Your AI strategy is a data governance strategy wearing a costume

The board signed it off on a Tuesday. Use cases on slide four, a vendor already on the framework, a four-week trial that had shown promise. Nobody asked who would still be allowed to delete the upload when the champion's secondment ended, or who could refuse further training once the embeddings had left the building. The pack had a colour. Green. The paper underneath it was not in the pack.

Your AI strategy is a data governance strategy wearing a costume. Folders have names. Models have federations. Someone is “on BIM”. Someone is now “on AI”. The plotter still sings. In that version of events the strategy is a slide: use cases, a sandbox, last Thursday’s ZIP, a steering pack that is already green.

They are not the same object. A strategy that cannot name issued-set ownership, the metadata that would survive a claims file, and the access rules that say who may retrieve, train and export, is not an AI strategy. It is a data-governance gap in a better jacket.

They already bought the system of record

The assumption underneath most of these packs is that the industry already bought the system of record.

A common data environment exists. A federated model exists. A document controller exists. Therefore the pile is governable, and the remaining work is to point a model at it. The plot is flattering because it treats last year’s information-management spend as a completed act. The CDE was the strategy. The AI programme is the upgrade.

It is also how organisations lie to themselves with a noun. “System of record” is a purchase. Issued-set ownership is a habit. Metadata is a fact you can still query after the organigram has been redrawn. Access rules are a named person who may refuse a retrieval. None of those arrive with the licence.

Construction AI pilots mostly die when they are not pinned to the issued set. Retrieval beats generation because finding the issued container is the job that survives contact with a claims file. A consultancy that buys the model without a rule for the corpus has bought a demo with a balance-sheet line. If the data is not governable — named, versioned, permissioned, retrievable — the AI strategy is a costume.

What the costume actually is

The model can be fluent. The room can be full. The licence can be current. Costume is a capability that looks complete from the front row and has no issued-set pin, no metadata that would stand as an exhibit, and no access rule that would survive the champion leaving.

The costume has a typical shape. Someone with a budget books a sandbox. A folder of last week’s PDFs is uploaded. A chat window summarises a drawing. A progress slide says the firm is evaluating. Then the live work — the transmittal, the RFI, the inspection, the daily report — continues in the systems that already issue and withdraw information. The unofficial set in the trailer is still the set people trust.

That is a type, not a vendor programme. Nobody needs a named rollout, a maturity ladder, or a field note from a particular cabin to recognise it. The industry already has the type. The work is to say what the type lacks.

It lacks three objects, and they are information-management objects, not model objects.

Issued-set ownership — who may treat a retrieval as if it had issued, and who must stop it.
Metadata — the facts that make a file into a container: identity, revision, state, suitability or purpose, recipients, withdrawn predecessor.
Access rules — who may see work in progress, who may retrieve a published pack, who may export embeddings and prompt logs, who is left holding the corpus when the job ends.

Miss those and you do not have an AI strategy. You have a chat window pointed at a picnic of files.

Issued-set ownership

An issued set is a dated act: named containers, a revision, a suitability or purpose, a list of recipients, and a withdrawn predecessor. Whatever sat in last Thursday’s download is not that act. ISO 19650’s states are unromantic here. Work in progress is not Shared. Shared is not Published. A chat window pointed at a ZIP of last Thursday’s export is reading a pile. It is not reading the set the site is allowed to build.

Ownership of that act is the first thing the costume lacks.

“Owner” is not the name on the kick-off slide. The costume has a champion: someone in digital, innovation, a curious quantity surveyor, a design manager who saw a demo. Champions are necessary. They are not ownership. Ownership is the authority to change how issued work is done — or to record, with a date and a person, that the model was wrong and the issued information stands. If the register of the strategy cannot name that person, you do not have a deployment. You have a mailing list of people who were in the room when the demo was shown.

There is a second owner the costume also lacks: the owner of the exception. Models are wrong. Drawings conflict. A retrieved clause is from the wrong contract. Someone has to have the job of saying stop — this output does not issue, this answer does not go to the workface, this revision is not the one. If that person is “the vendor”, you have outsourced the only decision that matters. If that person is “everyone”, you have no one.

