AI handover document automation in construction and infrastructure: from a pallet of manuals to a verified asset record

Stage 1 is not disorganised — it is sequenced wrongly. Everyone knows what the handover pack should contain, and most of the constituent documents genuinely exist somewhere. What does not exist is any mechanism that makes information arrive while the people who created it are still contactable. Handover therefore becomes an archaeology exercise conducted against a contractual deadline, and archaeology under time pressure produces a pack that satisfies a checklist and fails a maintenance engineer.

The structural cause is that handover information has no commercial weight until the very end. A subcontractor's cash is released against installed work, tested systems and defects closed; the O&M pack is a condition of the final account, which is negotiated months later by people who were not on site. So the information deliverable is rationally deferred, and the deferral compounds: by the time anyone reads it, the commissioning engineer who could have identified the unlabelled pump has moved to another job.

What makes this stage expensive is that the cost lands on someone else. The contractor absorbs a few weeks of chasing; the operator absorbs sixty years of not knowing. NIST's study of interoperability costs in the US capital facilities industry put a number on exactly this asymmetry — two-thirds of the quantified cost is borne by owners and operators, most of it during ongoing operation and maintenance. Stage 1 is not a failure of diligence. It is a transfer of cost across a contractual boundary, and it will keep happening until the information becomes a condition of payment rather than a condition of goodwill.

Stage 2 is the most common place to be and the easiest place to mistake for competence. The tracker is real, the folder structure follows the naming convention, the information manager can produce a percentage at the progress meeting, and that percentage rises through the job. Every visible instrument says the handover is under control. The instruments are measuring the wrong noun.

A document count is a proxy for completeness that fails in one specific and predictable direction: it cannot detect a document that is present, correctly named, correctly filed and about nothing you need. The manufacturer's 200-page generic catalogue satisfies 'O&M manual — mechanical, level 3' as thoroughly as a properly cut, asset-specific manual does. So a tracker at 96% is compatible with an asset record that answers no operational question at all, and the divergence between the two numbers is invisible until the FM team starts asking.

The other thing stage 2 hides is that compliance with the convention is itself unevenly distributed. The main contractor's own packages comply; the specialist sub-subcontractors two tiers down submit through whoever has a login. Because the tracker only knows about the file that arrived, information supplied through a tier-three route often arrives as a photograph of a certificate emailed to a site manager, and never enters the count at all. The tracker is not wrong. It is merely reporting on the part of the supply chain that reads the EIR.

Stage 3 is where handover stops being an end-of-job event and becomes a running process. The mechanism is unglamorous: a schema that states what must be known about each asset type, an engine that checks submissions against it, and a routing rule that sends each failure to the party who can fix it. Extraction is the part that gets the attention, but the schema and the routing are what convert a report into work that actually gets done.

The timing shift is the whole value. A validation engine that reports 1,400 missing attributes three days before practical completion has produced a liability register — everything on it is now a claim, a retention argument or a job for the operator. The identical engine run against the week-34 information drop produces a snag list. The M&E subcontractor still has commissioning engineers on site, the manufacturer's technical line still recognises the order number, and the fix costs an afternoon rather than a legal opinion. Everything on this page is ultimately an argument about moving that report earlier.

The discipline stage 3 introduces is citation. An attribute that arrives without a pointer back to the document and page it came from is indistinguishable from an attribute somebody guessed, and language models are extremely good at producing plausible values for fields they cannot find. Storing the citation alongside the value costs nothing at extraction time and is the difference between a record you can defend in a dispute and a spreadsheet with confident numbers in it. Teams that skip it discover the omission at the worst possible moment, when someone asks where a figure came from and the honest answer is that nobody knows.

Stage 4 is the point at which the record becomes a description of a building rather than a description of a document set. The binding artefact is the asset register: one controlled list where the model's object, the physical tag on the plant and the attributes extracted from the documents all resolve to the same identifier. Building it is mostly reconciliation work, and reconciliation is where the interesting failures live — the model has 44 AHUs and the tag survey found 42, or the commissioning certificates reference a numbering scheme the installer abandoned in month four.

Those discrepancies are the product, not the problem. At stage 3 a mismatch between a certificate's tag and the register's tag is a validation failure with no owner; at stage 4 it is an exception routed to a person who walks to the plant room and settles it. The exception rate falls steadily through the last quarter of a job, and it is the single most honest readiness indicator available — far more predictive than any tracker percentage, because it is measuring disagreement between independent sources rather than the presence of files.

The commercial consequence shows up on day one of operation. When the record is asset-true, the CAFM or CMMS load is an export with a defined schema, run in a morning, with the ability to reload if something is wrong. When it is not, the FM provider runs a data-cleansing project for the first year of the contract — which the operator pays for twice, once in the FM mobilisation fee and again in the maintenance that was deferred while the register was being rebuilt. The whole-life cost of getting this wrong sits with the owner, which is precisely the asymmetry NIST quantified across the facility life cycle.

Stage 5 is narrower than the phrase 'autonomous handover' would suggest, and deliberately so. What executes without a person is a bounded class of decision: accept this attribute, for this non-critical asset, where the value is cited to a page, conforms to the schema, and is not contradicted by another source. Everything else — anything touching fire, life safety, structure, pressure systems or a higher-risk building's safety case — escalates by policy and always will. The engineering to go further exists; the reason not to is that the consequence of a wrong acceptance is not symmetric with the saving.

What actually distinguishes this stage is that handover stops being an event with a date. Information is accepted progressively at defined drops, so at any point in the last third of the programme there is a current, honest statement of what is known about the asset and what is not. That statement is what soft-landings aftercare, an operator's mobilisation and — for higher-risk buildings — the golden thread obligation under the Building Safety Act 2022 all need to exist against. A record that is assembled once, at PC, cannot serve any of them, because all three are continuous requirements and PC is a moment.

Sustaining stage 5 is a governance discipline and it is the stage most likely to regress. The asset changes: a chiller is replaced, a floor is refitted, a fire strategy is revised. If those changes do not re-enter the same validation and citation path that construction information did, the record decays from the day it is handed over, and the decay is invisible until somebody relies on it. The operators who hold this stage treat the acceptance policy as a reviewed, versioned artefact and monitor record currency the way an engineering team monitors an error budget — as a number with a threshold and an owner, not as an assumption.