Skip to content
AI Metric

Chris M.

The future of construction take-offs: how AI could change measurement, estimating and cost planning

AI will change take-off less by measuring faster than by making every quantity traceable: where it was measured, from which drawing and revision, by what method, and how far it has been checked. Fast quantities nobody can trust are worse than slow ones. The opportunity is a take-off that reads the drawing as well as measuring it, shows its uncertainty, follows the UK's measurement rules and carries each quantity through to materials, prices and orders, with an estimator approving every step that matters.

Close-up of a printed structural drawing with dimensions and level marks, a sharpened pencil and a triangular scale rule.
Photo: Sven Mieke, Unsplash
2%Area error from a 1% scale error
60%How often AI models agree when both are wrong
9 Mar 2026RICS AI standard in force
P07A preliminary revision, not a contractual one

Part 1 | The problem

Why does a take-off take so long?

Because measuring is the smallest part of it. Before the first line is traced, the estimator has to establish which revisions are current, find the specification, read the general notes and understand how the elements relate.

Structural steelwork

Identify each member, check section sizes, calculate lengths, read the connection details, and total by section and grade.

Facade package

Measure cladding areas, deduct openings, count panels, review the interfaces, and check the specification for each system.

Then it all goes into a spreadsheet, where rates, labour, waste, access and method are applied. Then the drawings are revised, and a small design change can mean a substantial remeasure. The problem is not expertise. It is the manual work between a drawing and a reliable quantity.

A multi-storey structural steel frame under erection, with bolted connections and three mobile crane booms against an overcast sky.
Every member in this frame was measured, scheduled and priced from a drawing before anyone lifted it.Photo: Catia Climovich, Unsplash

Part 1 | The opportunity

Can AI read a drawing, or only measure it?

Today's tools, ours included, measure what the estimator points at. The next step is software that helps interpret the drawing. Given a set of steelwork drawings, an AI-assisted workflow could:

  • read drawing numbers, revisions, scales, notes and schedules;
  • recognise members and their labels, with section sizes and grades where stated;
  • flag missing dimensions and conflicting information;
  • produce a draft schedule with every line referenced to its drawing.

The estimator would start from a referenced draft rather than a blank spreadsheet, which only helps if every line can be checked.

Part 2 | The source

Does it matter what kind of drawing file you measure from?

More than anything else. A native CAD or BIM file, a vector PDF and a scan can look identical on screen and hold completely different information.

The file decides what software can know

SourceWhat software can readChecks still needed
Native CAD or BIMExact geometry, layers, objects and often their propertiesThat it is the current revision, and properties match the specification
Vector PDFPrecise line geometry and searchable textScale calibrated and verified against a known dimension
Scanned or flattened PDFAn image: no geometry, and no text until it is recognisedScale verification, dimension cross-checks and manual review

AI image enhancement can make a scan easier to read. It cannot recover geometry that was never captured, and scale is where small errors become large ones.

A 1% SCALE ERROR IS 2% ON EVERY AREA AND 3% ON EVERY VOLUMELENGTH ERRORAREA ERRORVOLUME ERROR0%4%8%12%16%0.5%1.0%1.5%0.5% SCALE ERROR1.0%2.0%3.0%1% SCALE ERROR2.0%4.0%6.1%2% SCALE ERROR5.0%10.3%15.8%5% SCALE ERRORHOW TO READ THISEach group is one error in the scale. Areas carry it twice and volumes three times.An A1 drawing printed on A3 halves every length and quarters every area: calibrate every sheet.
A scale error carries straight into lengths, doubles on areas and roughly triples on volumes.

The same numbers, as a table

Scale errorLengthAreaVolume
0.5%0.5%1.0%1.5%
1%1.0%2.0%3.0%
2%2.0%4.0%6.1%
5%5.0%10.3%15.8%

Part 2 | The record

What should a trustworthy measurement record show?

Where the number came from, how it was obtained and how far it has been checked.

