Checking a bill of quantities: common errors and AI support

An AI-assisted check finds in bills of quantities (BoQ) mainly inconsistencies that slip through when skim-reading: preliminaries that contradict an item, duplicate items, units that do not fit the work, gaps between trades and references to standards with a conspicuous year of issue. It delivers notes with a source reference and a query. It is not an expert opinion, and it says nothing about safety under procurement law. Responsibility for the bill of quantities stays with the designer.

Five types of error that slip through when reading

Hardly anyone reads a bill of quantities of forty pages in one go. The errors sit at the seams: between preliminaries and item, between two trades, between two stages of editing. The table shows what a check looks for. The examples are invented.

A contradiction between preliminaries and item can later lead to a dispute over a variation. How the check looks from the variation’s side is described in Checking variations. If the bill of quantities exists as a GAEB file, GAEB file explained helps with reading the file.

What section 7 VOB/A requires of the description

The VOB/A (the German award rules for construction works, Part A) regulates the award of construction works. Part 1 covers awards below the EU thresholds and exists in the 2019 edition (BAnz AT 19 February 2019 B2). Most recently, section 3a with the value thresholds was restated there, from 1 January 2026 (BAnz AT 16.12.2025 B7). The following citations come from sections 7 and 7b of the 2019 edition.

For Parts 2 and 3, the Federal Ministry for Housing, Urban Development and Building announced a new version on 22 July 2026, published on 24 August 2026 (BAnz AT 24.08.2026 B6). It has applied since 8 September 2026: on that day the ordinance entered into force that switches the reference in the VgV and in the VSVgV (the German public procurement ordinances) to this notice (Federal Law Gazette 2026 I No. 250). Section 7 EU of the new version is worded similarly to section 7 of Part 1. Anyone who does not have to apply the VOB/A can still use it as a guide to a clean description.

Under section 7(1), the work is to be described unambiguously and so exhaustively that all companies understand the description in the same way and can calculate their prices reliably without extensive preliminary work. All circumstances that affect price calculation are to be established and stated in the tender documents. No unusual risk may be placed on the contractor. Contingent items (Bedarfspositionen) are in principle not to be included, and the notes in section 0 of the ATV DIN 18299 ff. (the technical contract conditions) are to be observed. Under section 7b(4), only work that is technically and, for pricing purposes, of the same kind belongs under one item number. Work of different kinds may be combined only if one part has no significant influence on the average price.

A check translates this into questions to the text. Does item 05.03 combine three works that are priced differently? Does the bill contain contingent items although they do not in principle belong in it? Is there missing information on site conditions that determines the price? Whether a text in the end meets the requirement is a legal assessment. That belongs with your procurement lawyer.

How the check works

The check runs in six steps, and the AI takes on only two of them: the text comparison and the extraction of references. Everything else is done by a fixed program and a person.

In the text comparison, the model puts preliminaries and items side by side and marks where they contradict each other, where the same text appears twice and where a unit does not fit the work. This is reading work, where language models can help. They do, however, generate text from patterns, and they know neither your project nor the author’s intention. The Federal Chamber of Engineers (Bundesingenieurkammer) states that when AI tools are used, it is always engineers who in the end take personal responsibility, and that AI-based results must be checked critically for plausibility. For a bill of quantities that means every note is assessed by a specialist, and the designer’s liability is unchanged.

Calculation is not done in the language model. Totals, quantity times unit rate and surcharges are recalculated by a fixed program or a spreadsheet, using the quantities that appear in the bill. A language model also words wrong numbers convincingly. That is why a number appears in the list of notes only if the program has confirmed it; otherwise it carries the note “unchecked”.

References to standards: checkable only against a list

An example shows why. DIN 18299 belongs to the VOB/C (the German technical contract conditions for construction works) and contains the general rules for construction work of every kind. According to the entry at DIN Media, it exists in the 2023-09 edition in November 2026 and replaces the 2019-09 edition. If an older edition still appears in a bill of quantities, it may be an oversight. It may also be part of the contract.

What a language model believes it knows about current editions therefore does not count as evidence. The references are compared with a list of current editions that comes from a reliable source, for example from your standards management. The result is “outdated?” and comes with a query. The decision path in the second graphic shows when a note arises and when it does not.

A talk at the Day of Public Contracting Authorities (Tag der öffentlichen Auftraggeber) 2026 advises contracting authorities that use AI to check references to standards themselves and to confirm every named standard against the original source (slide 19). The same care belongs in an office that prepares a bill of quantities. Whether outdated references can be detected reliably by automatic means could not be substantiated. That is another reason why the comparison needs the list. How to trace an AI answer back to its source is described in Why AI answers need a source reference.

What a note looks like

Each note has four parts: source reference, observation, type and query. An example with invented data: the source reference is item 03.02.0040 and preliminary 3.1. The observation is that the preliminary includes disposal in the unit rates, while the item charges it separately per tonne. The type is “contradiction”. The query asks whether disposal is to be paid separately and whether the preliminary then has to be adjusted.

The notes are worded as queries, because the author knows the project and decides. At the end, the most frequent patterns are summarised. If contradictions with the preliminaries recur, the cause may lie in a template that was taken over. This can then be fixed in one place and not in forty.

What the check does not do

The standard check certifies neither conformity with standards nor safety under procurement law. It does not read drawings and does not assess bidders. It therefore does not test whether the quantities in the bill match the drawings. Taking off quantities from drawings is feasible, but it is a separate, additional step with its own effort and its own checking. It is not part of the check described here and has to be agreed separately.

Bills of quantities usually contain no personal data but are confidential before the contract is awarded. No file belongs in a consumer chat service for that. Which platform can be operated in line with the GDPR for your documents is something you clarify with your data protection officer.

What you can do now

1. Take a completed bill of quantities and read the preliminaries against five items. Note how long this takes and what you find.
2. Draw up a list of the standards, with their valid edition, that your texts refer to, and decide who maintains it.
3. Use the search function to look for words such as “included in the price” and “separately” (in a German BoQ: “eingerechnet” and “gesondert”) and check whether the items fit.
4. Clarify in advance where documents from ongoing awards may be processed.