What is as-built documentation?

The record of what was actually built, as opposed to what was designed. It can be a set of marked-up drawings, a measured survey, a 3D model, or all three. The distinguishing feature is provenance: every dimension should trace back to something observed on site rather than to the design intent.

What is the difference between as-built drawings and record drawings?

As-built drawings are the contractor's marked-up record of changes made during construction. Record drawings are the design team's tidy-up of those marks into the issued set. Neither is a measured survey, and neither is independently verified unless someone went and measured.

What is an as-built survey?

A measured survey of the finished condition, carried out by a surveyor with instruments, producing coordinates and dimensions with a stated accuracy. It is the only one of the three that carries evidence rather than assertion.

How accurate should as-built documentation be?

Accurate enough for what it will be used for, and stated. The USIBD Level of Accuracy bands are the usual language: LOA 20 (15 mm to 5 cm) covers most refurbishment and facade work; LOA 30 (5 mm to 15 mm) is for fit-out and construction documents. A deliverable with no stated tolerance cannot be accepted or rejected.

Can as-built documentation be produced automatically from a laser scan?

The geometry can, to a degree that is now worth measuring rather than arguing about. What cannot be produced automatically is the part that was never visible to the scanner: buried services, internal build-ups, and anything behind a finish.

As-built documentation: drawings, models and surveys, and which one you actually need

Three deliverables share the phrase “as-built”, cost between them varies by an order of magnitude, and scopes of work routinely ask for one while describing another. The confusion is expensive in both directions: a client pays for a measured survey and receives marked-up design drawings, or pays for marked-up drawings and expects to coordinate a refurbishment from them. This separates the three, says what each should contain, and gives the one clause that makes any of them checkable.

The three things

Deliverable	Who produces it	What it records	Evidence behind it
As-built drawings	The contractor, during and after construction	Changes from the issued design: substitutions, re-routes, dimensional deviations noticed on site	Observation and memory. Rarely measured.
As-built survey	A surveyor, after the fact	The finished condition, in coordinates, to a stated tolerance	Instruments. Traceable and repeatable.
As-built model	A modeller, from a survey or a scan	The finished condition as typed BIM elements with properties	Whatever the survey underneath it was, minus what modelling lost.

Record drawings are a fourth term and not a synonym. They are the design team’s tidied version of the contractor’s marks, reissued as a clean set. The tidying does not add measurement; it removes the evidence of where the marks came from.

Which one a job actually needs

The question to ask is not “what level of detail” but “what decision will be made from this”. Three common cases:

Handover and asset management. As-built drawings plus an asset register. Dimensional accuracy matters less than completeness and correct identification: the value is knowing what is there and where to find its manual.
Refurbishment or retrofit design. An as-built survey, and usually a model. Someone will design a new element to fit an existing opening, and a 3 cm error becomes a fabrication problem. This is the case where the marked-up drawings are not enough and get used anyway.
Dispute, insurance or compliance. A survey, with the raw observations retained. What is being bought is evidence, not a drawing; the deliverable has to be defensible by someone who was not there.

What an as-built survey should contain

A survey that cannot be audited is an opinion with coordinates. Four things make it auditable:

The raw capture, retained. The registered point cloud, the total-station observations, the control network. If the model is later questioned, this is the only thing that can settle it.
A control statement. What the survey is referenced to, how the network closed, and the residuals. “Site datum” without figures is not a control statement.
A stated accuracy, per element type. One tolerance for the whole model is almost always wrong: a scanned facade resolves openings better than it resolves a parapet edge seen at a grazing angle.
An explicit list of what was not captured. Occlusion is normal. Undeclared occlusion is what turns into a claim.

Specifying the accuracy so the result is checkable

Most as-built specifications name a Level of Development and stop. LOD says how far an element may be relied on — whether its geometry was modelled from a source or typed in as a placeholder — and explicitly does not set a tolerance. Pairing it with a tolerance is a separate clause, and the usual language is the USIBD Level of Accuracy:

LOA	Tolerance (95 % confidence)	Where it fits
LOA 10	5 cm to 15 cm	massing, feasibility
LOA 20	15 mm to 5 cm	most refurbishment and facade work
LOA 30	5 mm to 15 mm	fit-out, construction documents
LOA 40	1 mm to 5 mm	fabrication

The specification also separates measured accuracy (how good the capture is) from represented accuracy (how good the model is). These are not the same number, and the gap between them is where modelling error hides: a scan good to 8 mm carrying a model snapped to nominal 900 mm openings is not an 8 mm deliverable. Ask which one a quoted figure refers to. Our own take on how these two scales interact is in BIM LOD levels explained.

