Can AI build our golden thread for us?

It can help assemble and check one, and it cannot be responsible for one. The useful work is searching a large information set, pulling the entries behind a question, and finding what is absent. The judgement about whether the information is accurate, complete and safe to rely on stays with the dutyholder, because that is where the Act puts it.

Does using AI affect our competence declaration?

Competence requirements apply to people and organisations, so a tool does not add to or subtract from yours. What it can do is create a false impression of thoroughness. If a review was carried out by a model and signed by a person who did not read the material, the competence question gets harder rather than easier, and it is the person who is answerable.

Is it safe to let a model summarise safety information?

For orientation, yes. As the record of a decision, no. A summary is a lossy copy of something that already exists, and the golden thread needs the thing itself: the approved drawing, the test certificate, the instruction and who gave it. Use a summary to find the document. Do not let it replace the document.

What is the specific risk with AI search?

That it fails quietly. OpenAI publish an alignment test called broken search, covering exactly the case where a search tool fails and the model carries on as though it had not. On safety information a confident answer with nothing behind it is worse than no answer, because it stops somebody looking. Require the source entry with every answer and make it openable.

Where does AI genuinely help a dutyholder?

In finding gaps. Ask what is missing rather than what is there: which periods have no records, which changes were never formally approved, where an as-built does not match what was submitted. That is high volume, checkable work over your own information, and a gap found during construction is a correction rather than a problem for the safety case.

Do we need to record that AI was used?

It is sensible practice even where nothing requires it. If a register entry, a narrative or a gap analysis was machine generated, note that and note who checked it. The golden thread is supposed to explain why the building is safe and on whose authority, and how a piece of information came to exist is part of that.

AI agents and the golden thread: useful assistant, never the dutyholder

Why does this belong on a building safety site at all?

Because the golden thread is an information problem, and the tools for working with large sets of information changed materially this month.

Four new models were released in nineteen days in September 2026: three from OpenAI and one from Anthropic. The headline capability in all of them is that they search and act in a loop rather than answering once. You ask a question, the tool works out what to look up, reads what comes back, and goes again if the answer is not there. Anthropic describe their newest model as a reliable researcher that located hard to find sources in testing. OpenAI report their cheapest tier making about half as many factual mistakes as the version it replaced, at a tenth of the price.

Set that against what the golden thread actually is: a large, growing body of information that has to stay accurate as the building changes, and be findable by people who were not there when it was created. Searching it well is most of the work. So the capability is relevant, and the temptation that comes with it is obvious.

This page is about where the line sits.

What can these tools usefully do with a golden thread?

Three things, and they are not the ones usually advertised.

The first is finding the entry behind a question. Somebody asks when a detail was changed, or which drawing revision was on site in a given week, or who accepted a substitution. On a project with two years of records that is an afternoon of looking. It is now a question you can ask, provided the records exist in a form something can search.

The second is more valuable and gets less attention: finding what is absent. Ask which periods have no records at all, which instructions were never formally confirmed, which approvals have no paper behind them, where an as built condition does not match what was submitted. A gap found during construction is a correction. The same gap found when the safety case is being assembled is a problem, and the difference between those two situations is usually just when somebody thought to look.

The third is drafting. A gap analysis, a narrative, a register entry. Useful, and the least interesting of the three, because it is also where the risk sits.

Where does this break, and why does it matter more here?

Two things are worth stating plainly, because launch announcements do not lead with them.

The first is a number. On Agents' Last Exam, a test of agents on real professional tasks, the highest score in OpenAI's own published comparison was 59.3 per cent. That is a real capability and it also means roughly four attempts in ten came back wrong. In most settings that is an inconvenience. On information that explains why a building is safe, it is not.

The second is a failure mode with a name. Among the alignment tests OpenAI published with their newest models is one called broken search: the case where the search tool fails and the model carries on as though it had not. A tool that searches for you can be confidently wrong about what it found, and a confident wrong answer is worse than no answer, because it stops somebody looking.

That is the reason the practical rule below is not about model choice. Every answer needs the source entry attached and openable. Not a summary of the entry. The entry.

Can a duty be delegated to a tool?

No, and it is worth being precise about why, because the question is being asked in good faith by people who have a lot of information to assemble and not much time.

Duties under the Building Safety Act 2022 and the Building (Higher-Risk Buildings Procedures) (England) Regulations 2023 fall on named dutyholders during design and construction, and on the accountable person once a building is occupied. Competence requirements apply to people and organisations. None of that has a category for software, and none of it is discharged by explaining that a tool produced the output.

There is a sharper version of the point. The finding that produced the golden thread in the first place, set out in Building a Safer Future, was that nobody could reliably say what had been built, why, or on whose authority. A process where a model assembles the record and a person signs it without reading it reproduces exactly that condition, with better formatting. The information looks complete and no identifiable human judgement stands behind it.

Used well, these tools make the opposite easier. They make it cheap to look, and looking is what was missing.

Where does the line actually fall?

Task	Where a tool helps	Where a person must decide
Finding information	Searching the record and returning the entries behind an answer	Whether the entry found is the right one and the latest one
Gap analysis	Listing periods, approvals or documents with nothing behind them	Which gaps are material to safety and what is done about them
Change control	Flagging changes discussed but never formally recorded	Whether a change required notice or approval, and who authorises it
Handover and the safety case	Collating and structuring evidence against a naming convention	Whether the information is accurate, complete and fit to be relied on

The pattern is the same in every row. A tool is good at volume, retrieval and pattern. It is not a competent person, and the second column is where competence is exercised.

What should a dutyholder put in place first?

Five things, before any of this touches information that matters.

Decide what it may read, in writing. Specific folders, registers or systems, not blanket access. The connector standards that let a model reach internal systems are designed to be scoped, so scope them.
Require the source with every answer. If the tool cannot produce the entry it relied on, the answer is not usable. This single rule removes most of the risk in the previous section.
Never make it the last check. A machine generated gap analysis is a starting list for a competent person, not a clearance.
Record what was machine generated, and who checked it. The thread is meant to explain why the building is safe and on whose authority. How a piece of information came to exist is part of that.
Keep the thread somewhere you control. A vendor's chat history is not an information management system. The record has to outlive the tool, the contract and the people, which is the whole point of it.

What does good look like?

A dutyholder who uses these tools to look harder, not to look less.

The honest position is that most projects do not have an information problem a model can solve, because the information a model would need is not being captured in the first place. Decisions still get taken in conversations and never recorded, which is the subject of WhatsApp on site and the golden thread. Fix that and these tools become genuinely useful, because there is something for them to search. Skip it and the newest model in the world will confidently summarise a record that does not describe the building.

The duties are unchanged, and this page is information rather than legal advice. What has changed is the cost of checking your own information, and that has fallen a long way. The dutyholders who benefit will be the ones who spend it on finding what is missing while there is still time to put it right.

This page is information, not legal advice. It is written and maintained by a practitioner, verified against primary sources on the date shown above, and corrected fast when the regime moves. Spotted something out of date? Tell us.