One line of a steelwork take-off, with its evidence

Element
UB 457 × 191 × 67
Location
Level 02, grid B to C
Quantity
4 No.
Measured length
6,250 mm
Source
Drawing S-204, revision P07
Method
CAD geometry cross-checked against the dimension
Status
Verified measurement

The manufacturer's section tables give that beam 67.1 kg/m, so four at 6.25 m weigh 1.68 tonnes: a calculated quantity resting on verified inputs. Where something is uncertain, the record should say so rather than assume.

EVERY QUANTITY CARRIES HOW IT WAS CHECKEDA status is evidence of a check, not a percentage a model gives itself.VERIFIEDMeasured from thegeometry and confirmed bya second sourceEXAMPLEBeam length from thedrawing geometry,matching the dimensionstring: 6,250 mmCALCULATEDDerived from verifiedinputs by a stated ruleEXAMPLEWeight: 4 No. x 6.25 m x67.1 kg/m from thesection tables = 1.68 tNEEDS REVIEWSomething is missing,unclear or in conflictEXAMPLESection size is clear,but the length is notdimensioned and the gradeis not stated
Three statuses, each meaning a different kind of check has or has not happened. The estimator checks the orange items first.

A percentage a model assigns to its own output is not evidence. The aim is not to make every quantity look certain. It is to make the level of certainty visible and explainable.

Part 2 | The checks

Will several AI checks catch each other's mistakes?

Only partly. A system could split the work, with separate processes reading annotations, extracting geometry, calculating quantities and comparing them with schedules. But AI models are not independent in the way two estimators are: a 2025 study of more than 350 models found they choose the same wrong answer far more often than chance.

AI MODELS SHARE THEIR MISTAKES: ABOUT 60% AGREEMENT WHEN BOTH ARE WRONG0%25%50%75%100%If wrong answers were random33%What the models actually didABOUT 60%HOW TO READ THISMore than 350 models tested. On one benchmark, 97.5% of model pairs agreed on wrong answers more often than chance.A second model checking the first is not an independent check. Geometry, schedules and a person are.
When two models were both wrong, they agreed about 60 per cent of the time, against 33 per cent if their mistakes were random.

The same finding, as a table

When two models are both wrongThey give the same wrong answer
If their wrong answers were random33%
What more than 350 models actually didAbout 60%

So the checks that matter are the ones that work differently: deterministic geometry, the steel schedule, the dimension strings and a person. Multiple AI passes add rigour. They do not replace professional judgement.

Part 2 | The link to buying

How do quantities connect to materials, costs and buying?

Measurement is only the start. A facade take-off can feed a panel schedule for supplier pricing, steel members can be linked to section-weight tables and fabrication allowances, and excavation volumes to disposal and backfill. The estimator still sets the method, productivity, risk and commercial strategy.

FROM DRAWING TO ORDER, WITH A PERSON AT EVERY DECISIONSoftware carries the information between steps. People approve at the three gates.DRAWINGCurrentrevisionQUANTITIESEach traced toits markMATERIALSSchedule fromthe specSPEC CHECKAlternativescomparedENQUIRIESSent tosuppliersQUOTESPrice, leadtime,complianceORDERPlaced andrecordedEstimator approvesquantitiesDesigner approvesany substitutionBuyer approves theorder
Software carries the information from drawing to order. People approve at three gates: the quantities, any substitution, and the order itself.

The prize is a clear line between what was designed, measured, priced and ordered, which is where omissions and duplicated quantities hide. On a higher-risk building, a substitution can also trigger the Building Safety Regulator's rules on controlled changes, as set out in construction is under pressure.

Part 2 | Revisions

What happens when the drawings are revised?

Finding what changed on a large package is real work. AI-assisted comparison could produce a change report instead of a full remeasure:

Drawing S-204: revision P07 to P08

  • Two structural members changed
  • One additional opening
  • Three measured quantities potentially affected
  • One specification note revised

Not yet entitlement

Under NEC4 a change to the Scope is a compensation event when the Project Manager instructs it, unless it corrects a defect or the Contractor's own design. Under JCT, a Variation needs a written instruction from the Employer's side.

Tracking variations before they become disputes covers the paperwork that turns a change into money.

Part 3 | The rules

Which UK bodies set the rules an AI take-off has to follow?

Six sets of rules apply whether a person or software measured the quantity.