One clause worth adding to any as-built scope: “The supplier shall state, per element class, the proportion of elements whose dimensions were derived from measurement and the proportion assigned nominally, and shall deliver the capture data from which this can be verified.” It costs nothing to ask and it makes the difference between the three deliverables above impossible to blur.

Who owns it, and how long it has to last

As-built documentation outlives everyone who made it, which is the thing most contracts handle worst. Three clauses decide whether it is still usable in ten years:

Format. A deliverable in one vendor’s native format is a deliverable with an expiry date. IFC and the raw capture (E57 for scans) are readable without a licence; an .rvt from a version nobody runs any more is not. Ask for both if you want the working file, but require the open one.
The capture data, not just the model. Models get superseded; point clouds do not. The cloud is the only artefact that can answer a question nobody thought to ask at the time, and it is the cheapest part of the job to retain.
Licence and ownership. Surveyors frequently retain copyright and grant a use licence. That is normal and fine, but if the licence is limited to the current project the documentation cannot be reused for the refurbishment it was partly bought for.

How it goes wrong

The failure modes are consistent enough to list, and all of them are cheap to prevent at specification time and expensive to fix afterwards:

What happens	Why	The clause that prevents it
Openings are all nominal sizes	The modeller snapped to catalogue dimensions rather than the measured extents	Require the measured-versus-nominal split per element class
The model is accurate but incomplete	Occluded elements were silently dropped rather than flagged	Require an explicit list of what was not captured, and why
Two deliverables disagree	Drawings were produced from the design set, the model from the survey	Name one source of truth and derive everything from it
Nothing can be checked years later	Only the model was retained; the cloud and control were discarded	Require the capture data as part of the deliverable
The tolerance is unenforceable	An LOD was specified and read as an accuracy	State the level and the tolerance as two separate clauses

What it costs, roughly

Numbers vary by market and building, but the shape is consistent: capture is a day-rate exercise, modelling has historically been the larger half, and the modelling half is the one that has moved. A facade survey of a mid-rise block is typically a day or less on site; the manual modelling that followed it used to be several days. That ratio is why “we already have as-built drawings” has so often won the argument — the alternative was expensive.

Where modelling becomes a fraction of the capture cost, the calculation changes, and the honest comparison stops being “survey versus drawings” and becomes “measured model versus a record nobody verified”.

What automation changes, and what it does not

The modelling step — turning a registered cloud into typed BIM elements — is the part that has moved. It was the slow, expensive, error-prone middle of the process, and for facades it is now largely mechanical. What that changes is the economics of asking for a model at all: if the model is cheap relative to the survey, the argument for accepting marked-up drawings weakens.

What it does not change is the evidence problem. An automatically produced model is exactly as trustworthy as the measurement behind it and the review in front of it, which is why we publish per-element figures rather than a tolerance: on one named building of a surveyed estate, scored against the surveyor’s own IFC, 99 % of the windows found, 94 % of the windows we drew are theirs, 96 % of the wall surface covered, and every door found. On buildings nobody had read before, on clouds four times thinner, 75 % of the windows. The method and the current values are on the Accuracy page.

The part no pipeline reaches is the part the scanner never saw: services in a ceiling void, the build-up behind a render, a lintel above a closed window. As-built documentation that claims to cover those without opening anything up is describing the design, not the building.

In short

As-built drawings are a contractor’s record, an as-built survey is measured evidence, and an as-built model is one of those two turned into objects. Decide which the downstream decision needs, specify the level and the tolerance as two separate clauses, require the capture data alongside the deliverable, and ask for the split between measured and nominal dimensions. Everything that goes wrong with as-built documentation goes wrong because one of those four was left implicit.