The BIM model gives you 70% of the BOQ. The other 30% is your margin

The model is LOD 400. Fabrication-ready, every element with accurate size and position. The take-off still takes three days.

That isn't the modeller's fault, and it isn't the estimator's. There is a gap between the two that rarely gets discussed openly.

What the model gives you

The measurable part comes out of the model, and it comes out accurately. Air handling unit and chiller counts, ductwork area in square metres, pipe runs in metres, insulation area, fan coil and VAV terminal counts, valves and supports by number.

This is genuinely valuable. Every length, area and count is measured, not guessed. Roughly 70% of the bill of quantities.

But you cannot price a tender from it.

What it doesn't give you

The remaining 30% is manual, every single time:

• Material wastage and cutting allowance — typically 5–10%, and material-specific
• Consumables — gaskets, sealant, welding consumables, fixings
• Installation labour hours, access and site distribution
• Testing, adjusting and balancing (TAB)
• System flushing, pressure testing and commissioning
• Generic design objects — no part numbers, prices or lead times

That last one is the expensive one. "AHU, 12,000 m³/h" is not something procurement can buy. Someone still has to find the manufacturer, the price and the lead time — item by item, by hand.

That 30% is about three days per tender package. And because the deadline doesn't move, whatever gets missed is absorbed by the margin.

How the gap closes

An AI take-off assistant doesn't re-measure the model. It adds what the model cannot hold.

It reads the specification for consumables and fixings and adds them to the BOQ — on one real project that was 1,940 items. It applies material-specific wastage, not a blanket multiplier but a figure learned from that company's own completed projects: 5% on ductwork, 7% on pipework. It adds installation, TAB and commissioning hours from historical project data — 4,120 hours on the same project. Then it cross-checks the specification against the model and flags what's missing: seven items went to the estimator to confirm.

The decisive step, though, is matching. Every generic model item is matched to a real supplier product with a part number, unit price and lead time. "Duct, 600×400, class B" becomes a specific product at a specific price with an eight-day lead time.

What comes out is no longer a bill of quantities. It is a purchase-ready buying package, grouped by trade and supplier.

Where the wastage figures come from

Wastage percentage is the line where the most money rides on an assumption. The industry rule of thumb for ductwork is 5–10%, and most estimates use the upper end, because the upper end is safe.

Except that the safe number is a competitive disadvantage. If real wastage is 5% and you price at 10%, the tender is dearer than it needs to be — and the difference isn't margin, because a competitor submitted the same work for less.

So the system doesn't use an industry average. It uses actuals from that company's completed projects: how much material was ordered, how much was installed, what the real difference was. Broken down by project type and by material. That is where 5% on ductwork and 7% on pipework come from — measured, not estimated.

The estimator approves every number

This doesn't change, and it shouldn't.

The system measures and matches; the estimator reviews. On one real tender the estimator adjusted the hydronic installation rate from €32 to €34.50 — because they knew the site and the system didn't. The tender came to €665,350, every line approved by a person.

AI makes no pricing decisions. It prepares the ground so there is something to decide from.

What changes

• Take-off time: from three days to three hours
• Items missed: from several to effectively none
• Wastage allowance: from guessed to measured
• Tender submission: from at-risk to on time

Who owns the data

This question comes up first, and rightly. A take-off system sees a company's most sensitive commercial information: purchase prices, the wastage it allows for, the margin it applies, what it submitted a tender at.

If that data goes into a shared model serving multiple users, a competitor eventually benefits from it. Which is why the system runs in the client's own EU cloud environment, and why the data isn't used to train a third party's general models.

Learning still happens — it just stays inside the company's own data. The company learns from its own history, not the industry's.

Subcontract packages

The output isn't one long list. It is split by trade and supplier: ductwork, hydronics, plant, insulation and TAB, each as its own package.

That matters because you don't ask for quotes against a whole bill of quantities — you ask package by package. When the split is produced automatically, RFQs go out the same day rather than three days later, and the quotes that come back are comparable, because they are priced against the same line items.

Why now

Earlier attempts at automating take-off failed on two things: rule-based systems couldn't interpret a specification, and they couldn't match a generic object to a real product. Both of those have changed.

For MEP contractors and main contractors the question is simple: how much of your take-off time goes on the part BIM was never going to give you?