AI compliance site documentation in construction: from Friday write-ups to a live golden thread

Stage 1 is not the absence of documentation — construction produces more paper per pound of revenue than almost any industry. It is the absence of evidential value in the documentation that exists. The diary entry written three days after the pour, the briefing register signed in a batch at the end of the month, the photo that proves the rebar was inspected but sits undated in a leaver's phone: each record exists, and almost none of it would survive cross-examination, because nothing ties it to the time, place and person it claims to describe.

The tell is what happens when someone asks a precise question. A client's engineer asks for the evidence that the fire-stopping in riser 3 was inspected before it was boarded over. At stage 1 the answer is a reconstruction: emails are searched, subcontractors are phoned, a foreman is asked what he remembers about a Tuesday eleven months ago. The record set the project needs exists in fragments across twenty inboxes and eight phones, and the person assembling it is doing archaeology, not retrieval.

This stage is expensive in a way that rarely appears on a cost report. Disputes settle worse because the diary is thin exactly where the delay happened. Insurance claims drag because the condition photographs cannot be dated. And when the HSE investigates an incident, the gap between what was done and what can be proven was done becomes the story. The work was usually done; the site just cannot demonstrate it.

Stage 2 is where most of the industry sits, and it is a comfortable place to be stuck because every visible artefact of modernity is present. There are tablets on site, a CDE in the contract, forms with dropdowns, a dashboard for management. What has actually happened is that the Friday write-up has been typed instead of handwritten. The record is still authored after the event, still from memory, still filed by a person deciding which folder it belongs in — the medium changed and the epistemology did not.

The structural problem at stage 2 is that digitisation was scoped as a forms project rather than a capture project. Each compliance demand became another form; each form became another end-of-shift chore; and the quality of what goes into the forms degrades in exact proportion to their number. Site teams are rational: when the paperwork burden rises and the working day does not lengthen, the entries get shorter, later and more generic. A stage-2 CDE fills up with records whose consistency looks like rigour and whose content is boilerplate.

The second problem is coverage blindness. A shared drive full of PDFs can answer 'show me documents about riser 3' but not 'which requirements have no current evidence?' — because nothing links a record to the obligation it discharges. Stage 2 operators discover their gaps at exactly the wrong moment: in an audit, in a gateway submission, in disclosure. The records were searchable all along; what was missing was any definition of what a complete record set would look like.

Stage 3 inverts the relationship between work and record: documentation becomes a by-product of doing the work rather than a task performed after it. The foreman dictates ninety seconds at the gate and a model drafts the diary entry with plant, labour, weather and delays already structured. The engineer photographs the hold point and the image arrives in the CDE already classified — record type, location, ITP reference — because the capture app knew where it was and what was scheduled there. Nobody 'does the paperwork'; the paperwork condenses out of the workflow.

The discipline that makes stage 3 defensible rather than merely fast is the verification signature. A model's draft is not the record; the record is what a named person reviewed, corrected and signed, with the draft, the edits and the signature all retained. This distinction — draft versus record — is the single most important design decision on the whole ladder. It is what a lawyer will ask about first, it is what keeps a hallucinated sentence out of a document that might one day reach the HSE, and it is cheap: a review takes a fraction of the time the write-up used to.

Stage 3 is also where the requirement register appears, and it changes the site's self-knowledge more than any model does. The register is the list of what this project must be able to prove — by regulation, by contract, by ITP — with an owner and a retention rule per line. It is the first time the site knows what complete looks like, which means it is the first time anyone can honestly say how complete the record is. Most operators are startled by the initial number.

Stage 4 is the difference between a pile of good records and an evidence graph. Each captured record now carries machine-followable links — this photo evidences that ITP hold point, at this grid reference, released by this engineer, using this plant with this LOLER certificate. Once those links exist, the questions that used to take days become queries: show the full chain for fire-stopping on level 3; show every permit active in zone B on 14 March; show which lifting operations ran on certificates within thirty days of expiry. The record set has become a database with documents attached, rather than documents with a database aspiration.

Linking changes the temporality of compliance. At every earlier stage, gaps are discovered by audits — which is to say, too late, by the least sympathetic possible reader. At stage 4 a gap detector runs daily against the requirement register: a hold point released without its photographs, a RAMS briefing with no attendance record, a permit expiring on work that has not closed out. The gap surfaces as a task on a named person's list while the scaffold is still up and the subcontractor is still on site — when closing it costs minutes instead of a claim.

The frontier at stage 4 is the supply chain. A principal contractor's own records link cleanly; the steel erector's inspection sheets arrive as monthly PDF bundles and the piling contractor has its own app. Flow-down — putting the capture standard and CDE obligation into subcontract schedules, and providing the tools that make compliance easier than non-compliance — is slower than any engineering task on this ladder, and it is what separates a stage-4 project from a stage-4 business.

Stage 5 is narrower and less glamorous than 'AI-run compliance'. It means one specific thing: audit-readiness is a steady state, not a mobilisation. Coverage against the requirement register is a number computed daily for every live project; gaps have owners and ages; the golden thread on an HRB is a maintained dataset the BSR could query rather than a binder the project would assemble. Nothing about this requires exotic models — it requires the register, the graph and the verification discipline of stages 3 and 4 running as an operating rhythm rather than a project.

The work that sustains stage 5 is register maintenance, and it is a governance discipline, not an engineering one. Regulations move: Building Safety Act secondary legislation lands, approved documents are amended, a client's ESG reporting adds a duty. Each change must be triaged into the register — new lines, changed retention rules, new evidence types — on a cadence someone owns, or the system drifts into confidently computing coverage against last year's obligations. A stale register is more dangerous than no register, because it wears the costume of assurance.

The same rigour turns inward on the AI itself. By stage 5 the models that draft, classify and link records are part of the compliance estate and need their own file: what the model does, what data it saw, its error profile, its version history — the shape of discipline that ISO/IEC 42001 formalises for AI management systems and that the NIST AI RMF describes for risk. When a record is challenged, the operator must be able to say which model version touched it, what the human changed, and who signed. Stage 5 operators can. That answer, more than any capture technology, is what makes the whole ladder defensible.