There is a third, and it only shows up after practical completion. The client paid for the design and will operate the asset. The contractor built from a published revision and will carry defects. The consultant holds the federated model. The sandbox holds a private copy someone uploaded because the official cupboard could not be trusted on a Tuesday. If the strategy cannot say which of those four may delete the upload, export a bounded pack, and refuse further training, the model has a parent and the project has a landlord.

Without that pin, every answer is ungrounded. The model can be fluent about a beam depth that was superseded on Monday. It can reconcile a register against sheets that were never transmitted. It can summarise a method statement that is not the one on the permit board. The site will still build the issued set. That is what issued means.

The pin is expensive. It requires a register that is true, a transmittal that is complete, and a habit of withdrawing the old plot from the people who still have it. A folder dump is cheap. It is also how you train a confident liar. The unofficial set in the trailer is the same failure in paper: when the official record cannot be trusted to name the current instruction, people keep a local copy of what they think was agreed.

If the strategy cannot answer which issued containers, which revision, which recipients and which purpose, it is retrieving over a picnic of files rather than doing construction information work. Retrieval over a picnic is how the costume starts. It is also how it stays a costume.

Metadata that would survive a claims file

Metadata is not a folder name.

A folder named Issued is storage. A container that can still be queried for state, suitability, revision and recipient is a fact. The tree encodes last year’s organigram. The facts are what a claims file, an adjudication timetable, or a handover will ask you to produce when the people have left.

The costume treats title-block text as metadata because the model can read it. Reading is not binding. A revision letter in a PDF is a picture of a letter. It is not the project’s claim that this container, in this state, was issued to these companies, for this purpose, and that the predecessor was withdrawn. A model that extracts “Rev C” from a crop is doing optical work. A document controller who can still produce the transmittal that made C the live instruction is doing information management.

Without those facts the model cannot tell Work in Progress from Published. It cannot tell a shared coordination model from the pack the gang is allowed to build. It cannot tell the unofficial set from the issued set, because the unofficial set is often the more current picture and the metadata never said otherwise.

The typical pile is ugly in a predictable way. Title blocks. Superseded revisions that were never withdrawn. Transmittals that prove a send, not receipt at the workface. A federated model that froze at practical completion. Inspection photographs with no asset tag. A method statement that is not the one on the permit board. Point a fluent engine at that pile and you do not get a strategy. You get a confident reconstruction.

Retrieval, not generation, is the use case that survives contact with a claims file. See also Retrieval beats generation: the best AI use case in AECO is finding things. Finding is still costume if the thing found cannot be named as an issued container. A paragraph generated about the issued set is a second, unofficial record. A retrieval that returns the wrong revision is the unofficial set with a nicer interface.

The claim that dies for want of a contemporaneous file does not become easier because the file was embedded. Embeddings are not exhibits. Prompt logs are not transmittals. A vector store is not a register. If you cannot reconstruct the act — these containers, this revision, this purpose, this list of recipients, this moment — you do not have metadata. You have a search index sitting on a disagreement the project never recorded as a disagreement.

Access rules, not seat licences

Permissioned is the third word the slide pack skips.

A seat licence says who may open the chat. An access rule says who may see work in progress, who may retrieve a published pack, who may export the embeddings, who may keep the prompts, and who is left holding the folder when the trial ends. Those are not the same sentence. The first is procurement. The second is governance.

The costume collapses them because collapsing them is cheap. Invite the steering group. Upload the ZIP. Let the model improve the service. The click-wrap is silent on whether the owner of the building can still export last month’s issued set without a vendor screenshot. It is silent on whether other employers’ jobs, uploaded under the same improvement clause, were carved out or silently folded into someone else’s method.

Access is also a labour question. Document controllers and information managers already spend their week making the issued set true. After the costume arrives they spend a second week asking whether the model’s memory is the same set, or a private copy the project will not get back. That is not innovation. That is a second unofficial set, this time in a vector store.

