Language models: what they can and cannot do, for construction

A language model reads, sorts and drafts text well. It cannot calculate reliably, and it can invent facts. You should also be careful with drawings. The reason lies in how it works: the model calculates which word is likely to come next, and without a connected source it looks nothing up. This leads to a clear division of work. The AI prepares text, a fixed program calculates, and you check and decide.

What a language model actually does

The BSI (the German Federal Office for Information Security) describes it like this: large language models generate text on the basis of stochastic correlations that they learned in training. They use probability distributions to predict which character, word or sequence of words might come next in a given context. According to the BSI, the outputs typically have a high linguistic quality and are often not easily distinguishable from text written by people. The publication is dated January 2025 (version 2.0); on the BSI’s website this was still the current version in November 2026.

A sentence can be fluent, polite and professional in tone and still be wrong. The wording does not tell you whether a source reference stands behind it or only a plausible pattern.

Some applications connect a language model to external sources and programs. The BSI describes this extension of the model’s functions and points out its risks. The content then comes from the source or the program, and the model only puts it into words. That does not change the division of work in this article, because here too you need to know where a figure or a source reference comes from.

1. Your documentsContracts, minutes, emails, lists
2. Fixed programCalculating, comparing, checking rules
3. Language modelReading, sorting, drafting
4. PersonChecking, deciding, answering for it

Text version of the graphic

• Layer 1 from the bottom: Your documents, Contracts, minutes, emails, lists
• Layer 2 from the bottom: Fixed program, Calculating, comparing, checking rules
• Layer 3 from the bottom: Language model, Reading, sorting, drafting
• Layer 4 from the bottom: Person, Checking, deciding, answering for it

What works well: understanding and preparing text

Among the opportunities of large language models, the BSI lists drafting formal documents, spelling and grammar checks, classifying texts, pulling out names and details (entity extraction), and summarising and translating. In a building project that means:

• Turning a few notes from a meeting into a draft of the minutes.
• Pulling the points on one topic out of a long exchange of correspondence.
• Splitting a variation text into sections and naming the key statements.
• Drafting a covering letter in the tone you want.

In all four cases your documents are the basis, and the model works on them linguistically. The result is a draft. Sign-off stays with you.

What does not work reliably

Calculating is not one of the model’s strengths. The BSI names mathematics only as a limited field of use and states that generative models offer no or insufficient guarantees for the quality of their outputs. A model can get one sum right and the next one wrong, and you cannot tell from the answer which case you have. Quantities, costs and dates are therefore calculated by a fixed program or a spreadsheet.

The model also knows facts only to a limited extent. According to the BSI, a model without access to real-time data, for example through a search, knows only what already existed when it was trained. Many models still answer questions on current topics and invent content that was in neither the input nor the training data. The BSI calls this hallucination. Such outputs mostly appear credible, even when they refer to sources that may themselves be invented. If you ask for a deadline, a standard number or a court ruling, you may get a convincing invention.

For reading drawings I have no source, and no measurement of my own. It is the author’s assessment: you should not let the model alone take dimensions, levels and symbols off drawings. Whether and how well a particular model processes images and drawings depends on the model. Test it on one of your own drawings before you rely on it.

A model does not reliably give the same result twice either. According to the BSI, outputs are not necessarily reproducible because of random components, so the same input can lead to different outputs. A checking record that reads differently on every run does not work as the only evidence.

Why the model fills gaps

The BSI states that generative models offer no or insufficient guarantees for the quality of their outputs, partly because of their probabilistic character. If a detail is missing from your documents, the model can fill the gap with whatever fits its patterns best.

As a countermeasure, the vendor documentation from Anthropic recommends explicitly allowing the model to say “I don’t know”, which it says can drastically reduce false information. For long documents (the documentation names more than 20,000 tokens), have the model extract word-for-word quotes first and base the analysis on those quotes. The model can also support its statements with quotes and withdraw a statement if it finds no supporting passage. The same page states that these techniques do not eliminate hallucinations entirely. Critical information still has to be validated.

As a countermeasure, the BSI recommends that outputs with potentially critical effects be checked, compared with other sources where needed and, if appropriate, finished by hand. This applies especially when they go outside the company.

What this means for the division of work

Task	Who
Turn notes into a draft of the minutes	Language model, then a person
Find passages in contracts and correspondence	Language model, check hits against the original
Calculate quantities, totals, prices, deadlines	Fixed program or spreadsheet
Take dimensions from drawings	Person, with a program where needed
Assess entitlement, price, approval	Person

An invented example shows the division. The notes of a site meeting become a draft of the minutes with columns for topic, person responsible and date. The model sorts and words. Whether a date really appears in the notes, you compare with the original, and a spreadsheet calculates any duration or total. The model may prepare whatever is linguistic. Whatever has to be right is calculated by a program or decided by a person. The time you gain lies in preparation and writing up. The model does not replace the check.

How such a workflow is built with a single AI step and a fixed approval is explained in the article AI in a workflow: one AI step, fixed rules, human approval.

Legal questions, for example on deadlines or liability, are not answered by the AI in a binding way. Your legal adviser clarifies the individual case. Project data belongs only on a platform that can be operated in line with the GDPR, not in a consumer chat service. Which data may go into which tool is covered in the article Data zones: which data may go into which tool?. Talk to your data protection officer before the first use.

What you can do now

1. Take a finished set of minutes and have the model write a draft from your old notes. Compare it point by point with the original.
2. Ask the model a question whose answer you know, such as a deadline from your contract. Note whether it names the source passage or guesses.
3. Decide which calculations run only in spreadsheets or programs.
4. Write down for your office who checks an AI output before it is passed on.