read them there, you have been shown a download button rather than an exit. Embeddings are not exhibits. Prompt logs that die with the seat are not transmittals. A confidence score is not a suitability.

Somebody has to be able to say stop

Models are wrong. Drawings conflict, a retrieved clause turns out to belong to a different contract, and a plausible sentence walks into an RFI before anyone reads it twice. Somebody needs the standing to say that this output does not issue, that this answer does not go to the workface, that this is not the current revision. If that person is the vendor, you have outsourced the only decision that matters. If it is everyone, it is nobody. What you are looking for is a named closer, a residual duty written down, and a record of the misses.

When the model is wrong, who is contractually liable for the instruction that followed — the vendor, the appointing party, the designer who issued the sheet, or the person who pasted the answer into the RFI?
Is there a named closer on the project who may reject the model's sentence and leave the issued set standing, and is that rejection recorded as an act?
Does the order form describe the product as an aid to a competent person while the marketing treats the output as an issued instruction?
If the model misses — a superseded sheet treated as live, a hazard not flagged, a clause from the wrong appointment — who owns the reconstruction, and can they produce the pin and the prompt without three weeks of forensics?
Does the licence try to move a CDM duty, a Building Safety Act holding or a contractual obligation onto a click-wrap, and will your solicitor accept that sentence?

A click-wrap does not appoint a principal designer, and a model does not become a duty-holder by speaking in the first person. Regulation 8 still says a designer or contractor must not accept an appointment without the skills, knowledge, experience and organisational capability. Software can feed those duties. It cannot hold one.

The miss is the expensive gap, because it arrives dressed as clearance. Site teams over-read a green score: the model did not object, therefore the sheet is current; the chat summarised the method statement, therefore the isolation is good. Cost that failure first. Then read the indemnity, then name the closer.

Category	Questions	If this fails
Who owns the pile once it is uploaded	1–5	You are granting a licence over the jobnot buying a tool
What the answers are pinned to	6–10	Stop the tour — every answer is ungrounded
Getting the record back out	11–15	Export is a download buttonnot an exit
Somebody has to be able to say stop	16–20	The vendor has been handed the only decision that matters

Reading the marks afterwards

You do not need a weighted score. You need the holes.

Mark every question that was not Shown. A hole in the pin ought to end the argument about whether this is a construction information tool, whatever the steering pack says three months later. Holes in ownership or export belong in the order form — the exit, the retention window, the prohibition on training — before anybody's logo goes on a slide. A hole in liability is the reason an unofficial set will exist by week four, because people keep a local copy of what they think was agreed when the official answer cannot be trusted to name the current instruction.

Hand the marked list to the person who will run the register, not only to the person who signs. If your document controller cannot picture themselves asking question eight on a wet Tuesday, do not sign.

Run it beside the CDE list rather than instead of it. A model pinned to a permissioned cupboard is still reading a cupboard, and a CDE that can issue and withdraw properly, paired with a model that cannot see the pin, is two purchases that never meet.

What twenty answers will not buy you

ISO 19650, CDM, the Building Safety Act, NEC, FIDIC, JCT, COBie and MOMRA each have owners, and not one of them delegates to a checklist. Twenty answers of “shown” will not make a project ISO 19650-compliant, and they will not shift a duty by a millimetre. Compliance is a process your organisation runs. Software either holds the mechanics that process needs — a pin, an export, a named closer — or it declines to, quietly, in a window that always has something to say.

The person who will sign — not the champion who sat the demonstration — should treat an unshown pin, an unproven export or an unnamed closer as a reason not to sign. Put those three in the order form, with a price on the second export and a named human who may reject a sentence, and pay a solicitor to keep the sentences out of a click-wrap. That costs a slower purchase. Signing without them is how the unofficial set exists by week four.