There is a workface version of the same hole. Issue is not receipt. Receipt is not the package that will place the work. A model that can retrieve a published PDF for anyone with a login has not solved last-mile access. It has widened the set of people who can hold a copy the hut has not withdrawn. Permission to retrieve is not permission to issue. If the strategy cannot tell those two apart, the chat window is another plotter.

Who may delete the upload. Who may refuse further training. Who may produce a bounded pack from project custody when the champion’s secondment ends, or the financial year closes, or the person who “gets AI” leaves for a contractor that still plots from USB. If those names do not exist, the strategy is already written as a sunset. The sunset is just waiting for a calendar.

Why the costume is cheaper than the job

The costume survives because it was never hired to govern the record.

It was hired to prove curiosity. That proof is useful in a bid, a digital appendix and a year-end pack. It is cheap compared with the work that follows: pinning the corpus to issued containers, writing metadata the project can still query, naming an owner who can stop a bad answer, writing access rules that survive a leaver, withdrawing the workaround, and checking that the hut no longer holds the old depth.

The time-box also protects the people who would have to decide. A four-week trial is an alibi. No one person can be blamed for a sandbox that “showed promise”. A named workflow with two overdue exceptions is a management object. Projects prefer the alibi.

There is a third reason, and it is informational rather than moral. A living deployment crosses systems the sandbox does not enter. The answer lives in a chat. The decision lives in a comment or an email. The change lives in a native model. The instruction lives — if it lives — in a published PDF and a transmittal. Until those are one record, “in production” has no single place to be true.

Automation percentages without a definition of the issued revision belong in that same pack. A ceiling on a consultant’s taxonomy is not a forecast, and it is not a licence to skip the three objects. What the thirty-nine per cent automation figure does and does not mean is the evidence. If the figure does not touch an issued revision, it is costume arithmetic.

None of this is a request for another product. It is why sandboxes accumulate. The clash report nobody closed is the same failure with a different engine: output without a close-out. Detection is a count. A pilot is a demo. A strategy, if the word is to mean anything, starts when someone can still say which containers were current, who received them, who may still retrieve them, and which model answer was allowed to remain.

What would count as a strategy

You do not need a better adjective than “transformational”. You need three objects you can argue with in a coordination meeting.

Issued-set ownership. The corpus is pinned to current issued containers — these sheets, this model state, this revision, released to these companies, for this purpose — and a named person can stop an output and withdraw a wrong instruction. The predecessor is withdrawn from the people, and from the model, who still have it.
Metadata. Every container the model is allowed to see carries the facts a claims file would ask for: identity, revision, state, suitability or purpose, recipients, date of issue. A retrieval that cannot return those facts with the file is a guess with a citation that will not stand.
Access rules. Who may see work in progress. Who may retrieve Published. Who may export embeddings, prompt logs and fine-tunes. Whether those residues are project property, vendor method, or scheduled for deletion at practical completion. Whether a bounded pack can be produced from project custody without a letter to someone else’s counsel.

If those three answers exist, you may still have a bad model. You do not have a costume. A bad model can be switched off. A costume stays licensed until the crane comes down, because nothing in the trial made issued work the point of the demo.

Close-out, if the word is to mean anything, is one of two recorded acts. The production path changes — a named workflow now runs through the tool, with the same issued-set pin, the same metadata, and a named owner for exceptions — and the old path is withdrawn. Or a dated decision says the trial is rejected: wrong corpus, wrong owner, wrong rules, and the sandbox is switched off. If neither act exists, the strategy is still in the slide. Painting it green is how organisations lie to themselves with a colour.

The board that signed the pack can still refuse to treat it as approved until three names exist on paper: who may stop an output, who may delete the upload, and who may refuse further training after the job. Digital can run the sandbox; they cannot write those names without information management in the room, and information management cannot write them without hours the programme has not currently got. The cost is the pin — a true register, complete transmittals, withdrawn plots — not another licence. Until those names are in the pack, keep the costume on the table if you must. Do not call it a strategy.