RICS: NRM 1, 2 and 3

The New Rules of Measurement: cost estimating and cost planning (NRM 1), detailed measurement and bills of quantities (NRM 2), and maintenance (NRM 3). Current editions date from October 2021; RICS says earlier ones should not be used on new projects.

RICS: responsible use of AI

In effect since 9 March 2026 wherever AI output has a material impact on a service. A named surveyor records whether an output is reliable, and the client is told in writing before work begins.

CIOB: Code of Estimating Practice

The Chartered Institute of Building's guidance on estimating and pricing, linking the estimate to production planning, resources, time and cost control.

ICE: CESMM4 Revised

The Institution of Civil Engineers' standard method of measurement for civil engineering work, revised in 2019 and usable with NEC, FIDIC and ICC contracts.

UK BIM Framework: ISO 19650

The UK guidance on managing project information. Its revision codes mark a P revision as preliminary and non-contractual, and a C revision as contractual.

Building Safety Regulator

On higher-risk buildings in England, changing a product after approval can be a recordable, notifiable or major change, depending on its fire performance.

Part 3 | In practice

What do those rules mean for an AI-assisted take-off?

Does the RICS AI standard apply to take-off software?

For RICS members and regulated firms, yes, wherever the output materially affects the service. The firm needs a record of its AI systems, a responsible use policy, supplier due diligence and randomised checks on high-volume outputs, and a named surveyor accepts responsibility for each output relied on. A simple AI register holds most of that.

Does an AI take-off have to follow NRM 2?

If the bill says it is measured to NRM 2, every quantity in it follows NRM 2's rules for measurement and description, whatever produced it. The method comes from the tender documents, not the tool.

Can you price from a P revision?

You can, but you are pricing preliminary information. Record the revision against every quantity, as the record above does, and qualify the tender accordingly.

Who signs off an AI-assisted quantity?

A competent, named person, because the RICS standard, the contract and the client all expect someone to stand behind the number. Where that line sits for each task is in the human in the loop matrix.

Part 3 | The estimator

Will AI replace construction estimators?

Not in any future that matters to this year's tenders. Estimating is buildability, sequencing, temporary works, access, productivity, market conditions and commercial risk, none of which reduces to a measurement. What changes is the shape of the day.

Less time on

  • Transferring measurements into spreadsheets
  • Searching through drawing revisions
  • Preparing repetitive schedules

More time on

  • Checking quantities and challenging assumptions
  • Developing the construction method
  • Finding the commercial risk

Part 3 | What we are building

What is AI Metric building?

A browser take-off tool, Metric Takeoff, in private beta with invited teams, and the AI layer that comes after it. The principle throughout: an estimator can trace any quantity back to its source, see how it was calculated, and correct it.

AI MetricMetric Takeoff, private beta, sample drawing
Metric Takeoff showing a sample flat floor plan with measured rooms, walls and openings, one partition selected and highlighted on the drawing, and the bill grouped by NRM 2 work section on the right.
A sample drawing in Metric Takeoff. The selected partition is highlighted on the plan, and the bill on the right groups every quantity by NRM 2 work section. The rates are illustrative, not real prices.

In Metric Takeoff today

  • Scale calibrated, then verified against a second dimension
  • Snapping to real geometry in vector PDFs and DXF files
  • Every quantity linked to its mark on the drawing
  • NRM 2 work sections, priced from a rate library
  • Superseded sheets flag their measurements for re-checking
  • Export to Excel with a drawing reference per line

Planned, not built yet

  • Reading members, labels and schedules from the drawing
  • Statuses based on recorded checks
  • Revision comparison reports
  • Links to material schedules and supplier quotations
  • DWG files, BIM models and Uniclass coding

We also do this as consultancy: AI for estimating and take-off around the tools your team already prices in, starting with one measured workflow from £1,500. The take-off that wins will be the one that can be checked, explained and trusted.

Sources

Which sources is this article built on?

Every fact above resolves to one of these, each read at source before publication.

Photographs by Sven Mieke and Catia Climovich on Unsplash, used under the Unsplash Licence. The Metric Takeoff screenshot is our own, on a sample drawing.

AI Metric is a construction-native AI consultancy. If your team is spending more time operating software than doing their job, book a 30 minute call.